COMPREHENSION

A Paradigm for Cognition

Walter Kintsch
153
K62c
1998
 Digitized by the Internet Archive
in 2020 with funding from
Kahle/Austin Foundation

https://archive.org/details/comprehensionparOOOOkint
In this landmark volume, Walter Kintsch presents a theory of human text com¬
prehension that he has refined and developed over the past 20 years. Character¬
izing the comprehension process as one of constraint satisfaction, this compre¬
hensive theory is concerned with mental processes - not primarily with the
analysis of materials to be understood. The author describes comprehension as a
two-stage process: first, approximate, inaccurate representations are constructed
via context insensitive construction rules, which are then integrated via a process
of spreading activation.
In Part I, the general theory is presented and an attempt is made to situate it
within the current theoretical landscape in cognitive science. In the second part,
many of the topics are discussed that are typically found in a cognitive psychol¬
ogy text. How are word meanings identified in a discourse context? How are
words combined to form representations of texts, both at the local and global
level? How' do texts and the mental models readers construct from them repre¬
sent situations? What is the role of working memory in comprehension? The
book addresses how relevant knowledge is activated during reading and how
readers recognize and recall texts. It then draws implications of these findings for
how people solve word problems, how they act out verbal instructions, and how
they make decisions based on verbal information.
Comprehension is impressive in its scope, bringing in relevant ideas from all of
cognitive science. It presents a unified, sophisticated theory backed by a wealth
of empirical data.
- '
Comprehension
- '
Comprehension
A paradigm for cognition

Walter Kintsch
University of Colorado, Boulder

Cambridge
UNIVERSITY PRESS
PUBLISHED BY THE PRESS SYNDICATE OF THE UNIVERSITY OF CAMBRIDGE

The Pitt Building, Trumpington Street, Cambridge CB2 1RP, United Kingdom

CAMBRIDGE UNIVERSITY PRESS

The Edinburgh Building, Cambridge CB2 2RU, United Kingdom
40 West 20th Street, New York, NY 10011-4211, USA
10 Stamford Road, Oakleigh, Melbourne 3166, Australia

© Cambridge University Press 1998

This book is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without
the written permission of Cambridge University Press.

First published 1998

Printed in the United States of America

Typeset in Ehrhardt

Library of Congress Cataloging-in-Publication Data

Kintsch, Walter, 1932-

Comprehension : a paradigm for cognition / Walter Kintsch.
p. cm.
Includes bibliographical references and index.
ISBN 0-521-58360-8 (alk. paper). - ISBN 0-521-62986-1 (pbk. :
alk. paper)
1. Comprehension. I. Title.
BF325.K56 1997
153 - dc21 97-26406
CIP

A catalog record for this book is available from

the British Library

ISBN 0 521 58360 8 hardback

ISBN 0 521 62986 1 paperback

4PJ5 i
For Eileen
viii Contents

3.1.2. The macrostructure 64

3.1.3. The psychological reality of
propositions 69
3.2. The propositional representation of
knowledge 73
3.2.1. The construction of meaning:
Concepts 74
3.2.2. Emergent structures 82
3.3. Latent semantic analysis: A vector
representation 86
4. Modeling comprehension processes: The
construction-integration model 93
4.1. Construction and integration 96
4.2. Processing cycles in text comprehension 101
4.3. Textbase and situation model 103
4.3.1. Imagery as a situation model 108
4.3.2. A script-based situation model 111
4.4. Earlier versions of the model 118

II. Models of comprehension 121

5. Word identification in discourse 123

5.1. The construction of word meanings 123
5.1.1. From the retina to working memory 124
5.1.2. Homographs 128
5.1.3. Local and global priming in discourse:
What words mean 130
5.1.4. A simulation of priming effects in
discourse 136
5.2. Anaphora 144
5.2.1. The psycholinguistics of anaphora
resolution 144
5.2.2. Anaphora resolution in the Cl model 149
5.3. Metaphor 157
6. Textbases and situation models 167
6.1. Parsing 168
6.2. Macrostructure formation 174
6.2.1. Signaling the macrostructure 177
6.2.2. Dominance of the macrostructure:
Contradictions in the text 181
 Contents ix

6.2.3. Macrostructures as vectors in the

LSA space 184
6.3. Inferences 188
6.3.1. Classification of inferences 189
6.3.2. Inference generation during
discourse comprehension 193
6.3.3. Time course for constructing
knowledge-based inferences 198
6.3.4. The construction of situation
models 198
6.4. Spatial situation models 199
6.5. Literary texts 204
6.5.1. Versification as a source of
constraints 206
6.5.2. Multilevel situation models 209
6.5.3. Comprehension strategies for
literary language 211
7. The role of working memory in comprehension 215
7.1. Working memory and skilled memory 217
7.2. Long-term working memory in text
comprehension 221
7.3. Retrieval structures in the Cl model 224
7.3.1. Episodic text memory during
comprehension 224
7.3.2. The activation of knowledge
during comprehension 227
7.3.3. The short-term memory buffer 234
7.4. An LSA model of retrieval structures 235
7.5. Capacity differences in working memory? 238
8. Memory for text 247
8.1. Recognition memory 247
8.1.1. List-learning data and theory 247
8.1.2. Sentence recognition 251
8.2. Recall and summarization 264
8.2.1. Simulations of story recall 268
8.2.2. Evaluation of summaries 277
9. Learning from text 282
9.1. What makes for a good reader? 282
9.1.1. Decoding skills 282
X Contents

9.1.2. Language skills 284

9.1.3. Domain knowledge 287
9.2. Learning and memory 290
9.2.1. Textbase and situation model 291
9.2.2. Learning 293
9.3. The measurement of learning 295
9.4. A simulation of learning with the
Cl model 302
9.5. Using coherent text to improve learning 307
9.6. Improving learning by stimulating active
processing 313
9.7. Matching students with instructional texts 323
9.8. Educational implications 327
10. Word problems 332
10.1. Word arithmetic problems 333
10.1.1. Schemas for solving word
arithmetic problems 334
10.1.2. Error simulation 339
10.1.3. Situation models 345
10.1.4. The Cl model for word problem
solving 347
10.1.5. The emergence of schemas 354
10.2. Word algebra problems 358
10.2.1. Algebraic schemas 359
10.2.2. A model for word algebra problems 360
10.2.3. The ANIMATE tutor for solving
word algebra problems 364
11. Beyond text 372
11.1. Action planning 372
11.1.1. Routine actions 372
11.1.2. Novices and experts 381
11.1.3. Simulation of errors 387
11.2. Problem solving, decision, and judgment 394
11.2.1. The role of comprehension in
problem solving and decision
making 394
11.2.2. Impression formation 402
 Contents xi

11.3. The representation of the self in working

memory 409
11.3.1. Evaluative and motivational
functions of self representation 413
11.3.2. Cognition and emotion 419
11.4. Outlook 421

References 425
Name Index 449
Subject Index 459
-'
 Preface

A few years ago some of my friends, colleagues, and former students put
together a book entitled Discourse Comprehension, with chapters that
relate in some way to work that I have done in this area. They are excel¬
lent chapters and illustrate how far research on discourse processes has
advanced in the last few decades. Building on and incorporating earlier
work in linguistics and philosophy, cognitive scientists have gained new
insights about how language is understood and have refined their experi¬
mental methods and elaborated their theories. Indeed, although the con¬
tributors to that book did not represent a single or uniform theoretical
viewpoint, it seemed to me that the current research on discourse com¬
prehension in cognitive science, at least in the section represented there,
could be brought under a unifying, overarching theoretical framework. In
fact, I thought I knew what that framework could be - the construc¬
tion-integration (Cl) model of comprehension that I had been develop¬
ing. More than that, I believed that that framework was broad and pow¬
erful enough to include related areas in cognitive psychology beyond
discourse comprehension proper.
The only problem was that this framework was nowhere stated explic¬
itly and in detail. I had published the basic ideas, and several applications
of the model had also appeared in the last few years, but a complete and
coherent treatment was lacking. Although it seemed that the Cl model
could provide an account for many interesting findings in the study of
cognition, in many cases the actual work of developing and testing a
detailed simulation of the phenomena in question had not yet been done.
So I, could not even be sure that the model was really as powerful as I sur¬
mised. Worse, the applications that had been done were often published
in inaccessible places, making it difficult for a reader to gain an under-
xiv Preface

standing of the scope of the model and to evaluate its effectiveness. Fur¬
thermore, even if a reader studied all the available publications, it would
not have been easy to arrive at a coherent picture, because in the pub¬
lished papers each application of the model stood by itself, and the all-
important links between them were rarely made explicit.
Therefore, this book was written to present to my colleagues and stu¬
dents interested in discourse comprehension a coherent and broad theo¬
retical framework to better investigate its effectiveness. In so doing I had
to explore for myself just how adequate this framework was: Could it
really do all the things claimed for it? Readers are entitled to their own
answers, but I became more and more convinced of the power of Cl archi¬
tecture as I continued to explore new issues in discourse comprehension
within that framework. In fact, I came to realize that the framework I pro¬
posed applies not only to discourse comprehension but also to a broader
range of issues in the study of cognition in general.
Thus, my own goals changed as I engaged more deeply with the
research for this book. I realized that not only discourse comprehension
but also other cognitive processes may be viewed from the vantage point
of the Cl model. I became interested in just how far the domain of the
comprehension paradigm extends. Comprehension is modeled as a con¬
straint satisfaction process in the Cl model. It is apparent now that some
other cognitive processes, such as action planning and decision making,
can be modeled in the same way. Hence, the title of this book, which sug¬
gests that comprehension can serve as a theoretical paradigm for a wider
range of cognitive processes. It does not encompass all of cognition -
there are truly analytic thinking processes that are beyond the range of
the Cl model. Such is the case even in the discourse area itself. For
instance, certain kinds of inference processes require conscious, active
problem solving rather than mere constraint satisfaction. But on the
other hand, I was able to show that most of what goes under the label of
inference in discourse comprehension can in fact be accounted for within
the proposed framework.
Writing this book thus became a quest to determine the limits of the
comprehension paradigm. I am still not quite sure where these limits lie,
except that they are farther out than most psychologists had assumed.
Further study will be required to explore in depth some of the sugges¬
tions made here, but I think that the usefulness of the comprehension
paradigm for the study of cognitive processes has been established.
 Acknowledgments

The research reported here could not have been performed without the
long-term support afforded by a Merit Award from the National Institute
of Mental Health, grant MH-15872. Additional support has been
received in the past ten years for various special projects: from the
National Science Foundation (with James G. Greeno) for the work on
word arithmetic problems, from the National Science Foundation (with
Gerhard Fischer, Clayton Lewis, and Peter G. Poison) and the Army
Research Institute (with Fischer) for the research on action planning,
from the Mellon Foundation (with Eileen Kintsch) for the research on
learning from texts, and from the McDonnell Foundation (with Fischer
and Thomas K. Landauer), and from DARPA (with Landauer) for the
exploration of the Latent Semantic Analysis technique. I am deeply
grateful to these institutions for permitting us to develop and evaluate the
ideas about discourse comprehension that eventually led to this book.
The University of Colorado has been blessed with an exceptional
group of graduate students, several generations of whom participated in
this endeavor. Their dedication and enthusiasm, experimental and mod¬
eling skills, but above all their fresh ideas and infectious inquisitiveness,
have profoundly shaped this research program. It is one the great plea¬
sures of being a professor to be able to collaborate with an ever changing
but eternally young team of the brightest and the best. Among the stu¬
dents and post-docs who were associated most closely with the work
described here are, in roughly chronological order: Randy Fletcher,
Chuck Weaver, Ernie Mross, Susan Zimny, Suzanne Mannes, Walter Per-
rig, Denise Cummins, Kurt Reusser, Franz Schmalhofer, Stephanie
Doane, Jose Otero, Isabelle Tapiero, Mitch Nathan, Danielle McNamara,
xvi Acknowledgments

Evelyn Ferstl, Marita Franzke, Julia Moravcsik, Mike Wolfe, Dave Stein-
hart, and Missy Schreiner.
The Institute of Cognitive Science and the Department of Psychology
at the University of Colorado has provided an ideal environment for my
work. I am deeply indebted to my colleagues for many stimulating dis¬
cussions and arguments. I could not have worked without their feedback,
whether critical or encouraging. Over the years, Peter Poison has been a
most valuable partner in this respect. I also want to mention two specific
collaborations that had a significant influence on the work presented in
this book. My former colleague Anders Ericsson and I became intrigued
with the functioning of working memory, which resulted in the theory of
long-term working memory - ideas that put a whole new twist on our
understanding of discourse comprehension. Last, I gratefully acknowl¬
edge my debt to Tom Landauer who introduced me to latent semantic
analysis (LSA) and who has been my companion in exploring this excit¬
ing new technique.
My intellectual debts are owed not only to my collaborators at Col¬
orado but also to a much broader group of researchers in psychology, cog¬
nitive science, and education. I shall not try to list them here, but I want
to express my appreciation and my hope that in the pages that follow I
always have given credit where credit is due.
Special thanks, however, go to those who read and criticized one or
more of the chapters of this book: Anders Ericsson, Tom Landauer, Akira
Miyake, Peter Poison, Paul van den Broek, and Sashank Varma. I also
thank Julia Hough of Cambridge University Press for her interest and
support.
This book is dedicated to my wife and co-worker, Eileen Kintsch. She
has been my supporter, research collaborator, discussion partner, critic,
and editor — not only on this project but throughout many wonderful
joint years. This book would certainly not exist without her. I only hope
that she and the many others who helped me with this book will be happy
with what it has become.

Walter Kintsch
 1

Introduction

The study of thinking has been characterized by a multitude of different

approaches, first in philosophy and the arts, later in science. Many scien¬
tific disciplines today are concerned in one way or another with the study
of thinking, or higher-level cognition, the somewhat fuzzy term currently
preferred in cognitive science and cognitive psychology.
If one looks at books on thinking in those areas, they are most fre¬
quently compendiums of observations that people have made about the
phenomena of thinking, both anecdotal observations and laboratory inves¬
tigations, descriptive as well as prescriptive, interspersed with some more
or less formal and more or less speculative theories focused on particular
phenomena. During the last 100 years, scientists have mostly left global
theorizing about human cognition to philosophers. On the whole that has
proved to be a wise choice, in that this focus on the specific has helped us
to acquire much solid information about cognition in that time, unham¬
pered by too many false theoretical starts. Nevertheless, even good data
are not totally satisfying if they are not tied together within some theo¬
retical framework, and in recent years interest in global frameworks for
cognition - architectures of cognition - has resurfaced.
Indeed, Newell (1990) has argued that we might miss something essen¬
tial if we focus too much on the facts and unrelated minitheories. Expla¬
nations of complex cognitive processes have too many degrees of free¬
dom. That is, there are too many ways to explain a local phenomenon, so
that it is not possible to tell which one is right or best. One way out of this
dilemma is to develop a single theory that explains many different local
phenomena, preferably all the phenomena of cognition, in the same way.
That is, we look for a general architecture that would allow us to develop
principled models of particular phenomena in such a way that the indi-
2 Introduction

vidual models constrain and support one another. Newell’s own work and
the research program of J. R. Anderson are stellar examples of this
approach to cognitive science (Anderson, 1983, 1993; Newell, 1990).
Comprehension may be another paradigm for cognition, providing us
with a fairly general though perhaps not all-encompassing framework
within which cognitive phenomena can be explained. Since Newell and
Simon’s groundbreaking book, Human Problem Solving (1972), problem
solving has been the paradigm for the higher cognitive processes, in
particular these authors’ conceptualization of problem solving as search
in a problem space. The task of the problem solver is conceived of as find¬
ing a solution path in a large, complex problem space full of dead ends.
This has been an extremely successful way of thinking about human cog¬
nition. But it may be that for at least some cognitive processes there are
alternative conceptualizations that could be fruitfully explored. For
instance, one can look at cognition, at least certain forms of cognition,
as a constraint-satisfaction, or comprehension, process instead. In this
book, I work out a constraint-satisfaction theory of comprehension and
show how this kind of cognitive architecture can serve as a paradigm for
much of cognition.

1.1 Understanding and comprehension

The terms understanding and comprehension are not scientific terms but
are commonsense expressions. As with other such expressions, their
meaning is fuzzy and imprecise. To use them in a scientific discourse, we
need to specify them more precisely without, however, doing undue vio¬
lence to common usage. First, understanding and comprehension are
used as synonyms here. The choice of one or the other term is thus purely
a matter of linguistic variation.
But just what do we mean by either understanding or comprehension?
What seems most helpful here is to contrast understanding with percep¬
tion on the one hand and with problem solving on the other. In ordinary
usage, perceive is used for simple or isolated instances of perception, espe¬
cially when no specific action is involved. Understand is used when the
relationship between some object and its context is at issue or when action
is required. 1 he boundaries between the two terms are certainly fuzzy,
however Understanding is clearly the preferred term when a perception
involves language.
 Introduction 3

Both perception and understanding can each be described as a process

of constraint satisfaction. Given a certain environmental configuration
and a particular state of an organism, including its knowledge and goals,
the organism responds with either an overt action in the environment or
a mental event that in turn may be related to an action. The features of the
environment and relevant organismic factors are processed in parallel in
an automatic way, with relatively little demand on resources. Although
the end result is conscious, the processes themselves are not. Problem
solving, or thinking, in contrast, is typically under conscious control and
is resource demanding and sequential. Constraint satisfaction is still a
requirement, but the constraints are not directly given in the environ¬
ment and by the organism; they must first be generated through possibly
complex procedures. Perception and understanding are the processes
people normally use; when an impasse develops in perception or under¬
standing, they resort to problem solving as a repair process.
We see this process occurring in concurrent verbal reports during
reading. The reports yield little information beyond the actual content of
the text being read, as long as reading proceeds normally. When it breaks
down, however, because a reader does not understand something, rich
verbal reports are obtained about the problem-solving processes needed
to repair the impasse (Ericsson & Simon, 1993).

1.2 A paradigm for cognition

We have thus defined comprehension by pointing to phenomena that

people commonly label as such. It makes little difference that some peo¬
ple use the term comprehension in slightly different ways than others do.
Indeed, by the end of this book, the reader might be even less certain just
where comprehension starts and where it ends, for I argue that compre¬
hension provides a good paradigm for areas of cognition that have not tra¬
ditionally been viewed from this vantage point (e.g., in chapter 11, action
planning, which is usually treated as problem solving). The paradigm I
am speaking of is not the fuzzy commonsense notion of comprehension
but the theory of comprehension that is developed in the following pages.
What are the distinctive features of this theory? First, it is a psycho¬
logical process theory. That is, it is concerned with the mental processes
involved in acts of comprehension - not primarily with the analysis of the
material that is to be comprehended. Applied to text comprehension, this
4 Introduction

means that it is not a theory of text structure, or a text analysis. The text
structure is only indirectly important, in that it is one determinant of the
comprehension process and therefore of the product of this process, the
mental representation of the text and actions based on this construction.
Second, at the most general level the theory characterizes the compre¬
hension process as one of constraint satisfaction. Comprehension occurs
when and if the elements that enter into the process achieve a stable state
in which the majority of elements are meaningfully related to one another
and other elements that do not fit the pattern of the majority are sup¬
pressed. The theory thus must specify what these elements are and how
they reach a stable configuration.
All kinds of elements enter into the comprehension process. In com-
monsense terms, these may be perceptions, concepts, ideas, images, or
emotions. We need a way to deal with all these in the theory. A proposi¬
tional representation will be described that provides a common notation
for these elementary units of the comprehension process and for the
description of the relations among them.
A crucial consideration is where these elements come from: from the
world via the perceptual system, as well as from the organism in the form
of memories, knowledge, beliefs, body states, or goals. At the heart of the
theory is a specific mechanism that describes how elements from these
two sources are combined into a stable mental product in the process of
comprehension.
Roughly, the story goes like this. We start with a comprehender who
has specific goals, a given background of knowledge and experience, and
a given perceptual situation. The perceptual situation may, for instance,
be the printed words on a page of text. We mostly skip the question of
how the reader forms basic idea units from these words (though we deal
extensively with word identification in a discourse context and, at least
tangentially, with the question of how sentences are parsed into their con¬
stituents). Given these idea units in the form of propositions as well as
the reader’s goals, associated elements from the reader’s long-term mem¬
ory (knowledge, experience) are retrieved to form an interrelated network
together with the already existing perceptual elements. Because this
retrieval is entirely a bottom-up process, unguided by the larger dis¬
course context, the nascent network will contain both relevant and irrel¬
evant items. Spreading activation around this network until the pattern of
activation stabilizes works as a constraint-satisfaction process, selectively
 Introduction 5

activating those elements that fit together or are somehow related and
deactivating the rest. Hence, the name of the theory, the construction-
integration (Cl) theory: A context-insensitive construction process is fol¬
lowed by a constraint—satisfaction, or integration, process that yields if all
goes well, an orderly mental structure out of initial chaos.

1.3 The architecture of cognition

The theory presented here is a proposal for an architecture of cognition.

That is, it is a collection of specific models, all employing the same general
architectural framework. Thus, it is quite possible to develop alternative
models within that framework, that is, models that differ in their specific
implementations but that share a common architecture. Hence, a model
may be inadequate not because of its basic architecture but for the way
it has been realized within that architecture, for every model involves
assumptions that are situation-specific and not dictated by the basic cogni¬
tive architecture. On the other hand, the architecture itself is not directly
testable. Only cumulative experience with many different models involv¬
ing a broad range of psychological phenomena allows us to determine the
usefulness of a cognitive architecture.
The features that are most characteristic of the architecture that is
proposed and explored here are (1) assumptions about the mental repre¬
sentation of texts and knowledge and (2) processing assumptions about
comprehension. Both texts and knowledge are represented as networks of
propositions (or, loosely equivalent, semantic vectors). Such associative
networks are relatively unstructured and hence differ considerably from
assumptions about the representation of meaning in logic or in frame¬
works based on formal semantics. The mind in this view is not a well-
structured, orderly system but is a little chaotic, being based on percep¬
tion and experience rather than on logic, being Aristotelian rather than
Cartesian. The assumptions made about cognitive processes are similar:
The construction of mental representations does not involve the applica¬
tion of precise, sophisticated, and context-sensitive rules; instead, con¬
struction rules may be crude and relatively context-free and may yield
only approximate solutions full of irrelevancies and redundancies that
need to be cleared up by constraint satisfaction - specifically, a spreading
activation mechanism.
In recent years, since the pioneering work of Marr (1982), we have
6 Introduction

learned a great deal about the different levels at which information

processing theories can be developed (Anderson, 1990, Newell, 1990,
Pylyshyn, 1981). Marr and his followers, always using a new and differ¬
ent terminology, distinguished three levels of psychological theory, the
middle level being once more subdivided into two sublevels. T he most
general level of analysis is at the rational level (I follow Anderson’s termi¬
nology). An information-processing theory at the rational level must ask
what the goals of the computation are and must analyze how these goals
can be carried out. The lowest level of analysis is the biological level,
which consists of an account of cognition in terms of the organism s
underlying brain processes. In between these two extremes is the more
typical psychological information-processing theory that is concerned
with the nature of mental representations and the computations that are
performed on these representations. At this level, the theorist is con¬
cerned with cognitive algorithms and their implementation. A cognitive
algorithm is a specification of all the steps in a cognitive process, the states
of the system at each point in time, and their transformations. The imple¬
mentation is a particular mechanism that achieves these transformations.
A cognitive architecture can be described as a recipe for constructing
implementations of a cognitive algorithm. In the case of the Cl architec¬
ture, it requires the theorist to use simple, bottom-up rules to construct
a preliminary but incoherent propositional network, followed by a
spreading activation process that integrates this network into a coherent
mental representation.
The problem for information-processing theories has been and con¬
tinues to be the identifiability of the mechanisms postulated. There are
many ways to achieve a goal, and more to implement them. How can we
be sure the one we happen to think of is the right one? It may explain the
data, but so might many other mechanisms. As Anderson (1990) points
out, this is a particular problem at the implementation level. At the algo¬
rithm level, there are at least behavioral data that can be directly com¬
pared to the proposed steps of the algorithm, but the implementational
mechanisms of our theories can be inferred only from the behavioral
observations at the algorithm level. For instance, we may be able to deter¬
mine that a given model correctly predicts whether readers make a cer¬
tain type of inference on line, but the mechanism that produced the infer¬
ence may be only partially constrained by our observations.
Psychologists have tried various ways to deal with this dilemma,
 Introduction 7

including ignoring it. An obvious solution is to look to the biological level

for the constraints that would disambiguate information-processing the¬
ories. Indeed, in some areas such as mental imagery and attention, some¬
thing like this is already happening. Perhaps the rapidly improving meth¬
ods of cognitive neuroscience will soon also yield information about such
complex processes as reading comprehension and language understand¬
ing.. However, it is not obvious that such primarily culturally determined
processes are sufficiently distinct at the biological level to warrant their
study at this level.
The rational level of analysis provides another set of constraints that
are very important for information-processing theories, as both Newell
and Anderson have argued and demonstrated in their work. Thus, the
logical requirements of information retrieval can serve to constrain theo¬
ries of memory retrieval. Certainly, information-processing theories need
to find help wherever they can get it. But one need not give up hope of
solving the identifiability problem at the architecture level. Although a
particular model may not be identifiable, it may be possible to design a
cognitive architecture capable of accounting for a wide variety of behav¬
ioral data over the whole range of cognitive psychology. If the same archi¬
tecture explains many different phenomena, both simple and complex, it
becomes much harder to think of plausible alternatives. Or at least there
may be only a few architectures that are reasonable candidates. Newell
(1990) has most convincingly made this argument in his call for unified
theories of cognition. Information-processing theories of isolated cogni¬
tive phenomena, whether simple or complex, are of much less impor¬
tance, however elegant and successful they may be empirically, than the¬
ories that are able to account for a broad range of phenomena with the
same set of principles.
I see cognitive architectures as languages that scientists develop to talk
about (describe, predict, postdict) cognitive phenomena. Some theorists
may aspire to a loftier epistemological status for their theories than just a
useful and convenient language. It is not a goal to be despised, however,
and it is perhaps all we can do for the time being.

1.4 The goals and scope of the book

The research reported in this book focuses on text comprehension.

Although I do not neglect comprehension processes in general, most of
8 Introduction

the empirical work and modeling have been done in the area of text com
prehension. Furthermore, the theory of comprehension presented here
is a computational theory. It describes a sequence of steps that, given cer¬
tain initial conditions, yields outcomes comparable to human compre¬
hension. Because computation presupposes a representation on which
computations can be performed, the concept of mental representation
and the role that mental representations play in cognition is discussed in
the second chapter. Parts' of this chapter are highly speculative, and a
wiser author may have refrained from touching on these issues. It is
important, however, that we attempt to place our work within a broader
framework.
In the third chapter, propositional representations are introduced and
justified. The fourth chapter discusses the computations performed on
these propositional representations that form the basis of text compre¬
hension. This first section of the book concludes with a discussion of the
processing theory, the construction-integration model.
The second part of the book consists of seven chapters touching on
many of the topics that are typically discussed in a cognitive psychology
text. How are word meanings identified in a discourse context, including
anaphora and metaphors? How are words combined to form coherent rep¬
resentations of texts, at both the local and global level? How do texts and
the mental models readers construct from them represent situations?
What is the role of working memory in comprehension? How is relevant
knowledge activated during reading, and how is the information provided
by a text integrated with a reader’s knowledge? How do readers recognize
and recall texts? What is the distinction between remembering a text and
learning from a text, and what principles govern remembering and learn¬
ing? What are the implications of these findings for how people solve word
problems, how they execute verbal instructions, and how they make deci¬
sions based on verbal information? Thus, the book is concerned with
attention and pattern recognition, knowledge representations, working
memory, recognition and recall, learning, problem solving, and judg¬
ment- almost the complete range of topics in cognitive psychology. I have
tried to discover for myself and to show others how comprehension theory
can be applied to these phenomena. Some of this work has been published
in scholarly journals or book chapters before; much of it is presented here
for the first time. But even when I describe previously published work, it
 Introduction 9

is set here into a general framework that did not exist earlier or that could
not be as fully explicated as I have done here.
Some of the work reported in these chapters consists of large-scale,
carefully executed, and well-documented experimental studies and theo¬
retical simulations. But I have not resisted the temptation to include pre¬
liminary results and simple examples illustrating how the theory could be
applied. Such illustrations are like Gedankenexpenments exploring the
implications of the comprehension theory. They are no substitute for sys¬
tematic experimentation, theoretical proof, or large-scale simulation. Yet
we cannot do without them if we wish to explore the full scope of a the¬
ory, and although the reader may not want to take them as ultimate evi¬
dence, they may seduce the reader with their promise. Thus, many results
are presented on the following pages, some with more assurance than oth¬
ers. If these pages convince the reader that comprehension is a useful par¬
adigm for cognition, a conceptual framework will have been provided for
use in following up on the issues that have been treated here only super¬
ficially, as well as for exploring the even more numerous issues that have
not even been considered here at all.
 2

Cognition and representation

Cognitive science and cognitive psychology focus on cognition and per¬

force neglect other aspects of human behavior. Nevertheless, we must ask
where cognitive research is situated in the study of the human mind and
human behavior. I argue here, and also in the last chapter of this book
where I return to these arguments in a somewhat different way, that an all
too narrow focus on cognition places intolerable restrictions on cognitive
science and that progress beyond a certain point depends on our ability
to redintegrate the cognitive and emotional-motivational aspects of
human behavior. It is imperative to begin with a clear idea of the range of
psychological phenomena that must be considered and an understanding
of how they fit together.
Thinking, comprehending, and perceiving presuppose representa¬
tion - of some kind. The theory of text comprehension that is the primary
concern of this book mostly assumes that these representations are
propositional (a related vector representation is also used on some occa¬
sions). In the first section of this chapter, I discuss the variety of mental
representations that appear to play a role in psychology in general and
cognition in particular. In the second section, I argue that in spite of the
significance of multiple forms of representations, a theory of text com¬
prehension that relies primarily on propositional representations is nev¬
ertheless feasible.

2.1 Multiple representations:

How the mind represents the world

Geometry was one thing and algebra was quite another before
Descartes discovered the isomorphism between these two fields:
Part I

The theory
14 The theory

Algebraic expressions can be represented as geometric figures

and vice versa. For instance, a circle that is centered at the origin
with radius r = 2 can be represented algebraically by the equa¬
tion x2 + y2 = 22.
What is gained by this transformation? For some purposes, it
does not matter much. We can see in Figure 2.1 that the circle has
been displaced from the origin (0,0) to a point in the upper right
quadrant (4,2): We can express this change of location by the
equation (x + 4)2 + (y + 2)2 = 4, but the algebraic representation
is not obviously superior to the geometric. The picture is as good
as or better than the equation in this case. But consider a differ¬
ent scenario. Suppose we want to find a square centered at the
origin that has the same area as the circle. This is notoriously dif¬
ficult to do in the picture domain but quite trivial in the algebraic
representation: The area of the circle is r2 * 7t; therefore, the side
of a square with the same area must be (22 * n)]/2. Once we have
calculated this value, we can translate it back into the picture
domain and draw the square as was done in Figure 2.1
Some computations are easy to perform in one domain but dif¬
ficult or impossible in another. Mathematics has been enormously
enriched by the possibility of representing an object from one
domain in another domain and operating on it with procedures
peculiar to that domain. Cognition employs the same trick.

Objects and events in the environment are characterized by certain

properties and relations. Perception, comprehension, and problem solv¬
ing generate mental models of the environmental objects and events, and
operate on these models. The cognitive system transforms the original
structures, merging the current environment with the organism’s previ¬
ous experience of it. The mental models generated in this way are iso¬
morphic to the environmental structures and hence provide a basis for
the interaction of an organism with the environment.
What is achieved by these representations that could not be achieved
otherwise? Much as the possibility of writing an equation instead of
drawing a circle permits the mathematician to use operations that could
otherwise not be used (the operations of algebra), representing the
human environment in a mental model allows us to make mental compu-
 Cognition and representation 15

Figure 2.1

tations that are not possible in the actual environment. The environment
rarely provides all the information necessary for an action. It is usually
the case that human knowledge and experience are required in coordina¬
tion with the environment to guide action. The knowledge representation
makes this possible. At a minimum, it supplements information in the
environment by filling in gaps that are unspecified. Often, however, the
environmental input must be transformed in complex ways to ensure an
optimal action.

2.1.1 Types of mental representation

A definitive account of mental representations does not yet exist, nor is

one offered here. Instead, I briefly indicate my own views on this issue,
and then discuss related or complementary positions in cognitive science
and describe some of the empirical evidence for these positions, empha¬
sizing the diversity of mental representations and their interrelations.
 16 The theory

One can think of mental representations as forming a hierarchy of

abstractness and increasing independence from the environment. The
most basic forms of mental representations are procedural and percep¬
tual representations that are tied directly to the environment. Episodic
memory representations are at the next higher level. Intentional but non¬
verbal representations, including forms of imagery, make up the next
layer of mental representations. The final two layers achieve the most
independence from the environment and are both verbal: the verbal nar¬
rative level and the verbal abstract level. The hierarchy is not one of com¬
plexity, however: There are highly complex perceptual representations
and very simple abstract representations. The defining feature of the
hierarchy is that it changes from direct representations of the environ¬
ment to ever more indirect, flexible ones that permit more and more arbi¬
trary, unconstrained computations. The layers of the hierarchy are
ordered in terms of their appearance on the mental stage, both phyloge-
netically and ontogenetically. As new forms of mental representations
arise, the older forms do not disappear but remain embedded within the
newer layer. Thus, the picture is one of gradual unfolding of the full
capacity of the human mind. It is a picture the details of which are still
quite vague, and as these details become better understood, some of the
picture s features may have to be corrected. But the implications of the
framework suggested here for cognitive theory remain relevant, even if
the framework is altered in detail.
As one moves from the bottom to the top layers of mental representa¬
tion, the general character of the representations changes in the follow¬
ing ways.

Most significantly, the degree of environmental control weakens; sto¬

ries and mathematical theories liberate man from the web of environ¬
mental dependencies.
Representations change from sensorimotor and analog in the lower lay¬
ers to symbolic and arbitrary in the upper layers.1 In between, an image
01 a gesture may be used in a symbolic but nonarbitrary way.

At the same time, the degree of consciousness increases. Similarly, the

1 The termy symbol has at least two senses: (1) a symbol is defined by its referent, and
(Z) a symbol is abstracted from its referent and defined by its conceptual role I am
using the term here in its second sense.
 Cognition and representation 17

degree of intentionality increases. Thus, we have at the one extreme

procedural representations that are acquired through incidental,
unconscious learning, and at the other extreme formal school learning
that requires intentional and conscious acts.

Direct procedural and perceptual representations. These forms of repre¬

sentations involve largely innate systems. Many different types of affor-
dances, abilities, and actions exist, as well as different biological mecha¬
nisms. These forms of representation can be modified through experience.
Responses to environmental affordances can be modified by perceptual or
procedural learning. Such learning is tightly coupled to the environment
(e.g., how to tie a shoelace). Repetition and reinforcement determine the
learning process within constraints imposed by the abilities of the organ¬
ism and the affordances of the environment. This capacity is shared by all
animals.

Episodic representations. Cognition at this level is based on episodic

memory representations, that is, generalized event representations of
experience that are created to guide action and anticipate changes in the
environment. Event memory, unlike procedural memory, is accessible to
recall and reflection, because it is a form of declarative memory. Out of
event memory emerges the recollection of particular experiences, involv¬
ing a certain level of consciousness and self-awareness. It permits the
analysis and breakdown of perceptual events. Event memory is shared
with higher animals. What is remembered and represented in memory
are concrete events and scriptlike sequences of events. Learning occurs
through experience and is incidental, unintentional, and goal-directed.
Organisms, such as apes, that must rely on event memory because they
do not have higher forms of cognition, use signs and have a repertoire of
social skills that depends on a rich episodic memory, but their actions are
directly linked to the environment. Thus, cognition at this level is analytic
and reflective but still environmentally bound. Human event memory is
different from the “pure” event memory of apes, because humans can
code their memories linguistically. Human event memory is embedded in
other forms of representations. As a consequence, human event memory
is tied in with higher cognition, that is, with linguistic and symbolic
thought.
18 The theory

Nonverbal, imagery, and action representations. These representations

are sensorimotor in character but are used intentionally, often but not
necessarily for the purpose of communication, especially the communi¬
cation of emotions (e.g., body language). An established social commu¬
nity is a prerequisite for representations of this type.

Narrative oral representations. These representations are one of two

types of linguistic representation. They are verbal but not abstract. Their
structure is linear, and information processing at this level is analytic and
rule-governed, as in semantic memory, propositional memory, discourse
comprehension, analytic thought, induction, and verification.
Much of what we know and what we learn is in the form of stories - for
example, our cultural and historical knowledge. Stories are narrative
mental models that allow us to learn about the world. The world becomes
more comprehensible to us when we are able to tell a coherent story about
it. There is again a social component to narrative learning because stories
are told by someone to someone (including to oneself). Socially elabo¬
rated and sanctioned stories are the cognitive structures that hold a cul¬
ture together.

Abstract representations. Abstract representations are required for cate¬

gories, logical thought, formal argument, deduction, quantification, and
formal measurement. Abstract symbols are dependent on visuographic
invention: written language, maps, calendars, clocks, artistic and scien¬
tific graphing, and other forms of external memory storage. Knowledge
is primarily stored in the world, not in the individual brain. Biological
memory carries around the code for the use of external memory, whereas
the specifics are found in external symbolic storage systems. Pedagogy
has always been directed at this abstract level. This is where learning
problems arise and where most special instructional efforts are needed.
This is what school learning is about.
To illustrate these distinctions with an example, consider the way peo¬
ple assign meaning to one object by transferring to it aspects of meaning
from another. At the nonverbal level, an example is children’s pretense
play: A banana becomes a “telephone” by the child’s talking into it, as
adults do with a telephone. An example at the narrative linguistic level is
metaphor: One concept assumes certain semantic features from another.
 Cognition and representation 19

For example, sermons assume some properties of sleeping pills. Analogy

is the corresponding example at the abstract linguistic level: Objects are
no longer directly compared, but the meaning transfer depends on an
abstract categorical system (the analogy between electricity and water
flowing through pipes requires a precise mapping of categories).
The neat classification presented here is complicated by the fact that,
in the adult human, lower forms of cognition are encapsulated by the
higher forms. The types of cognition we have discussed are found in pure
form in the phylogenetic development of man or in the ontogenetic
development of cognition, but not in the adult human mind. Thus, peo¬
ple have incidental event memory, as a dog might have, but they also have
explicit, language-coded event memory that is uniquely human. Even
procedural memory may be language-coded, though not very well, as is
testified by the inefficiency of verbally describing a tennis stroke or a ski¬
ing turn. Action representations, too, are coded with language. Thus, a
modern apprenticeship situation is not purely imitational but supple¬
mented by oral language and symbolic thought, as, for example, a gradu¬
ate student doing work in some laboratory learns skills.
Conversely, the higher layers of representation that characterize the
modern adult human mind do not exist in a symbolic vacuum but are
based on the sensorimotor substratum that humans share with other
forms of animal life. Hence, language cannot be understood only as an
arbitrary disembodied linguistic system and cannot be separated from its
nonlinguistic substratum.

2.1.2 The evidence for levels of representation

To keep it brief and focused, no references have been provided in the

foregoing discussion, but there actually exists considerable evidence sup¬
porting a scheme like the one suggested. Indeed, that scheme is culled
quite directly from the existing literature. One problem with that litera¬
ture is that everyone employs a different terminology, which creates a
great deal of confusion, and which is why I choose to disregard so far the
literature in this discussion.
Despite differences in terminology, there exists agreement among
widely varying sources for distinguishing two basic types of representa¬
tion: a habit system and a cognitive system. The evidence comes from
20 The theory

behaviorism, animal learning, perception, situated cognition, and cogni¬

tive neuroscience. I do not claim that all these people talk about exactly
the same things, and if they do, they certainly do so with very different
words, but there are some commonalities worth considering.

Behaviorism: S-R connections and mediated responses. The behaviorists

originally had no use for mental representations, but soon the mind crept
back into behavioral theories in the form of mediating responses. Whereas
orthodox behaviorism (Skinner, 1938) was interested in only the func¬
tional relationships that hold between various stimulus classes, S, and
responses, R, most of the great theorists of behaviorism in the 1930s and
’40s acknowledged the necessity of inserting a mediating representational
level between the S and the R, the mediating response, r. Thus, the
behaviorist paradigm changed from S-^RtoS-^r^R (Hull, 1943;
Spence, 1956; Tolman, 1932; for a modem account, see Kimble, 1996).
Little-r is clearly a kind of mental representation, but it is far from a full-
fledged symbolic representation. For Hull and Spence little-r was still a
response, at least potentially observable. A great deal of research effort
was in fact devoted to measuring and observing little-r. For instance,
Witte and I (Kintsch & Witte, 1962) spent many months observing how-
conditioned salivary responses develop in dogs that are learning an
instrumental response - bar pressing for a food reward. We were able to
show a lawful relationship between the act of bar pressing (R) and the
mediating salivary response (r), depending on the schedule of reinforce¬
ment. It is not too strong to say that our dogs “thought” with their sali¬
vary glands! I argue later that similar concrete representations play a role
in human cognition as well.
The main limitation of the mediating response concept is its simplic¬
ity. The kinds of computations people perform with their mental repre¬
sentations are much more complex than the mediating response concept
allows for. Unlike dogs, people have language and symbolic thought,
requiring mental representations at a different level of complexity.

Animal learning: Direct and indirect representations. Gallistel (1990)

brings an ecological and biological view to the study of animal learning. A
basic distinction in his approach is that between direct and indirect rep¬
resentations. Gallistel uses the term direct representation when others
would say that no representation is involved at all. This is because he
 Cognition and representation 21

wants to emphasize that even in the simplest cases of learning, some

change in the organism is occurring, that is, some representation of the
environment is formed in the organism’s neural system. Originally, a
stimulus S elicits a response Rj, but as a consequence of learning the
organism is changed (has formed a “direct” representation), so that S
now elicits R2. The direct representation does not involve a mediating
levej — it cannot be used for computational purposes. It is still direct
action, tied to the environment. Only indirect representations break up
the S-R link, providing the power to plan and prepare and to control the
environment.
Gallistel’s approach is explicitly representational. He is concerned
with the “formal structure of the represented system and the formal struc¬
ture of the representing system that enables one to predict results in the
represented system on the basis of operations conducted within the
representing system” (Gallistel, 1990, p. 582, italics in original). Among
the learning processes he has analyzed, some can be described as the
acquisition of direct isomorphisms between the environment and the
organism’s representation thereof, whereas others require indirect repre¬
sentations. Classical conditioning, for instance, involves the formation of
indirect mental representations, according to Gallistel. It is not a direct
associative process but reflects the operation of higher-order processes
and involves the computation of temporal intervals, rates, and statistical
uncertainties.

Perception. Gibson’s pioneering work on direct perception (Gibson,

1977) has recently been extended by Neisser to include both direct and
representation-based perception. Neisser (1994, p. 228) has elaborated
his position to include three perceptual modules. Two modules are direct
systems and one is representational, as follows:

1. Direct perception/action, which enables us to perceive and act effectively

on the local environment.
2. Interpersonal perception/reactivity, which underlies our immediate
social interactions with other human beings.
3. Representation/recognition, by which we identify and respond appropri¬
ately to familiar objects and situations.

Module 1 is the field first defined by J. J. Gibson (1977); module 2 has

been the domain of social and developmental psychologists; and module
22 The theory

3 is what information-processing psychology and cognitive science have

focused on.
Direct perception (module 1) links the organism to the surrounding
optic array. The full optic array includes movement-produced informa¬
tion and is highly redundant, forming a single, mathematically tightly
constrained complex. Any part of that complex allows the reconstruction
of the whole. The perceiver’s location and movement are a central com¬
ponent of that complex. The organism perceives not only what is but what
might be done: “Every purposive action begins with perceiving an affor-
dance” (Neisser, 1994, p. 235). The knowledge obtained by direct per¬
ception is immediate, bottom-up, and cognitively impenetrable; it does
not have to be constructed.
Interpersonal perception (module 2) is similar in these respects, but its
object is not the relationship between self and the environment but the
relationship between self and another person. It is highly interactive.
Neisser cites an experiment by Murray and Trevarthen (1985) in which
babies were observed to interact successfully with their mothers over
closed-circuit television, as long as the mother directly responded to the
baby. In a noninteractive condition, where the baby was shown the
rewound tape of its mother, the babies quickly became distressed. What
they had enjoyed before apparently was the coordinated interaction
between themselves and their mother, not the mother herself.
Interpersonal perception always engages emotion. Mother and baby,
or for that matter any human partners, form a dynamic, finely attuned
affective system. The continuous flow of emotions in interpersonal inter¬
actions, but undoubtedly also in less direct interactions, such as occur in
reading a story, functions as a modulator and motivator of cognition.
The representation/recognition system is the third module of percep¬
tion described by Neisser. It ranges from classical conditioning to pattern
recognition, language understanding, expertise, and problem solving.
Neisser points out that recognition is always dependent on past experi¬
ence, whereas direct perception and interpersonal perception are based
on currently existing situations. Recognition depends on individual
details and can be dissociated from direct perception. Neisser illustrates
this point with the famous Kohler studies of adaptation to inverting
prisms (Kohler, 1962): After 18 days of wearing left-right reversing
prisms, Kohler could ride his bicycle through town, but all letters and
number still looked backward to him. At that point, the direct perception
 Cognition and representation 23

system had returned to normal, allowing him to act normally in his envi¬
ronment, but the recognition system had not yet adapted.

Situativity: Direct action and symbolic representations. Another exten¬

sion of the Gibsonian system has been suggested by Greeno (1994), who
combines Gibsonian ideas on direct action with the situated cognition
research coming from anthropology (Lave, 1988; Suchman, 1987).
Properties of both the environment and the organism are relevant for
the analysis of the interaction between an organism and its environment.
Mental representations are needed to keep track of and compute the
implications of the constraints between environmental situations that an
organism has experienced.
Properties of the environment that determine the organism-environ¬
ment interaction are called affordances, after Gibson. Properties of the
organism that allow it to be attuned to the environmental affordances are
called abilities. The interaction between an organism and its environment
can be described in terms of affordances and abilities that reflect the con¬
straints existing in the environment. Affordances and abilities are there¬
fore inherently relational terms, one being defined in terms of the other.
Affordances and abilities yield smooth, efficient performance in a well-
attuned interactive system.
Consider, for example, a professor who has just finished his lecture.
The students are leaving. The professor gathers his notes that are dis¬
persed on the table, closes and unplugs the overhead projector, grabs the
notes and the projector, and walks out the door, all the time conversing
with a student who had asked for the clarification of a point that had not
been presented well in the lecture. The notes on the table, the projector,
and the open classroom door provide affordances to which the professor
responds without thinking. He directly understands the environment
and interacts with it, cleanly picking up the notes and projector and leav¬
ing the room, without needing to recognize his notes, projector, or the
door as the objects they are. He is certainly conscious of what he is doing,
but he is not thinking about it, he is just doing it. He perceives the door
and the floor, not as door or floor but simply as something to walk on or
walk through. It is of course possible for him to stop and say, “Aha, this is
a floor,” but normally he just walks, which leaves him with plenty of
resources to carry on his conversation with his student. This is direct
action. On the other hand, the conversation itself involves symbolic
24 The theory

representations. The professor must access his own memory of his lec¬
ture, realize just what confused the student, and reformulate his explana¬
tion in a more judicious manner. This is no longer a direct response to the
environment based on some well-established ability - but neither does it
require a deliberate act of problem solving. The professor “understands”
what the student is saying and why, and responds appropriately.
Greeno emphasizes the social origin of mental representations. People
do not think and act alone but as part of a social and cultural community.
Concepts evolve out of the discourse of communities of practitioners in
some particular domain. For instance, the concept of the turning radius
of a car was constructed in response to certain constraints experienced
by drivers and automotive engineers. It may be used quite differently by
different groups of people. It may be an implicit concept - affordance
plus ability of an experienced driver. Or it may be an explicit symbolic
concept - for example, for the driving instructor who must explain it to
students, or for the engineer who describes it with a mathematical for¬
mula. And if the engineer also drives a car, it may be an implicit and an
explicit concept at the same time.

Cognitive neuroscience: Habit and cognition systems. The idea that there
are two distinct psychological systems involved in cognition and learning,
direct action and representation, has attracted a great deal of attention
within cognitive neuroscience. Many patients with damage to certain
areas of the brain have been studied who are unable to recognize objects
but behave appropriately with respect to these objects (e.g., the Hindsight
phenomenon first reported by Weisskratz, Warrington, Sanders, & Mar¬
shall, 1974). Leibowitz and Post (1982) provide a succinct summary of
this research that supports a distinction between visual processes devoted
to object recognition on the one hand and orientation on the other. They
demonstrate this distinction by the ease with which it is possible to walk
while simultaneously reading. Attention is focused on the reading mater¬
ial, but locomotion in most environments is smooth and troublefree.
Focal vision is primarily involved in reading, whereas orientation engages
peripheral vision. Recognition is conscious, whereas the ambient func¬
tions usually operate without awareness. Evidence that supports these
common observations from both anatomical studies involving ablations
and perceptual experiments is reviewed by Leibowitz and Post.
 Cognition and representation 25

Mishkin and Petri (1984) and other neuropsychologists (Squire,

Knowlton, & Musen, 1993) distinguish between two learning systems: a
cognitive memory system and a habit system. The cognitive memory sys¬
tem stores information from past experience in locations where it can be
reactivated by new sensory inputs. The brain areas involved are the rhi¬
nal cortex, connecting to the thalamus and mamillary bodies, and from
there to the prefrontal cortex. The habit system is activated by sensory
input from the sensory processing systems to the caudate nucleus and to
the putamen, where the probabilities of particular stimulus-response
connections are stored and updated as a result of new experiences. Habit-
regulated behavior is obtained through the sequential activation of fur¬
ther structures in the brainstem that project to the ventral portions of the
thalamus and eventually to motor portions of the frontal cortex. The con¬
tent of this habit store are not neural representations of objects or events
but merely response tendencies varying in strength. Thus, the two sys¬
tems have fundamentally different learning and retention properties, use
different circuits in the brain, and store different aspects of experience.
Closely related dual-learning models have been proposed by Squire
(1992), who uses the terms declarative and nondeclarative, and by Graf
and Schacter (1985), who talk about explicit and implicit memory. Thus,
there are many observations in the field of neuroscience that point to the
same dichotomy between processes based on direct representations and
those based on symbolic representations.
The literature referred to is much too rich and complex to be discussed
here in detail. Instead I describe a prototypical study by Squire and
Knowlton (1995), which nicely makes the main points of interest. Squire
and Knowlton gave normal and amnesic subjects a probability learning
task. Subjects saw on each trial several cards with unusual designs and
had to predict the “weather” - sun or rain - on the basis of these cards.
Each card was associated in an arbitrary, probabilistic way with a predic¬
tion; for example, a double X on top of a card might predict sun 70%
of the time. The subjects looked at each display, made a prediction, and
then were told whether they were right or wrong. After a large number of
trials, they reached a performance level that was well above chance, as is
typical in such experiments. There were three interesting results in this
study. First, although the subjects responded well above chance, they
insisted that they were merely guessing and that they did not know what
26 The theory

they were doing; this is also typical for studies of this kind. Second, and
more surprisingly, normal and amnesic subjects performed about equalh
well. This is a striking result, because Squire and Knowlton’s amnesic
subjects had severe memory deficits. In the extreme case, such subjects
could remember nothing at all; even after many experimental sessions,
they did not remember ever having performed the experimental task -
although they had learned the reinforcement contingencies quite well, as
well as normal subjects. Third, once subjects reached their performance
asymptote on the prediction task, amnesic subjects remained at this level,
but normal subjects showed a small further improvement with continued
training. This improvement was correlated with verbalizations such as
“Now I understand that the double X on top is more likely to predict sun
than rain.”
The probability learning task did not involve the formation of mediat¬
ing mental representations. Learning was slow, based on reinforcement,
and understanding played no role. The brain areas that perform this
function were unaffected by the brain damage that destroyed the event
memory of the amnesic patients. Thus, although the amnesic patients
could not remember events such as having participated in this experi¬
ment before, their acquisition of a direct isomorphism (implicit learning,
habit formation) was unimpaired. The superior representational facilities
of normal subjects were of no use in this task - the experimenters clev¬
erly manipulated the environment in such a way that there was nothing
to learn except rote stimulus-response contingencies, thus depriving the
normal subjects of a chance to use their superior abilities. Only with pro¬
longed training did normal subjects start to reflect on the habit they had
acquired and to verbalize their experiences and formulate rules, which
the amnesics could not do. The task was rigged in such a way that reflec¬
tion and rule formulation were of little help, however. Nevertheless, it is
of great interest theoretically. It demonstrates that the lower habit system
is encapsulated within the higher cognitive system. Our habits are not
merely habits, as in a dog or an amnesic patient; we can reflect on them,
talk about them, mathematize them, and so on. In Squire and Knowlton’s
experiment, this ability was of little consequence, but in everyday life this
ability to link habits with language and symbolic thought can have pro¬
found consequences. It permits the higher cognitive systems a certain
amount of control over the lower ones. If one asks a tennis player to ana¬
lyze a certain stroke, he will be at a loss and ask to be allowed to demon-
 Cognition and representation 27

strate what he does; the tennis coach, however, has learned how to talk
about the player’s stroke, and the kinesiologist may be able to provide a
scientific analysis of it.
The fact that so many investigators in such different areas have found
it useful to distinguish two broad categories of mental representations
suggests that we are on fairly solid ground in differentiating a habit and a
cognitive system. In addition, there seems to be fairly general agreement
about the need to separate procedural representations from episodic memory.
The best evidence comes from studies of child development (Karmiloff-
Smith, 1992; Nelson, 1996). Procedural learning tends to be unconscious
and is fully under environmental control. Episodic memory may involve
different degrees of consciousness and provides an organism with the
earliest opportunities to weaken the direct link between action and envi¬
ronment.
The distinctions within the cognitive system that have been made here
are based partly on Bruner (1986), who argued for a distinction between
what he termed the narrative and paradigmatic forms of language use,
and partly on studies of cognitive development, both its ontogeny and
philogeny. In particular, it is the work of Donald (1991) on the evolution
of human cognition and the related work of Nelson (1996) on child devel¬
opment that have strongly influenced my thinking in these matters. I
briefly discuss their ideas, both as justification for the types of represen¬
tations assumed here and to further elaborate them.

Cognitive development: Phytogeny and ontogeny. Donald (1991) argues

that the kind of mental model of the world that an organism can construct
depends on its representational facilities. He describes a sequence of four
cultures, or ways of modeling the world, in the evolution of the modern
mind: episodic culture, mimetic culture, narrative culture, and theoretic
culture.

1. Episodic culture. All animals learn and have procedural memories. At

some point in evolution, certainly at the level of primates but perhaps ear¬
lier, animals reach a level of awareness that makes possible declarative
memory and certain limited mental representations. Episodic memory is
an event memory (rather than an object memory). Nelson (1996) calls
this the general event memory and shows that it is ontogenetically prior
to both specific episodic memory and semantic memory. Mental models
28 The theory

in the episodic culture are representations that are generalized records of

past experience. Such models suffice to guide actions in the present and
to plan for the future, although specific memory episodes are not differ¬
entiated in this model. The organism acquires a habit to react in a certain
situation in a certain way, which summarizes its past experience with that
situation, but the past experiences are not separately retrievable and no
abstract semantic categories are formed. As biological evolution contin¬
ues, a new representational system emerges allowing for the development
of:
2. Mimetic culture. Intentional imitation is the characteristic achieve¬
ment of this stage of development of the earliest humans. This imitation
is for an audience, that is, social communication is the driving force
behind it. Speech appears at this stage; its function is not representation
but social communication. The representation is through action, includ¬
ing speech action. Cultural achievements of this period are trades and
crafts, as well as games, ritual, and art. Individuals are self-aware (some¬
thing very rare in apes) and have a high degree of social intelligence, but
the representation-through-action limits the kind of mental representa¬
tions they can achieve.
At this point in human evolution, language begins to be the decisive
factor in the further evolution of mankind. Biological evolution is
replaced by cultural evolution. The sensorimotor apparatus of humans is
not very different from that of the other primates, but the possibility of
sharing knowledge, first by imitation then by the representational use of
language, makes humans unique.
3. Narrative culture. Narrative culture is an oral narrative culture, the
characteristic product of which is myth. Myth, in Donald’s catalogue, is
a kind of mental model that is fundamentally linguistic. As a story about
the world, this kind of representation is removed from the actual world
and hence is less dependent on it. More flexible computations can be per¬
formed with models of this kind. The human mind can be dissociated
from its environmental context to achieve symbolic control over the
world. Words and thoughts are like two sides of a coin and are insepara¬
ble. Most human cultural achievements are within the purview of narra¬
tive mental models.
4. Theoretic culture. This final stage is characterized by formal thought,
 Cognition and representation 29

abstract analyses and taxonomies, formal measurement, formal science,

and logic. Whereas the source of mythic culture is the oral narrative, the
source of theoretic culture lies in written language and external memory
devices. Thus, evolution has reached a technology-driven stage. Books,
records, maps, the establishment of libraries, all have made possible for¬
mal models of the world. The human mind expands; in addition to the
engrams of memory, there are now the exograms procured by our tech¬
nology (though both of them are dependent on the human mind,
Schonpflug & Esser, 1995). With external symbolic storage comes the
need for formal education; it is no longer enough to imitate the master at
work, or to listen to the stories of the elders; instead, formal schooling
becomes an individual’s admission ticket to the culture.

Cultural needs drive the unfolding of mental representations. One

cannot have symbolic thought in an individual mind, only in a mind that
is part of a certain culture. We think the way we think only because gen¬
erations before us have invented the ways of thinking we now use and
have accumulated much of what constitutes the specific contents of our
thought. Greeno (1994; 1995) has pointed out that people more typically
think and solve problem in teams than individually, but even when we
think as individuals, our thought processes are shaped by our cultural
environment and traditions. Bartlett (1932), one of the direct ancestors of
modern research on memory, was well aware of these social aspects of
cognition. He pointed out that memory is a constructive process and as
such, a social act.
There exists, thus, almost a consensus about multiple levels of mental
representation. The few serious attempts to argue for a single represen¬
tation have lost favor among the majority of today’s cognitive scientists.
Skinner’s behaviorism is no longer an intellectual force today. Newell and
Simon (1976) once proposed as a working hypothesis that all cognitive
representations are symbolic, the “Physical Symbol System Hypothesis,”
but Simon (1995) readily accepts the reality of multiple representations.
Only some of Gibson’s followers (though not Greeno or Neisser) and
some of the more radical members of the situated cognition camp seem
to cling to a unitary, direct action position. On the other hand, there is
much less agreement among cognitive scientists about the specific types
of mental models that have been distinguished here.
30 The theory

2.1.3 Implications

Embedding is a fact of cognition at every level. Thus, event memory is

not basically linguistic, but frequently it is linguistically encapsulated. It
is not the event itself that is remembered but the story told about it
(Barsalou, 1993). For example, in walking through an unfamiliar hotel
lobby, I notice a telephone booth and say to myself, “Good to know, I
might need that tomorrow,” as well as a softdrink machine, which I pass
by without awareness. Later, when I need to make a phone call and am
thirsty, I may remember both the telephone booth (intentional, goal-
directed learning — event memory plus linguistic code) and the drink
machine (incidental, non-goal-directed learning - pure event memory).
Karmiloff-Smith (1992) makes another important point: that mental
representations are not only embedded but that they are subject to repre¬
sentational redescription. Experiences that are encoded at a lower level of
representation can be intentionally redescribed at a higher level. Repre¬
sentational redescriptions of episodic memory at the narrative-language
level is probably quite automatic and, for the adult, requires few or no
resources. However, redescription at the abstract level may be difficult
and resource demanding, something we learn laboriously in school, not
something that comes naturally.
A number of other familiar distinctions between types of representa¬
tions that are made in cognitive psychology can also be mapped into this
framework. For example, the procedural declarative distinction. Percep¬
tual memory can be considered as the most elementary form of repre¬
sentation. All the other representations described are forms of declarative
memory. However, it must be remembered that lower forms of represen¬
tations are encapsulated within higher forms, so that in adult humans,
procedures may be represented at multiple levels (the example of the
turning radius of cars discussed earlier).
The semantic episodic memory distinction can also be captured:
Episodic memory is memory for events. All other declarative knowledge
representations are semantic.
How can we distinguish between propositional and imagery repre¬
sentations? Event memory as well as concrete, nonlinguistic mimetic
representations may involve imagery in the broad sense, that is, not re¬
stricted to visuospatial imagery. Propositions are the basic representa-
 Cognition and representation 31

tional units at the narrative-language level, to which I shall turn in the

next section.
Just as we need to broaden our concerns beyond the individual mind to
the social group and culture, we must explicitly include in our analysis the
environment with which the individual is interacting. This is by no means
a novel observation. Simon in the Sciences of the Artificial (1969) has made
this point very clearly with his image of the ant on the beach: If we want
to understand why the ant takes a particular path along the beach, we need
to know something about the ant (its motor and perceptual abilities), but
mostly we need to study the sand, where the big grains are that force a
detour and where there is smooth going. It is the structure of the envi¬
ronment, the affordances the sand provides, that determine the behavior
of the wandering ant. Simon has analyzed the Tower of Hanoi problem in
this way, and others have provided similar analyses (e.g., how to use a cof¬
fee pot, Larkin, 1983; see also chap. 11 of this book). Donald (1991)
extends Simon’s image by adding culture to it: As generations of ants move
along the beach, they leave a scent trace that tells other ants where to go.
However, the typology of mental representations has even stronger
implications. If symbolic thought, for instance, is indeed made possible
by external devices, such as written language, then these external devices
must be part of any analysis of cognition. The boundaries between the
internal and external become obscured in such an analysis, as Bateson
(1972) argued with his example of the blind man and his stick: The cog¬
nitive process must be ascribed to the system, not to the blind man alone
or the stick by itself. This point has been elaborated effectively by
Hutchins (1995), who provides a wealth of examples and detailed scien¬
tific analyses of the cognitive processes involved in ship navigation. It is
simply not possible to understand navigation as an individual cognitive
process. It involves team performance, and it includes the use of tools
and instruments. Navigation is the product of a system that consists of a
team of people, their instruments, and their social organization.
The representational system discussed here thus has the virtue of pro¬
viding a framework within which some of the current limitations of cog¬
nitive science may be overcome. That, of course, is more a program for a
future cognitive science than a current achievement, but it is a goal worth
keeping in mind in discussing what we have achieved so far in the study
of cognition and comprehension.
32 The theory

2.2 Propositional representations:

How science represents the mind

Les langues sont le meilleur miroir de l’esprit humain.

(Languages are the best mirror of the human mind.)
Gottfried Wilhelm (Leibniz, 1765/1949)

Der Mensch lebt mit den Gegenstanden hauptsachlich, ja, da

Empfinden und Handeln in ihm von seinenVorstellungen abhan-
gen, sogar ausschliesslich so, wie die Sprache sie ihm zufiihrt.
(We interact with objects mainly as they are represented by
language; indeed exclusively so, since perception and action
depend on memory images.)
Wilhelm von Humboldt (1792)2

We can take for granted that several different types of mental represen¬
tation play a role in behavior and cognition and that some representa¬
tions are embedded in others. How are we as scientists to represent these
layers of embedded representations in our theories? It is not easy to
envisage a theory of complex cognition that provides an optimal formal¬
ism for each type of representation and that adequately describes their
interactions, though for some purposes that may be required. However,
for a theory of comprehension, especially one that focuses on text com¬
prehension, a simpler solution is available: Find one form of representa¬
tion that fits everything. Obviously, a single format cannot fit everything
precisely, or we would not need to talk about different types of repre¬
sentation in the first place. However, there may be a form of representa¬
tion that allows us to adequately approximate the various types of repre¬
sentation. One does not have to look far to find a candidate for such a
superrepresentation: language. Language has evolved to enable us to talk
about all the world and all human affairs and is as suitable a tool for our
purposes as we have. Not that expressing everything through language
does not sometimes distort things. It simply means that on the whole
language is the best tool we have available.
What are the difficulties that arise when we represent the complexity

2 Both quotes are from Cartesian Linguistics (Chomsky, 1966).

 Cognition and representation 33

of the human mind by using language? There is the inevitable problem

of distortion. Text comprehension involves primarily language represen¬
tations at the narrative level. Therefore, if we find a good formalism for
the narrative level, we ought to be able to deal quite well with text com¬
prehension. However, even at the level of text comprehension, we often
must deal with nonnarrative representations, especially imagery and
emotion. I argue later that there are language-based representations of
imagery that are serviceable, that allow us at least to deal with visuospa-
tial imagery in text comprehension, not perfectly, but sufficiently for
some purposes (section 2.2.4). Emotion is harder to deal with, in part
due to the dearth of research in that area. I return to this topic only in
the final chapter of this book with some very tentative suggestions.
There exists, however, a second potential problem that is more subtle.
I have emphasized in the previous section the embeddedness of mental
representations. Thus, the narrative—language level that I suggest here
as the common format for all forms of mental representation does not
stand on its own but is constructed from nonlinguistic, episodic, and pro¬
cedural building blocks. A formalism suited for narrative-language rep¬
resentations does not by itself reflect this somatic, experiential basis of
language.
Choosing the narrative-language level as the theory language is not
what is commonly done in philosophy and linguistics. Instead, some form
of logical system, that is, the abstract language level, is usually the choice
for a formalism to describe cognition. However, to squeeze all human
cognition into a logical formalism greatly compounds the distortion
problem noted earlier: Compared with narrative language, logic (and
comparable, abstract mathematical formalisms) is a very inflexible sys¬
tem, not notably suited for the representation of natural language, and
even less so for lower levels of mental representation.
What form do narrative-level representations take? Natural language
itself is not an adequate medium for our theory of language. What we
need is a format suitable for the expression of meaning, of the semantic
relations that are also expressed in natural language. Natural language
serves many other functions besides the expression of semantic rela¬
tions — for example, communicative and social functions. It has long been
recognized in psychological research on language that it would be desir¬
able to have a formalism for the representation of meaning that is at least
34 The theory

to some extent independent from the way this meaning is expressed in

natural language on a particular occasion and in a particular context.'
The problem is that the same meaning can be expressed in words in many
different ways. Of course, just which way is chosen in a specific context is
by no means arbitrary and, indeed, is an important subject for study, but
for other purposes researchers must be able to abstract from the particu¬
lar words and phrases and to deal with meaning relations directly. For
most of the research described in this book, it is not the words themselves
that matter but the meaning they convey.

2.2.1 Features, semantic nets, and schemas

The representation system that we need must fulfill multiple functions:

It must serve for the mental representation of texts but also for other
memory structures, such as general knowledge, concepts and word mean¬
ings, and personal experiences (episodic memory traces). Because the
mental representations of text are in part derived from knowledge and
experience, it is desirable that both can be described in the same format.
Four kinds of systems have been widely used for the representation of
meaning: feature systems, networks, schemas, and propositions.
Feature systems were developed in philosophy and linguistics but have
become enormously popular in psychology. The original goal of feature
analysis was to find a finite set of basic semantic features that then could
be combined by semantic composition rules to form complex semantic
concepts, much as the 100-plus chemical elements are combined to yield
all the substances in this world. Thus, Katz and Fodor (1963) defined
“bachelor” as +HUMAN, 4-MALE, and +HAS-NEVER-MARRIED.
Apart from the obvious objection that there is more to the meaning of
“bachelor” than what is captured by these features, this program failed
because it proved impossible to come up with a list of basic semantic fea¬
tures corresponding to the chemical elements. However, psychologists
were satisfied with a weaker version of feature systems, allowing in addi¬
tion to defining features ad hoc characteristic features as needed (Smith,
Shoben, & Rips, 1974). In a similar vein, Smith & Medin (1981) have
suggested a probabilistic feature representation.

3 Among others, Kintsch (1974), pp. 244ff.

 Cognition and representation 35

Almost all psychological models of categorization (e.g., Estes, 1986;

Rosch & Mervis, 1975; Smith & Medin, 1981) assume a feature repre¬
sentation, as do the dominant models of episodic memory (Gillund &
Shilfrin, 1984; Hintzman, 1988; Murdock, 1982). Connectionist models
typically rely on a feature coding of their input (Rumelhart & McClel¬
land, 1986).
Feature representations are popular because they are so simple. But
they are deceptively simple because they do not explicitly represent con¬
ceptual relations. People and even animals are, however, highly sensitive
to relations. It not sufficient to assign to “rose” the features RED and
FRAGRANT; these features are the values of the attributes COLOR and
SMELL and hence stand in a particular kind of relationship to the con¬
cept that must be represented. Similarly, “break” is not defined by fea¬
tures such as BOY, VASE, and HAMMER, for this neglects the fact that
BOY is the agent, VASE the object, and HAMMER the instrument of the
action. The features MOUNTAIN, SKI, and BINDING do not define
“skiing,” no matter how many additional features are added to the list,
because their relationship to the to-be-defined concept as well as their
relationships among each other are neglected. It is true that feature rep¬
resentations capture some aspects of the meaning of concepts, as their
successful use in so many psychological theories attests. For purposes that
require more than a rudimentary representation of meaning, however,
feature representations are inadequate.
The use of feature representations may be questionable on additional
grounds. Features are often thought of as something out there in the real
world - namely, the feature - that can be attended to, selected, and used
for various psychological purposes. This seems not to be the case, however.
Wisniewski (1995) has argued convincingly that features are psychologi¬
cal constructions and that what is constructed to be a feature depends on
context, goals, and experience. Thus, a line starting from the neck of a
crudely drawn human figure may be constructed as a “tie” by a subject try¬
ing to differentiate city and farm dress, but as a “row of buttons” by some¬
one attending to the presence or absence of detail in the drawings. Theo¬
ries of categorization ought to explain how people construct features; they
should not start with features as representational primitives.
Associative networks have concepts as nodes and unlabeled links. They
are the oldest form of knowledge representation, going back to Aristotle.
In an associative net, knowledge is represented as a net of concepts linked
36 The theory

by associations of varying strength. For Aristotle, these associations were

based on temporal and causal contiguity. Later associationists elaborated
on these ideas. A wide variety of experimental data support the psycho¬
logical reality of such structures. For example, associative strength is com¬
monly estimated by the frequencies of responses in free-association exper¬
iments. When people are asked in a lexical decision experiment to say
whether a particular letteustring is an English word or not, they are 85 ms
faster when they have just seen a strongly associated word rather than an
unrelated control word (Meyer & Schvaneveldt, 1971), which suggests
that these associative links play a role in lexical access. On the other hand,
associative nets are clearly limited in scope, beause not all knowledge can
be represented by unlabeled links between simple concepts.
Semantic networks have concepts as nodes and labeled links. Links are
formed by relations such as class inclusion (IS-A) or PART-OF. In the
simplest case, well-ordered hierarchies are obtained in this way. The great
advantage of semantic nets lies in their ability to account for the inheri¬
tance of properties and hence for economy of storage. The classic data
supporting the psychological reality of semantic nets were reported by
Collins and Quillian (1969), who showed that sentence verification times
for sentences like A canary has skin were longer than for sentences like A
canary can fly, which in turn were longer than for sentences like A canary
can sing. These results were interpreted to mean that the verification time
for statements that were derived by chains of arguments such as Canaries
are birds — birds are animals — animals have skin is proportional to the num¬
ber of steps in the hierarchy that have to be traversed (two for skin, one
for fly, none for sing). Later results modified and qualified this interpre¬
tation in several ways. The number of steps in the hierarchy seems to be
only one factor determining verification times, and it is often overridden
by saliency, typicality, and frequency (e.g., Conrad, 1972). In particular,
very frequent properties of an object appear to be stored directly, even
when they could be derived through property inheritance in a semantic
net. Various forms of semantic nets are widely used in Artificial Intelli¬
gence (AI) (Barr & Feigenbaum, 1982).
Schemas, frames, and scripts are structures used to coordinate concepts
that are part of the same superstructure, or event. Well-known examples
are the room frame of Minsky (1975) and the restaurant script of Schank
and Abelson (1977). Scripts, frames, and schemas have proved to be pow¬
erful computational devices in AI. Their psychological status has been
 Cognition and representation 37

investigated in such studies as Bower, Black, and Turner (1979), which

showed that readers given parts of a schema readily infer the missing
components. Indeed, they often believe that they have actually read what
they infer.
Schemas in one form or another seem to be indispensable theoretical
constructs for cognitive theory. There has been an important change,
however, in the way schemas are conceived of. For Schank and Abelson
(1977), a schema was a fixed, complex mental structure that was retrieved
from memory when needed and used to organize some experienced
event. This concept proved to be difficult because a fixed schematic
structure had to be imposed on the contextually variable, fluid events that
humans actually experience. Therefore, schemas came to be thought of,
not as fixed structures to be pulled from memory upon demand, but as
recipes for generating organizational structures in a particular task con¬
text (Kintsch & Mannes, 1987; Schank, 1982; Whitney, Budd, Bramuci,
& Crane, 1995). Context-sensitive generation ensures that the structure
that is generated is always adapted to the particular context of use.

2.2.2 The predicate—argument schema: Networks of propositions

The various knowledge representations discussed above all have their

strengths and all have their uses, but none of them is sufficient for our
purposes by itself, for the reasons mentioned. Networks of propositions
provide an alternative that combines and extends their advantages and
avoids some of their limitations.
The predicate—argument schema can be regarded as a basic unit of lan¬
guage. It is commonly referred to as a proposition, a term borrowed from
logic but used here in a different sense.4 For the purpose of text repre¬
sentation, a proposition is simply a predicate-argument schema.
Atomic propositions (e.g., Kintsch, 1974) consist of a relational term,
the predicate, and one or more arguments, written as PREDICATE
[ARGUMENT, ARGUMENT, . . .]. The predicate determines the num¬
ber and kinds of arguments that may fill argument slots, that is, the seman-

4 Some of the borrowers did not intend originally to stray beyond the boundaries of the
field of logic (e.g., Bierwisch, 1969; van Dijk, 1972), but others (including Kintsch,
1974) extended the meaning of the term, in spite of the possibility of confusion with
its original logical meaning.
38 The theory

tic role of the participants. For example, the predicate GIVE may have
three argument slots, as in GIVE[agent:MARY, object:BOOK, goal:
FRED], Atomic propositions may be modified by embedding one propo¬
sition within another, as in GIVE[agent:MARY, object:OLD [BOOK],
goahFRED]], or INADVERTENTLY [GIVE[agent:MARY, object:
OLD[BOOK], goahFRED]].
Complex propositions (van Dijk & Kintsch, 1983) are compounds
composed of several atomic propositions that are subordinated to a core
propositional meaning. The general schema for complex propositions is
given by

Category (action, event or state):

- Predicate:
_ Arguments (agent, object, source, goal, . . .);

- Modifiers:

Circumstance:
- Time:
- Place:

Thus, the sentence Yesterday, Mary gave Fred the old book in the library
would be represented as:

- Action:
—Predicate: GIVE
— Arguments:
“Agent: MARY
“Object: BOOK
Modifier: OLD
-Goal: FRED

Circumstance:
-Time: YESTERDAY
L Place: LIBRARY

Not all expressions in the surface structure are represented in this nota¬
tion (e.g., tense is not, nor is the definite article, which is a discourse sig¬
nal that identifies that that book is known to the hearer). In general, surface
sti ucture may express pragmatic, rhetorical, stylistic, cognitive, or inter¬
actional properties, as well as additional syntactic and semantic properties
 Cognition and representation 39

that are neglected by this notation. The choice of which sentence proper¬
ties to represent in the propositional notation is pragmatic: Whatever
seems of little importance for a given theoretical or experimental purpose
is omitted. Hence, significant differences may be found in the explicitness
and completeness of propositional representations constructed for differ¬
ent uses.
The van Dijk and Kintsch (1983) notation in terms of complex propo¬
sitions has several advantages over the earlier notation of Kintsch (1974)
in terms of atomic propositions. In the earlier notation, the sentence
would have been represented as follows (using [.] as an abbreviation for a
list of arguments when it is clear what these arguments are):

GIVE[M ARY,BOOK,FRED]
INADVERTENTLY[GI VE[. ]]
OLD[BOOK]
YESTERDAY[GIVE[.]]
IN-LIBRARY[GIVE[.]]

or, equivalently,

PI GIVE[MARY,P3,FRED]
P2 INADVERTENTLY[P 1 ]
P3 OLD[BOOK]
P4 YESTERDAY[P 1]
P5 IN-LIBRARY[P 1 ]

Complex schemas preserve more of the actual structure of the sentence

and hence some of the discourse signals the speaker wanted to convey. For
example, the speaker’s choice of sentence subject determines the subordi¬
nation relations within a complex proposition. They also are generally eas¬
ier to use for such psychological purposes as scoring recall protocols.
Texts consist of more than one complex proposition that may be
related in different ways. Three levels of relationship among propositions
have been distinguished by van Dijk and Kintsch (1983). By definition,
unrelated meaning units do not form a text or discourse. The three levels
are the following:

1. Indirect coherence. The meaning units are part of the same episode.
That is, they share a time, place, or argument.
2. Direct coherence. The same as indirect coherence, but in addition the
40 The theory

coherence is specifically marked by separate clauses or sentences. Sen¬

tence adverbials such as therefore, then, so, as a result, and so on, com¬
pound sentences, or explicit coordinating connectives may be used when
the meaning units are linearly ordered or when pairs or ^-tuples are
formed (e.g., cause-consequence pairs).
3. Subordination. Various devices indicate that one meaning unit is sub¬
ordinated to another. Whep one predicate-argument schema is taken as a
specification of another (e.g., a condition), a complex sentence with a full
embedded clause is typically used, represented propositionally as two
underlying complex propositions. When only one aspect of a schema is
specified rather than a whole clause, such as a specification of the manner
of action or the property of a participant, restrictive relative clauses are
used. Adjectivization signals an even stronger degree of subordination.
The latter kinds of subordination are considered part of the same mean¬
ing unit, or complex proposition.

Consider the following minitext:

(1) The snow was deep on the mountain. The skiers were lost, so
they dug a snow cave, which provided them shelter.

The propositional representation would be as follows:

pDEEP LOST rDIG

Lsnow Lskiers 1-SKIERS
Lsnowcave PROVIDE
-SNOWCAVE
-SKIERS
-SHELTER
•-MOUNT <MOUNT> <MOUNT>

The coherence between the first two meaning units is indirect, through
the shared place, because the skiers presumably were lost on the mountain.
The coherence between the second and third proposition is direct,
expressed by so, but in addition coherence is provided by argument over¬
lap via agent and place, for the skiers presumably are still on the moun¬
tain. The fourth meaning unit, PROVIDE[SNOWCAVE, SKIERS,
 Cognition and representation 41

SHELTER], is embedded into the third as a specification for one of its

argument. (Propositionalizing is discussed in detail in section 3.1.1.) The
predicate-argument representation of meaning has advantages over the
alternative forms of representation discussed in the previous section
because it is flexible enough to subsume these systems. Predicate—
argument schemata are a more general form of representation within
which all the alternative meaning representations described can be
expressed, as follows:

1. Features can be expressed in a predicate-argument schema as in the

following example: RED[ROSE,COLOR] is a meaning unit that
expresses a value—object-attribute relation. (To simplify things, the
redundant attribute is often omitted in practice.) We could even add a
modifier, like .6-PROBABILITY.
2. Associations, of course, are basic to the kind of networks advocated
here. Meaning units may be linked by unlabeled links varying in strength.
That is, a propositional network is an associative network, the nodes of
which consist of predicate-argument units.
3. Semantic networks are networks with labeled links. Such networks
can be constructed by creating units that label the relation between two
nodes. For instance, the IS-A relation between BIRD and ROBIN in a
semantic net can be mimicked by a predicate-argument unit IS-A
[ROBIN,BIRD], linking the nodes BIRD and ROBIN in a network.
4. Frames, scripts, and schemas can also be expressed in this notation.
Thus, a concept like ROBIN can be a node in the network that is linked
to all kinds of other nodes that specify our knowledge of “robin.”

Writing the restaurant schema as a predicate-argument unit with

embedded arguments that are themselves schemas is quite straightfor¬
ward:
RESTAURANT
props[OBJECTS[. . ,],ROLES[. ..],.. .]
event-sequence[ENTER[. . . ],ORDER[ .],
EAT[. . ,],LEAVE[. . .]]

where the ellipsis dots indicate information that has been omitted for the
sake of brevity. A complex proposition like this does the same work as a
conventional schema: It specifies relevant attributes and lists their default
values.
42 The theory

5. It is also worth noting that production rules, which are the preferred
representational system for procedural knowledge (Anderson, 1983;
1993; Newell, 1990), can also be written as schemas with two slots IF[. . .]
and THEN[. . .].

The predicate-argument notation, therefore, is sufficiently flexible to

serve as a notational variant for other forms of representation, such as
features, frames, and productions. Important aspects of other forms of
knowledge representations, such as the notion of defaults in schemas or
the executability of productions, can be imported into this notation. The
advantage that is gained by this rewriting is that all these variants can be
brought together in the same format - a propositional network. Good
psychological reasons can be advanced for the use of every one of these
alternatives for representing meaning, but none of them is adequate by
itself. By using a uniform format, one can take advantage of what all
these systems have to offer without having to accept the limitations of
any one.

2.2.3 Encapsulated meaning

For all its advantages, the propositional representation proposed here

also has some limitations, even when it is used primarily to represent
the meaning of texts. One problem arises from the fact that higher levels
of representation encapsulate lower levels. Hence, separating propo¬
sitional representations from the layers of representation that are em¬
bedded within them introduces some distortion.The kind of mental
models people construct changes with the level of their phylogenetic and
ontogenetic development. The adult human mind is a hybrid system that
relies on all types of mental models simultaneously. Abstract thought
does not displace storytelling or generalized event memory. Rather, they
exist side by side, each with its own function. A cognitive representation
of meaning should capture relevant aspects of meaning at all these dif¬
ferent levels.
How people understand a proposition, therefore, also depends on the
meaning that the proposition encapsulates. Consider the meaning of a
simple statement like red rose. It originates from a perceptual experience
of certain flowers of a certain color range. The verbal expression origi¬
nally acquired meaning by the procedure of looking at a certain object
 Cognition and representation 43

and experiencing a certain color sensation. Looking at many objects of the

same color, a generalized perceptual procedure red(X) is generated. Once
the procedure red(X) has been acquired, it can be applied to other words,
even without the actual sensory experience: red bird is meaningful even
though I may have never seen such a bird, because I can apply the per¬
ceptual procedure red to the object in question. The perceptual proce¬
dure .m/ has become extended, so that we now have a linguistic procedure
red. Assigning red to light of a certain wavelength creates an abstract pro¬
cedure that gives meaning to red in the symbolic system of physics. Sim¬
ilarly, there is a perceptual—motor procedure that is the basis for my
understanding offlying bird and an perceptual-emotional experience that
is the basis for my understanding of angry dog, and the extension to fly¬
ing dragon and angry dragon, which I need not actually experience in real¬
ity. The point is that meaning is rooted in perception, action, and emo¬
tion. But it does not stay there, words become meaningful because of
their relation to other words, abstract concepts become meaningful
because of their relation to words, and so on. Lakoff (1987), among oth¬
ers, provides many illustrations of how abstract concepts are embodied in
the pattern of bodily interactions with the world. He is concerned with
the origins of abstractions, because their origins explain how we under¬
stand and use abstractions. But although words as well as abstract con¬
cepts are perceptually based and embodied, one must not forget that as
words and abstractions they take on a new role and function within
higher-order mental models, which have properties of their own that are
not expressible by and reducible to the models at a lesser level of the cog¬
nitive hierarchy. Cognition is not all symbolic thought, but neither is it all
direct action.
The cognitive-meaning representations in different types of models
are not independent, however. The units remain the same at each level, or
at least highly interdependent. What changes are the procedures used to
give these units their meaning, varying from perceptual-motor proce¬
dures to relations among words, to abstract relations. Red has a proce¬
dural interpretation at the level of perceptual experience, but the linguis¬
tic red or the red of the physical model has emergent properties that are
not present in the basic perceptual experience. Red Guards derives its
meaning from a linguistic, metaphorical extension; red shift depends for
its meaning on a complex, abstract argument.
44 The theory

Propositional representations of meaning are designed to capture

salient semantic relations at the narrative-linguistic level of encoding,
but they do not necessarily express some of the encapsulated, lower-level
meaning relations. However, one should not forget that language did not
evolve in a vacuum but that it is a tool humans invented in the service of
action. Thus, from the very beginning, language was attuned to the
world in which we live apd act. Language developed to reflect the con¬
straints of human action and human perception, probably as well as they
can be reflected in any medium. Barsalou (1993) and Glenberg (1997)
argue for cognitive representations that are closely tied to human per¬
ception and action - perceptual symbols in Barsalou’s case and a spatial-
functional medium in the case of Glenberg. However, their arguments
do not speak as much against the use of propositional units for mental
representations as for the need to interpret these units in terms of per¬
ceptual and spatial-functional considerations, not purely in terms of a
linguistic or abstract level. One should be wary of potential distortions
introduced by the linguistic medium, but we talk successfully all the time
about what we do and perceive - we ought to be able to do science in a
similar way.

2.2.4 Imagery

Predicate-argument schemas impose a propositional format on all mean¬

ing units. Although this format is not very limiting and capable of
expressing a wide variety of structures, predicate-argument schemata are
clearly most suitable for the representation of propositional information,
and human knowledge and cognitive processes are not restricted in this
way. Both observation and experiment have yielded incontrovertible evi¬
dence for the importance of nonpropositional representations in cogni¬
tion. For instance, the well-known mental rotation experiments of Shep¬
ard and Metzler (1971) and the various studies of Kosslyn (1980) clearly
show that spatial imagery is used to represent certain perceptual infor¬
mation. In addition, Santa (1977) provided evidence that shows that non¬
propositional representations may be either spatial or linear. Anderson
(1983) claims that we must deal with at least three forms of mental rep¬
resentations in cognition - linear, spatial, and propositional.
Whether nonpropositional representations are really needed or
whether they can and should be translated into propositional representa-
 Cognition and representation 45

tions (Anderson, 1978; Kosslyn, 1980; Pylyshyn, 1981) has been the sub¬
ject of a long-standing debate. No definitive conclusions have been
reached in this debate on the basis of either logical argument or behavioral
data. However, Kosslyn (1994) has tried to resolve the imagery debate by
considering the evidence from brain studies. The processes underlying
mental imagery in the brain appear to be closely related to perceptual
processes and can be clearly differentiated from verbal and symbolic
processes.
Nevertheless, for many purposes — including modeling text compre¬
hension with the construction—integration model - it is desirable to
work with a uniform medium, that is, the predicate-argument schema.
Therefore, imagery and linear strings will be translated into this format -
not because all information is by nature propositional but because of
practical considerations. We know how to work with predicate-argument
units, and it is not clear how to interface linear or spatial analog repre¬
sentations with such units. On the other hand, the translation from the
linear or spatial analog to propositional form should be done in such a
way that the unique properties of the analog representations are main¬
tained in the propositional form. That is, correspondences must be
established between direct forms of reasoning that are possible in these
analog systems in the propositional domain (e.g., Furnas, 1990). Thus,
ideally, representational properties that are characteristic of imagery but
not of verbal or propositional information (some such properties are the
existence of a canonical view, symmetry biases, the hierarchical structure
of space, framing effects, alignment, and perspective) should be repre¬
sentable in the propositional format. This is not an easily achieved ideal,
however.
The best-known approach to relating propositional and imagery rep¬
resentations is that of Kosslyn (1980). Kosslyn distinguishes a deep, non¬
pictorial level of representation (names and lists ol spatial coordinates)
and a surface, quasi-pictorial representation in a spatial medium. Spatial
information can thus be translated into propositional form through the
use of spatial predicates. This procedure does not necessarily satisfy the
criterion stated above that the properties of spatial reasoning be pre¬
served in the transformation. It is not entirely satisfactory, therefore, and
can only be considered as a stopgap solution in the absence of a more ade¬
quate procedure.
To illustrate this treatment of spatial imagery, a brief example from
46 The theory

Kintsch (1994a) will be described. Consider the sentence John traveled by

car from the bridge to the house on the hill. This sentence involves the con¬
cepts JOHN, CAR, BRIDGE, HOUSE, and HILL and expresses the
proposition TRAVEL[agent:JOHN, instrument:CAR, source:BRIDGE,
goahHOUSE, modifier:ON[HOUSE,HILL], The visual image associ¬
ated with this sentence contains somewhat different information. In par¬
ticular, John's car might be on a road, and a river might be under the bridge.
Thus, the representation that contains both the verbal, propositional
information derived from the sentence, and the propositional translation
of the image that was generated by that sentence, might look like Figure
2.2.
The spatial predicates are shown above the arguments of the proposi¬
tions, and the verbal predicates below. There is no claim made that every
reader will form just such an image, just as we cannot be certain that every
reader will interpret a sentence in exactly the same way. The point to be
illustrated here is merely that (some) spatial information can be translated
into a propositional format by using spatial predicates and thus becomes
an integral part of the text representation. Furthermore, the verbal and
imagery information are in part redundant (the house is on the hill) but not
entirely so. In some cases the verbal information is richer (John travels);
in other respects the image is richer (the inclusion of the road and the
liver).
Thus, although we are unable to deal with imagery representations in
a fully satisfactory manner, there is at least some way to deal with
 Cognition and representation 47

imagery in modeling comprehension processes. The problem is by no

means restricted to the present model but afflicts much of cognitive sci¬
ence today. It is to be hoped that the results that are beginning to become
available from researchers focusing on this question (e.g., Chan-
drasekaran & Narayanan, 1990) will provide better solutions in the near
future.
Among the building blocks of cognition are perceptual symbols,
imagery, and actions. The predicate-argument schema that we use for the
representation of cognitive structures, on the other hand, stems from the
linguistic domain. By extending it from the linguistic domain to all of
cognition, from abstract thought to concrete action, we have created a
unitary representational format for all of cognition. This has some obvi¬
ous advantages, but it also carries some risks: The predicate-argument
schema does not necessarily highlight relations that are significant in the
realm of action and perception in a direct, analogous manner. The theo¬
rist has to be more careful to make sure that salient perceptual informa¬
tion is indeed represented. The representation format invites a verbal
bias, which the theorist must try to compensate for. Nevertheless, the
predicate-argument schema is not only the best we have, it is also rea¬
sonably satisfactory on an absolute scale; it is quite satisfactory, indeed,
for work on text comprehension, which has a strong verbal component;
and it provides an at least feasible approach for research more directly
related to perception and action.

The study of cognition entails the need for some way of repre¬
senting mental structures. The difficulty is that there are many
different kinds of mental structures and that a good format for the
representation of mental structures should be suitable for all
kinds of structures. The basic linguistic meaning unit is the pred¬
icate-argument schema, and this schema can be used to represent
other types of structures as well. It is a very general format that
subsumes feature representations, semantic nets, production sys¬
tems, and frame systems. However, predicate-argument units
must not only represent abstract units of meaning but must also
represent meaning at multiple levels, including the perceptual,
action, linguistic, and abstract-symbolic levels. Although predi¬
cate-argument units are more suitable for the representation of
The theory

meaning at some levels than at others (simply put, it is more nat¬

ural to represent a sentence that way than as an image), they can
be used generally with proper care, and they appear to be the best
format available for representing mental structures in a general
theory of cognition.
 3

Propositional representations

3.1 The propositional representation of text

Psychologists studying language processing need a better representation

of meaning than is provided by the words and sentences of the language
itself, that is, a representation that more directly reflects the semantic
relations that are crucial for how people understand, remember, and
think with language. Language itself serves too many other goals, so that
a representation focused on meaning is often needed for the empirical
study of language processing. Propositional representations of text serve
that purpose. They make explicit those aspects of the meaning of a text
that are most directly relevant to how people understand a text. Such rep¬
resentations fall short of a complete formal analysis, but then, that is not
what we need for our purposes. A cruder, but robust representational sys¬
tem is sufficient.
As in previous work (van Dijk & Kintsch, 1983), we distinguish on the
one hand between the microstructure and the macrostructure of a text,
and on the other between the textbase and the situation model. The
textbase-situation model distinction refers to the origin of the proposi¬
tions in the mental representation of the text. Those propositions that are
directly derived from the text constitute the textbase. However, only in
rare cases is the result of comprehension a pure textbase; usually in order
to understand a text, comprehenders supplement the information pro¬
vided by a text from their knowledge and experience (long-term memory )
to achieve a personal interpretation of the text that is related to other
information held in long-term memory. The complete structure that is
composed of both text-derived propositions (the textbase) and proposi¬
tions (this includes imagery and action, which we also represent as propo¬
sitions) contributed from long-term memory is called the situation model.
50 The theory

The micro- and macrostructure distinction is orthogonal to the

textbase-situation model distinction. The microstructure is the local
structure of the text, the sentence-by-sentence information, as supple¬
mented by and integrated with long-term memory information. The
macrostructure (van Dijk, 1980) is a hierarchically ordered set of proposi¬
tions representing the global structure of the text that is derived from the
microstructure. It is sometimes directly signaled in a text, but often it
must be inferred by the comprehender. An ideal summary is (or should
be) a text expressing the macrostructure. Because summaries can vary in
their level of generality, the macrostructure is also hierarchical, so that
one may have macropropositions at different levels of generality (e.g.,
corresponding to major headings and subheadings in a text).
Thus, the textbase, with its micro- and macrostructure, is obtained
from a semantic analysis of a text and its rhetorical structure, as the
author of the text intended it. It is the sort of analysis linguists and
semanticists perform. 'The mental representation of a text a reader con¬
structs includes the textbase (not necessarily complete or veridical) plus
varying amounts of knowledge elaborations and knowledge-based inter¬
pretations of the text — the situation model. Neither the micro- nor the
macrostructure of the situation model is necessarily the same as the
micro- and macrostructure of the textbase, for the reader may deviate
from the author’s design and restructure a text both locally and globally
according to his or her own knowledge and beliefs.
An example may help illustrate these distinctions (after E. Kintsch,
1990). Suppose we have a text of four paragraphs, each comparing the
geography, agriculture, industry, and population of two countries,
Argentina and Brazil. And suppose we have a reader who knows a lot
about and is much interested in Argentina but neither knows anything
nor cares much about Brazil. The textbase formed by such a reader would
consist of a microstructure derived from the sentences and phrases of the
text, in the order they were presented in the text. It is the translation of
the verbal text into a corresponding propositional structure. The top-
level macrostructure would have four major subdivisions corresponding
to the four topics discussed in the text. At a lower level there would be two
macropropositions generalizing the information presented on each topic
for both countries. The microstructure propositions would be directly
subordinated to these second-level macropropositions, yielding a hierar-
 Propositio nal rep re sen ta tio ns 51

Figure 3.1

chical structure, as in Figure 3.1 (hypothetical text propositions are

denoted as Pj; only a single branch of the text hierarchy is filled in). What
sort of situation model might be constructed by our hypothetical reader?
Because our reader knows nothing about Brazil and does not care much
about it, textbase and situation model would be the same in this part of
the text. However, the reader might add information from his own knowl¬
edge to the Argentina part of the text, elaborating it as in Figure 3.2,
where S1-S4 are the knowledge elaborations added to the textbase. In
Figure 3.2, textbase and situation model are not much different, and the
macrostructure remains unaffected by the Argentina elaborations we
have added. However, a knowledgeable and interested reader may reorga¬
nize his mental representation much more thoroughly. For instance, the
52 The theory

reader may structure his situation model according to countries, rather

than topics as was done in the text, arriving at something like Figure 3.3.
In addition to the new macrostructure, the microstructure in Figure
3.3 has also been reorganized. Although the Brazil portion of the text
remains unchanged, the Argentina part has been restructured: The prop¬
ositions the reader has added from his knowledge about the country are
not just added elaborations but are integrated with the text and impose a
different structure from the one suggested by the text.
Which of these structures, or innumerable other variants, will readers
 Propositio nal rep re sen tatio ns 53

Figure 3.3

actually generate? If readers have all this freedom to reorganize and dis¬
tort a text according to their ideas and beliefs, how can we ever hope to
predict what they will do? Actually the situation in text comprehension is
not that bad, and better than in some other areas of cognitive science,
notably problem solving. When we give a subject a reasonably complex
problem to solve, we really have no way of telling how the problem will be
approached. In text comprehension, on the other hand, the subjects in
our experiments tend to follow the text we present more or less faithfully.
Many texts are designed so that the subject’s situation model will not dif¬
fer much from the textbase and hence is fairly well predictable. When we
do studies in which prior knowledge and beliefs play a larger role, we usu¬
ally have some idea what these are and are able to formulate hypotheses
about the way they should affect text processing. Thus, although we do
not have total control over our experimental subjects, we can constrain
their behavior reasonably well in our experiments. Real-life comprehen¬
sion episodes are much less constrained, and instead of prediction we
54 The theory

often must be satisfied with postdiction based on the principles of com¬

prehension derived from studies that are better controlled.

3.1.1 The microstructure

Even though the propositional representations of the meaning of a text

that are used here fall short of the requirements of a full formal semantic
analysis, or even a more informal linguistic analysis, they pose a major
unsolved problem. No fully automatic parser has yet been constructed
that is capable of deriving a propositional microstructure from arbitrary
English text. It is fairly simple to do that for limited domains (e.g., first-
grade word-arithmetic problems, as in Cummins, Kintsch, Reusser, &
Weimer, 1988), but the task is formidable if no restrictions are placed on
the nature of the text.
An escape route from this dilemma that I have used since my original
work in this area (Kintsch, 1972, 1974) is to ignore the parsing problem.
This provides no solution and hardly does justice to a highly important
problem, but it allows us to go on with research that would be impossible
unless we find some way to finesse the parsing problem. I am by no means
the only researcher in psychology, linguistics, or AI who has taken this
approach. Others have made the parsing problem the focus of their work,
and much progress has been reported in recent years. Nevertheless, my
hope that I could just pick a parser developed in some other laboratory
from the shelf and adapt it as a front end for my comprehension models
has so far not been realized.
Even if one decides to do without a formal parser, research of the kind
reported later in this book requires a reliable and objective system of
deriving propositions from any given text. Hand coding is slow and cum¬
bersome, but it serves its purpose as long as it is objective. A brief guide
for propositionalizing texts is given in the section that follows (“A Brief
Guide to Propositionalizing Texts”). This guide is then followed by a
sample analysis of a few sentences from a biology text.

A brief guide to propositionalizing texts. The meaning of a simple sentence

can be represented by a complex proposition (section 2.2.2) consisting of a
predicate with several arguments, time and place circumstances and
optional modifiers. Predicates are relational terms, frequently expressed
in language as verbs, adjectives, or adverbs. Each predicate is character-
 Propositional representations 55

ized by a predicate frame that specifies how many and which arguments
that a predicate takes. A verb frame, for instance, may specify that a par¬
ticular verb must have an agent, an object, and an optional goal, with
restrictions on these categories, such as that the agent must be human, or
the goal must be a location, Not only can concepts be arguments, but a
proposition may have as an argument another embedded atomic proposi¬
tion. • Sentence connectives, for instance, are predicates that require
propositions as arguments. Circumstances refer to the whole propositional
frame, specifying place and time. All propositional elements may be mod¬
ified by additional atomic propositions.
Thus, to repeat the example discussed in section 2.2.2, the sentence
Yesterday, Mary inadvertently gave Fred the hook in the library is repre¬
sented as a complex proposition structured around the predicate GIVE.
The arguments of this predicate are an agent MARY, an object BOOK,
and a goal FRED. Inadvertently modifies GIVE, and old modifies BOOK.
Time and place are specified by yesterday and in the library.
A brief list of frequently encountered propositional constructions fol¬
lows. This list parallels the one provided by Bovair and Kieras (1985) to
facilitate comparison. In each case, only a simple text example and its
propositional form are given without much comment. When more than
one version is given, these are notational variants of varying explicitness
to be used as convenient.

1. Verbs as predicates. Verb frames specify the arguments that can go

with each verb and are thus building blocks for propositions. It is usually
not necessary to indicate the case role of an argument:

The hemoglobin carries oxygen.

CARRY[HEMOGLOBIN, OXYGEN]

Prepositions indicate the case role of arguments and may be included in

a proposition for clarification:

The blood from the body arrives at the atrium through the veins.

ARRIVE[BLOOD,FROM-BODY, AT-ATRIUM,
THROUGH-VEINS]
56 The theory

or, abbreviated

ARRIVE[BLOOD,BODY,ATRIUM, VEINS]

2. Propositions as arguments. Propositions can be embedded as argu¬

ments in other propositions. For instance, an argument may be specified
by a modifier:

The blood arrives at the right atrium.

ARRIVE[BLOOD,RIGHT[ATRIUM]]

The first chamber is the right atrium.

IS[FIRST[CHAMBER],RIGHT[ ATRIUM]]

A sentence complement can be expressed as the argument of a superor¬

dinate proposition:

Purplish blood tends to lack oxygen.

TEND
-BLOOD
L PURPLISH
Llack
- BLOOD
L PURPLISH
L- OXYGEN

or more compactly:

TEND [PURPLISHf BLOOD],LACK[ PURPLISH [BLOOD],

OXYGEN]]

or, once again, using a different notation:

PI TEND[P2,P3]
P2 PURPLISH [BLOOD]
P3 LACK[P2,OXYGEN]

Which of these alternative notations is employed is a matter of conve¬

nience and preference.
 Propositional representations 57

The doctor discovered that the defect was congenital.

DISCOVER[DOCTOR,CONGENITAL[DEFECT] ]

Relative clauses are expressed as separate atomic propositions that are

assigned to a modifier slot in a complex proposition:

Blood that has been drained of oxygen arrives at the right atrium.

ARRIVE[BLOOD],RIGHT[ ATRIUM]]
I— DRAIN [BLOOD,OF-OXYGEN]

Modals are also expressed as modifiers in a complex proposition:

Heart attacks may be fatal.

FATAL[HE ARTATTACK]
•— POSSIBLE

or

POSSIBLE[FATAL[HEARTATTACK]

3. Modification. Adjectives and adverbs are treated as the predicates of

atomic propositions:

septal defect

SEPTALfDEFECT]

which is abbreviated in the context of a complex proposition as

DEFECT
1-SEPTAL

The blood returns to the heart quickly.

QUICK[RETURN[BLOOD,HEART]]

or

RETURN [BLOOD, HEART]

1- QUICK

Predicate nominals are treated similarly:

58 The theory

The doctor was a cardiologist.

CARDIOLOGIST[DOCTOR]

Multiple embeddings indicate the scope of modifiers:

the two major chambers of the heart

POSS ESS [HE ART,TWO[MAJOR[ CHAMBERS]]]

Negated propositions are treated as special predicates. This use of nega¬

tion is nonstandard but avoids some complications that arise when negated
propositions become part of a network that is integrated in the compre¬
hension process.

The heart rate is not constant.

NOT-CONSTANT[HEART RATE]

Paul did not expect Mary to come.

NOT-EXPECT[PAUL, COME [MARY]]

Superlatives, comparatives, and questions are treated in as simple a man¬

ner as possible:

The left ventricle is the largest chamber of the heart.

IS[LEFT[ VENTRICLE],LARGEST[CHAMBER]]
1— OF-HEART

OF-HEART is used here as an abbreviation for a POSSESS (or HAVE)-

proposition:

POSSESS[HEART,CHAMBER]

The right ventricle is smaller than the left ventricle.

MORE-THAN[SMALL[RIGHT[VENTRIC]] SMALL
[LEFT[VENTRIC]]]
 Propositional representations 59

or, simplifying some more,

SMALLER-THAN[RIGHT[VENTRIC]],
[LEFT[VENTRIC]]

How does the heart pump blood?

HOW[PUMP[HEART,BLOOD],?]

4. Problems and issues. Cognitive states (hope to, decide to, expect to, fail
to, be afraid to, begin to, hesitate to), causal verbs {cause, bring about, result
in), and verbs of saying, thinking and believing are treated as modifiers of
the sentence complement:

Paul hoped Mary would come.

COME
LMARY
L HOPE[PAUL]

Paul said Mary would come.

COME
lMARY
- SAY[PAUL]

John left early. This shocked everyone.

LEAVE
-JOHN
Learly
LSHOCK
LEVERYONE

or simply

PI LEAVE[JOHN]
P2 EARLY[P1]
P3 SHOCK[P2, EVERYONE]
 60 The theory

Sentence connectives take complex propositions as arguments:

CON[P 1 ,P2]

Examples of connectives include causal connectives {because, so, thus,

therefore), condition {if-then), purpose {to, in order to), contrast {on the one
hand, in contrast), concession {but, however), conjunction {and, comma,
also), and temporal connectives {then, next, second, following).

Note that Mary ran quickly and Mary might have run are represented
in the same way here, as a proposition RUN[MARY] modified by either
QUICKLY or POSSIBLY. This (and many similar “problems” one
could come up with) is not a weakness of the system but a source of its
strength. Quickly modifies ran, whereas might have applies to the whole
proposition, which is a valid distinction and crucial when we are con¬
cerned with the truth value of sentences. But we are not; we want to
model comprehension, to count elements in working memory, to score
recall protocols, and so on - for which purposes we do not need a logically
consistent system or a detailed semantics. We need a way to represent the
meaning of a text sufficiently independently of the words used that pre¬
serves those aspects of meaning that are important for our purposes and
glosses over those that are not. Propositions as defined here best fulfill
that need.

Although the foregoing examples do not cover every case encountered

in the propositional analysis of text, they should enable researchers to
construct simple but consistent propositional representations for many
texts. It should also be remembered that for many experimental purposes
particularly troublesome constructions can be avoided.

A sample analysis. In this section the following three sentences from a

junior-high biology textbook dealing with the functioning of the heart are
analyzed using the foregoing method. This is a detailed theoretical analy¬
sis that would be the output of an automatic parser if we had one, but for
most research purposes, a shorthand version can be used, as illustrated at
the end of this section.1

1 This section is based on Kintsch (1985). An excellent discussion of the general issues
involved in such an analysis is provided by Perfetti and Britt (1995).
 Propositional representations 61

(1) The first of the heart’s four chambers, the right atrium, receives
purplish blood, short of oxygen and laden with carbon dioxide.
This used blood arrives through the body’s two major veins, the
superior and inferior venae cavae, and from the many minute
blood vessels that drain blood from the walls of the chamber
itself.

Each sentence is divided into text elements, labeled Ei? composed of a

content word together with preceding function words. Thus, The first/of
the heart's/ . . . yields the first two text elements El and E2. Each text ele¬
ment is then annotated as to its function: For example, first modifies text
element E4, and is also the subject for the predicate E7; chambers becomes
the patient of the proposition to be formed from the verb phrase E4;
atrium is linked to El as a specification, and so on. We do this by our intu¬
itions; a parser would infer these annotations automatically. The anno¬
tated text elements are then used to construct propositions.
As each text element is read, either a new proposition is constructed
(possibly incomplete, with the missing parts here indicated by the still to
be processed text element or simply by an X - an automatic parser would
have to be somewhat more precise), or an already existing proposition is
modified. For instance, when the text element E3,four, is encountered, a
new proposition is constructed with FOUR as the predicate and an
unknown argument to be derived from E4. When E4 is processed, the
incomplete proposition is filled in with the appropriate argument, obtain¬
ing FOUR[CHAMBER], When E7, receives, is read, a new proposition
with the predicate RECEIVE is constructed that has the argument
FIRST[CHAMBER], a proposition already constructed for this purpose
at step E4. Flowever, RECEIVE needs at least one more argument, w hich
we indicate by an X. It is filled in at step E9 with PURPLE[BLOOD],
In order to construct propositions in this way, a parser must have
access to a considerable amount of lexical knowledge. For instance, it
must know that the predicate FOUR can be used with count nouns only
and that the verb frame for RECEIVE requires an object. It must also
have optional slots for a source (“from the body”), a goal (“the heart”),
and a medium (“through the veins”). The knowledge sources needed to
construct the propositions shown in Figure 3.4 are not shown explicitly
here, however (but see Kintsch, 1985).
62 The theory

Figure 3.4 A sample analysis. Text elements are shown in the first column-
the second column shows the syntactic tags required for building proposi- ’
tions; propositions under construction but still incomplete are shown in the
third column; the completed propositions are shown in the fourth column.
Propositional representations 63

Figure 3.4 (cont.)

 64 The theory

In the analysis given in Figure 3.4, Column 1 presents the complete

text decomposed into individual text elements. The semantic and syntac¬
tic roles of these elements are shown in column 2. Column 3 shows the
atomic propositions that are constructed as the reader processes the text.
For example, a reader processing E3 knows only the predicate of a propo¬
sition, but the argument is still unspecified; only when the reader reads
E4 can the proposition be completed. Completed atomic propositions are
shown in column 4. Additional arguments may be added to these propo¬
sitions as the reader acquires more information.
From the atomic propositions that have been formed (column 4), com¬
plex propositions can now be constructed by assigning modifiers to their
proper places in the propositional schema (the Circumstance slots are
empty for all the complex propositions shown in Figure 3.5).
The first two complex propositions constructed in this example are
linked by an indirect coherence relation. It is marked in the surface struc¬
ture by this (El4), which indicates that the used blood in P2 is identical
with the purplish blood in PI. The relation between the P2 and P3 is one
of direct coherence via the sentence connective and (E25). Another link is
established by the elliptic subject (used blood) of the phrase upon which
P3 is based.
Thus, the piopositional representation PI— P3 could be constructed in
an objective and straightforward manner by considering the syntactic
relations among the various text elements as well as generally available
lexical knowledge. What was done here by hand is, more or less, what a
parser would have to do. Having such a parser would be most desirable,
but not having one need not stop our research efforts.
In practice there is no need to go through this detailed step-by-step
analysis. The step-by-step, real-time construction of the textbase is often
not of interest. All we need to know (for instance, in studies of text recall)
is the final product, the representation in terms of complex propositions.
Flence, it is sufficient for most research purposes to construct the final
propositional representation, and its genesis can be neglected.

3.1.2 The macrostructure

Texts have a global structure as well as a local structure. The microstruc¬

ture of a text consists of the complex propositions that comprise the text
Propositional representations 65

Figure 3.5 The three complex propositions constructed from the analysis in
Figure 3.4.
66 The theory

and their interrelationships, for example, the three propositions P1—P3 in

the example analyzed earlier. The macrostructure organizes the proposi¬
tions of the microstructure hierarchically. It consists of macroproposi¬
tions and their hierarchical relations. Macropropositions may or may not
be expressed explicitly in the text. They are related to the microstructure
by three rules, called macrorules by van Dijk (1980):

• Selection: Given a sequence of propositions, propositions that are not

an interpretation condition for another proposition may be deleted.
• Generalization: A proposition that is entailed by each of a sequence of
propositions may be substituted for that sequence.
• Construction: A proposition that is entailed by the joint set of a
sequence of propositions may be substituted for that sequence.

An example for the operation of selection may be obtained from the

analysis in section 3.1.1. The macroproposition of the text fragment ana¬
lyzed there may be obtained by deleting P2 and P3 as well as inessential
elements of PI, as follows:

PI

P2 Ml RECEIVE
- RIGHTf ATRIUM]
P3 B ,OOD
“ SHORT-OF[BLOOD,OXYGEN]
_ LADEN-WITH[BLOOD,CARBONDIOXIDE]

An example for the operation of generalization would be

RIGHTf ATRIUM] ~
RIGHT[VENTRICLE] CHAMBER
LEFT[ATRIUM] 1— OF-HEART
LEFT[VENTRICLE] —

An example for a construction would be

EXPAND[MUSCLE]

|= OF-HEART -» PUMP[HEART]
CONTRACT[MUSCLE]
 Propositional representations 67

The macrooperations may be performed recursively on macropropo¬

sitions as well as on micropropositions, resulting in a hierarchical struc¬
ture, such as is shown in Figure 3.6.
Macrostructures thus represent the global organization of a text. As
such, they are linguistic structures, like the phrase structure of a sen¬
tence. For the psychologist, the important questions concern the psycho¬
logical reality of such structures, as well as how they are formed during
comprehension (for the macrorules operate on linguistic descriptions
and are not in themselves a psychological process model).
As to the psychological reality of macrostructures, there exists over¬
whelming evidence of their significance (e.g., van Dijk & Kintsch, 1983).
Indeed, for comprehension and memory, the gist of a text - expressed
formally by the macrostructure - is usually what matters most. As has
been described by van Dijk and Kintsch (1983), a number of interacting
factors are crucial for the formation of macrostructures. These include
general cultural knowledge, including the sociocultural context of the
comprehension episode - its situation type (e.g., a party conversation vs.
a court trial vs. reading a chapter in a textbook), the participant categories
(e.g., a teacher, a friend, their interests, gender, social status, and so on),
the type of interaction (e.g., a tutorial dialogue vs. a chat in the grocery

Figure 3.6
68 The theory

store), conventions, and habits. There are, however, also rich signals in
the texts themselves that enable the formation of macrostructures. These
may be structural signals, such as titles, initial topic sentences, summary
statements, and the like, as well as a great variety of syntactic and seman¬
tic signals that are mainly used to indicate local importance in a discoui se
but that may achieve macrorelevance through cumulative inference. Text
elements that have beeo repeatedly signaled to be of local importance
become important for the macrostructure, too. Syntactic signals for dis¬
course relevance include phonetic stress, cleft sentences, passives, the
clause structure of sentences, and other foregrounding and topicalization
devices. Semantic signals in a text include topic change markers (such as
change in time or place, introduction of new participants, change of per¬
spective, and so on) and various cues that indicate local coherence or a
break thereof that subjects can use either to maintain or to change their
current macrohvpothesis (Givon, 1995; van Dijk & Kintsch, 1983).
Textual schemas, what van Dijk and Kintsch (1983) have called the
rhetorical superstructures, also play an important role in the formation
of macrostructures. Readers know that particular text types tend to be
organized in certain ways and employ this knowledge to construct
schema-based macrostructures. Thus, narratives in our culture have a
basic exposition—complication—resolution structure, with the possibility
of embedding. By assigning macropropositions to these schematic cate¬
gories, readers can simplify for themselves the task of deriving the over¬
all organization of a story. Indeed, as Poulsen, Kintsch, Kintsch, and
Premack (1979) have shown, even four-year-olds make good use of this
schematic knowledge in story understanding. Conversely, when the
schema is misapplied because the narrative is constructed according to
some other cultural schema, comprehension may be distorted accord¬
ingly (Bartlett, 1932; Kintsch & Greene, 1978).
Rhetorical schemas play a role in understanding descriptive texts that
is as important as their role in understanding stories. If familiar schemas
are present, comprehension is facilitated but only at the macrolevel, not
locally. For instance, Kintsch and Yarborough (1982) wrote brief essays
according to four familiar rhetorical schemas: classification, illustration,
compare-and-contrast, and procedural description. The texts conformed
as closely as possible to these schemas. The texts were then rewritten, and
although their schematic structure was destroyed their content remained
unchanged. This was done by a partial reordering of sentences and the
 Prop o sit ion a l rep re sen tatio ns 69

deletion of explicit rhetorical signals from the original texts. Subjects

performed much better on topic and main idea questions, which require
good understanding at the macrolevel, when they read the rhetorically
well-structured texts (.62 and .43 correct in one experimental condition)
than when they read the same texts without the rhetorical cues (.23 and
.05 correct, respectively). On the other hand, on a test that evaluated their
local Understanding (a cloze test), there was no significant difference
between the rhetorically good and poor text versions (.33 and .39 correct,
respectively, for the good and poor versions).

3.1.3 The psychological reality of propositions

Why do we need propositions at all for our theorizing, what is wrong with
just plain natural language? In doing research on text understanding it is
frequently the case that propositional representations of the texts are
more readily related to the behavioral data than are the texts themselves.
Presumably this is the case in those situations in which meaning matters
most; propositions are designed to capture those semantic relations that
are most salient in text comprehension, whereas natural language serves
many purposes other than the expression of meaning and hence is often
less suited for our purposes than a representation that is focused on
meaning. Propositions appear to be the semantic processing units of the
mind and hence the most useful form of representation for our studies.
What is the empirical evidence that propositions are in fact effective
units in text comprehension? Because complex propositions are defined
in such a way that they (generally) correspond to sentences, all the evi¬
dence that shows that sentences are psychological units also supports the
notion of complex propositions. But sentences are structured according
to the rules of syntax, whereas the internal structure of complex propo¬
sitions is quite different and only indirectly related to syntax. Complex
propositions have a central proposition and various associated elements -
modifiers of various sorts and/or circumstantials. The question then
becomes whether these smaller propositional elements, or atomic propo¬
sitions, can be shown to function as psychological processing units. What
is the evidence that atomic propositions are the functional processing
units of text comprehension?
A discussion of the psychological reality of propositions can be found
in most contemporary cognitive psychology textbooks as well as in van
70 The theory

Dijk and Kintsch (1983, chap. 2). Hence, no more than a brief summary
is presented here. In a sense, every experimental study that successfully
employs a propositional analysis either to design the experimental mate¬
rials or to analyze the data provides evidence for the usefulness of propo¬
sitional analyses. However, one would like more direct evidence as to the
psychological reality of propositions. Such evidence comes from three
sources: recall experiments, reading times, and priming studies.

Cued and free recall. If subjects study sentences that are based on more
than one proposition and then are given a word from the sentence as a
recall cue, words from the sentence that belong to the same atomic propo¬
sition as the cue word are recalled better than words from the sentence
that do not come from the same atomic proposition. This result has been
obtained in several studies, (e.g., Wanner, 1975). Thus, consider the two-
element sentence that consists of a main proposition and a propositional
modifier

(2) The mausoleum that enshrined the tzar overlooked the square.

OVERLOOK[MAUSOLEUM, SQUARE]
1-ENSHRINEfMAUSOLEUM, TZAR1

Subjects who are given the word overlooked as a recall cue will more often
recall square than tzar, which are arguments of the same atomic proposi¬
tion, in spite of the fact that tzar is closer in the surface structure to the
recall cue than square.
Barshi (1997) studied how people were able to execute verbal instruc¬
tions that differed in the number of propositions and/or words. His sub¬
ject listened to messages that instructed them to move in a simulated
three-dimensional space on a computer screen. They repeated these
instructions orally and then attempted to follow them - as airline pilots do
when receiving real air traffic control instructions. The instructions were
given either in complete English sentences (“move two squares to the
left”) or in abbreviated form (“two left” - an unambiguous command in
the context of Barshi’s experiment). The number of propositions a subject
had to remember and respond to mattered greatly: Doubling the number
of propositions from two to four caused an increase in errors from 3% to
52%. The number of words used to express these propositions, on the
other hand, had no effect at all in Barshi’s experiment: A two-proposition
instruction was as hard to remember and execute whether it was expressed
 Propositiona l rep re sen tatio ns 71

with four words or with eight words. But there was a striking difference in
difficulty whether eight words expressed two propositions or four.
There are also a number of studies that demonstrate that when recall¬
ing sentences based on complex propositions, subjects tend to recall the
various elements of the complex proposition as units. That is, they either
recall the whole element or omit it entirely. The most extensive of these
studies is one by Goetz, Anderson, and Schallert (1981). In this study
subjects read eight sentences and then recalled them. The sentences were
either based on a single propositional element, such as

(3) The customer wrote the company a complaint.

WRITE[CUSTOMER, COMANY, COMPLAINT]

or a proposition with two additional elements, such as

LECTURE[PROFESSOR]
1— FAMOUS
— LOC: CLASSROOM

(4) The famous professor lectured in the classroom.

Almost all the free recall of these sentences, 94%, consisted of complete
propositional elements. That is, if a subject recalled anything at all from
sentences like (3), the whole sentence was recalled. In contrast, sentences
like (4) were often recalled partially, that is, subjects may recall The pro¬
fessor lectured. The professor was famous or The famous professor was in the
classroom, for instance. However, Goetz et al. argued that the holistic
recall they observed may not have been owing to the fact that subjects
represent the meaning of sentences in memory propositionally, but that
they may be a consequence of the familiarity, concreteness, and high
imagery value of these units. Therefore, they constructed similar sen¬
tences in which unlikely elements were combined and that would be
much harder to image, such as

(5) The comedian supplied glassware to the convicts.

(6) The bedraggled, intelligent model sang.

Their results, however, remained the same. As before, the overwhelming

majority of recall, 84% this time, consisted of whole propositional ele¬
ments. Thus, the unfamiliar, low-imagery sentences (type 5) were recalled
72 The theory

in an all-or-none fashion, much like type (2) sentences, whereas, sentences

of type (6) broke down into their propositional elements on recall.

Reading times and recall. In a number of early studies, sentence recall was
found to be a function of the number of propositional elements in the
sentence when reading time was controlled. Thus, Forster (1970) pre¬
sented six-word sentences at a constant rate. All sentences were six words
long but contained either one or two propositional elements:

(7) The kitten climbed over the fence.

(8) The truck Susan was driving crashed.

The average number of words recalled for sentences of types (7) and (8)
were 4.41 and 3.09, respectively.
Conversely, if reading time is subject controlled, subjects take longer to
read propositionally complex sentences, even when the number of words
is the same. Kintsch and Keenan (1973) constructed sentences that were
all of approximately the same length (14—16 words) but that contained any¬
where from four to eight propositional elements. Of course, the mere fact
that a sentence contains many propositional elements does not guarantee
that a reader constructs an equally complex representation. A fast reader
may simply skip over some propositional elements. Hence, it is inappro¬
priate to plot reading times against sentence complexity. However, if one
plots reading time against the number of propositional elements that sub¬
jects recall on a free recall test, a rather striking approximately linear rela¬
tionship results. For each additional propositional element recalled, read¬
ing times increased by 1.5 s in the Kintsch and Keenan study.
One needs to be cautious about generalizing this finding, however.
Kintsch and Keenan examined only a very limited range - between four
and eight propositional elements. It is not really possible to extend this
range, because as sentences or texts get longer, retrieval and forgetting
problems complicate the interpretation of the data. Even more critically,
when sentences are constructed that vary in propositional complexity but
not in length, these sentences necessarily also vary in many other ways —
lexical items, syntax, familiarity, the number of repeated concepts, and
so on.
Since these early studies, reading times have been analyzed repeatedly
using multiple regression analyses in the attempt to untangle these con-
foundings. In these studies, propositions in the text, not propositions in
the head (as reflected by subjects’ recall) have invariably been used as the
 Prop o sit ion a l rep re sent at io ns 73

variable of interest. Hence, these studies are not entirely relevant for the
question at hand — whether or not propositions are meaningful psycho¬
logical units of comprehension. Nevertheless, considerable support for
the usefulness of this notion can be derived from these studies (which are
reviewed in section 9.1.2).

Pnmirtg studies. Additional support for the psychological reality of propo¬

sitions comes from an experiment on recognition priming reported by
Ratcliff and McKoon (1978). These authors showed their subjects four
sentences for 7 s each. The sentences contained either one or two propo¬
sitional elements, such as sentence (2) above. After reading a set of sen¬
tences, the subject was given a recognition test with single words from the
sentences and unrelated distractor words. Subjects were required to
answer “yes” or “no” whether a word had appeared in one of the study
sentences or not. Priming effects are commonly observed on such a task:
When two words from the same sentence are tested one after the other,
recognition time for the second word is lower. In the Ratcliff and Mc¬
Koon experiment, this priming effect amounted to 111 ms when the two
words were both from the same sentence and the same atomic proposi¬
tion (square and mausoleum), but to only 91 ms when the two words were
from the same sentence but from different atomic propositions (e.g.,
square and tzar). The difference is small but statistically significant. It
indicates that not only sentences (and complex propositions) are psycho¬
logical units but that this is also true for the elements of complex propo¬
sitions (atomic propositions).
Atomic propositions, therefore, appear to be the proper building
blocks for comprehension structures. They have proven their usefulness
in numerous studies, and they will do so again in the experiments and
analyses reported here.

3.2 The propositional representation of knowledge

Propositional text representations are helpful for describing the semantic

characteristics of experimental texts, and they are useful for scoring recall
and other data from experimental subjects. Furthermore, they are widely
used for modeling purposes. Although they may not be ideal for every
purpose, they have proven to be a reliable workhorse. In contrast, propo¬
sitional knowledge representations remain largely untried. As I have
74 The theory

argued in chapter 2, propositional knowledge nets seem quite appropri¬

ate for the purpose of simulating knowledge use in the comprehension
model, but the simulations that have been done to date employ fairly
small-scale, ad hoc knowledge bases. No large-scale, general knowledge
base using the knowledge net format has been constructed yet, and it is
impossible to know beforehand how adequate the proposed format would
turn out to be. Given the problems that seem to be encountered univer¬
sally in the construction of large-scale knowledge representations, it
would not be surprising to find knowledge nets as envisaged here to have
some shortcomings, too. For our present, more modest purposes of sim¬
ulating knowledge use in comprehension in special cases, knowledge nets
appear to be quite satisfactory, however. I discuss in this section how two
of the central problems in knowledge representation might be dealt with
within such a framework. Of course, this discussion cannot substitute for
an actual simulation of a large knowledge net, in which the processes I can
only speculate about are realized computationally.

3.2.1 The construction of meaning: Concepts

In the model proposed here, knowledge is represented as a network of

propositions. Such a network is called a knowledge net. The nodes of the
net are propositions, schemas, frames, scripts, production rules - which
can all be written in a formalism based on the predicate-argument
schema, as argued in chapter 2. The links are unlabeled and vary in
strength, that is, a knowledge net is a type of associative net.
The meaning of a node is given by its position in the net, that is, by the
strengths with which it is linked to its neighbors, immediate ones as well
as neighbors many steps apart. This definition of meaning is an abstract,
linguistic one, not a psychological one. Psychologically, only those nodes
that are actually active (that is, are held in working memory) contribute
to the meaning of a node. Because the capacity of working memory is
severely limited, any node at any point in time has only a few neighbors;
its meaning is sparse, therefore. However, it can be readily elaborated,
almost without limit, in many different directions, as the situation
demands, because most nodes in a knowledge network are connected with
powerful, stable links - retrieval structures - to other nodes in the net that
can be brought into working memory. Thus, very complex meanings can
be generated automatically and effortlessly, although at any particular
time only a few nodes can be active in working memory.
 Propositional representations 75

Concepts do not have a fixed and permanent meaning. Rather, each

time a concept is used, its meaning is constructed in working memory by
activating a certain subset of the propositions in the neighborhood of a
concept node. The context of use determines which nodes linked to a
concept are activated when a concept is used. Goals, prior experience,
emotional state, situational as well as semantic context all influence which
nodes are activated and hence what the meaning of the concept will be on
this particular occasion.
In constructing the meaning of a concept, the concept node and any
other currently active proposition in working memory serve as the re¬
trieval cues, both individually and as parts of a compound cue. Hence,
what will be retrieved in the process of elaborating the meaning of a node
will depend not only on the node itself but on the state of working mem¬
ory as a whole.
The substructure from which the meaning of a concept is con¬
structed - the knowledge net - is relatively permanent (experience and
learning create and continually modify this structure). The meaning -
that is, the portion of the knowledge net that is activated - is flexible,
changeable, and temporary, however. Because meaning construction is
based on the same substructure, there will be a certain amount of consis¬
tency in the meaning of concepts on different occasions. The likelihood
that certain meaning elements will be sampled will always be greater than
for other elements, but the context in which this sampling occurs will
ensure a great deal of variability in the outcome.
Knowledge nets have some advantages over alternative forms of knowl¬
edge representation. First, there are good psychological data that argue for
the psychological validity of each one of the knowledge structures men¬
tioned, from associative nets to production systems, as was described in
chapter 2. But at the same time, there are equally good data that show that
neither of these systems is sufficient by itself for the representation of
knowledge. For instance, one can show, as Collins and Quillian (1969) did,
that certain psychological predictions that can be derived from semantic
nets can be verified experimentally, but as an army of their detractors have
demonstrated and as we can read in every cognitive psychology textbook,
these predictions are wrong in many ways because there is much more to
human knowledge than a semantic net. A knowledge system must account
for the inheritance of (some) properties, but it also must include schema¬
like structures with default slots and procedural knowledge that link cog¬
nition and action, and so on. In such a net one can make distinctions
76 The theory

between episodic and semantic memory, or procedural and declarative

knowledge, but it is not always necessary to do so. These are merely dif¬
ferent types of nodes in the same network, interacting with each other.
Knowledge nets, therefore, freely combine features from other knowledge
representation systems that have been shown to be useful for computa¬
tional purposes in AI research or to be valid in psychological experiments.
Consider a particular pode, P, in such a net. It may be linked to a set of
nodes, P,; the Pj nodes in turn are linked to another sent of nodes, P7; and
so on in ever increasing concentric circles until the whole network is
included. To say that the meaning of P is the relation between P and these
concentric sets of nodes is true in some abstract sense but beside the point
psychologically. At any given moment, in any given knowledge net, only
those nodes contribute to the psychological meaning of P that are actu¬
ally activated in working memory and linked to it. Thus, typically, instead
of a very large set of nodes, only a limited number of nodes constitute the
effective meaning of P at any time, perhaps only five or six nodes in the
case when the meaning of P is only superficially elaborated. However,
because P is embedded in a network of strong stable relationships with
other nodes in the knowledge net, further elaboration via retrieval struc¬
tures is readily achieved should there be a need for it. I may think of a few
things concerning P and you may think of a few things, and there may be
a little overlap and a miscommunication may result. However, if there is
some context to guide us, we are more likely to construct similar mean¬
ings for P and communicate effectively.
Knowledge nets thus imply a commitment to a radical constructionist
position in the controversy about the mental representation of word
meanings. In a mental lexicon, one looks up the meaning of a word. In a
knowledge net, there is nothing to look up. Meaning has to be con¬
structed by activating nodes in the neighborhood of a word. This activa¬
tion process is probabilistic, with activation probabilities being propor¬
tional to the strengths of connections among the nodes, and it may
continue for a variable amount of time, spreading outward into the
knowledge net from the source node. The meaning of the source word is,
then, the set of activated nodes in the knowledge net.
The knowledge net serves as a retrieval structure in the sense of Erics¬
son and Kintsch (1995; also chap. /). If any element of a knowledge net is
in working memory (focus of attention, consciousness), other elements
directly connected with it can be retrieved with a single 400 ms retrieval
 Propositio nal rep re sen tatio ns 77

operation. These directly retrievable elements make up what Ericsson and

Kintsch called long-term working memory. Once a particular element
from long-term working memory has actually been instated in the focus
of attention, it, too, will provide access to its neighbors in the knowledge
net, thereby further increasing the size of long-term working memory. In
addition, the pair of nodes will function as a compound cue, markedly
changing the retrieval probabilities in long-term working memory.
Words, in this view, have a potential meaning given by concentric shells
of ever expanding neighborhoods in the knowledge net. The most restric¬
tive potential meaning of a concept would be given by its immediate
neighbors, the most complete by the total net. There is no sense in asking
how many steps away from a concept the expansion has to go to give us
“the meaning” of a concept. Meaning may be more or less elaborate. And,
most important, this expansion process provides only a potential mean¬
ing. The real, actual meaning of a word is not the set of all nodes that
might be activated in long-term working memory but rather the nodes
that have actually been activated in the particular context of use. Thus,
meanings are not nearly as elaborate as they could be, because normally
only an insignificant fraction of a concept’s neighboring nodes in a
knowledge net enters consciousness (though many more are readily avail¬
able in long-term working memory). A linguist, semanticist, or psychol¬
ogist studying the meaning of a concept will come up with a very rich and
complex structure. That is not, however, what is actually operative when
that word is used on specific occasions in a specific context, where mean¬
ing is much more sketchy and incomplete.
Contextual word meanings are not only shallow but are dynamic and
fluctuating. Somewhat different word meanings are constructed on dif¬
ferent occasions, even if the knowledge net and discourse context remain
the same, simply because of the probabilistic nature of the sampling pro¬
cess that determines which of the many possible knowledge elements
actually enter consciousness. But the discourse context is in continuous
flux, and different persons operate with different knowledge nets. Hence,
there must be considerable variability in effective word meanings.
Before continuing this discussion on the construction of meaning, it
seems only proper to ask whether there is any psychological evidence that
would support such a theory. Indeed, there is. In fact, the psychological
evidence overwhelmingly favors the view that concepts are temporary
constructions in working memory, generated in response to task demand
78 The theory

and subject to the constraints exercised by the underlying knowledge base

and the situational context. Cognitive scientists might as well discard the
traditional notion that concepts are stable entities to be retrieved from
long-term memory - a view that we have inherited from philosophy and
linguistics.
Barclay, Bransford, Franks, McCarrell, and Nitsch (1974) were the
first to point out the role of encoding variability in memory retrieval.
They gave to subjects words like “piano” to study in the context of play¬
ing music or moving furniture. On a later memory test, they gave either
“loud” or “heavy” as a retrieval cue. The former was a better retrieval cue
when “piano” had been presented in the music context, but the latter was
the better cue when “piano” had been studied in the furniture context,
suggesting that a context-specific concept of “piano” was encoded.
In several studies that demonstrate the flexibility and context-depen¬
dency of concepts, a sentence verification paradigm was used. A repre¬
sentative experiment is that of McKoon and Ratcliff (1988). They argued
that two equally known properties of tomatoes are that they are round and
red. Nevertheless, the actual availability of these properties strongly
depends on the context in which tomato was used. Following a brief para¬
graph about painting a still life containing a tomato, Tomatoes are red was
verified faster than Tomatoes are round. But after a paragraph describing
a child rolling tomatoes around the floor, Tomatoes are round was verified
faster. The average difference between matching and nonmatching sen¬
tences was 120 ms. In a neutral context that emphasized neither color nor
shape (about eating tomatoes), both target sentences were responded to
equally fast.
Kintsch and Welsch (1991) have shown that these results follow
directly from the assumptions about meaning construction that have
been discussed. They simulated the results of the McKoon and Ratcliff
experiment using the construction-integration model to be descrtibed in
the next chapter. Memory strength in this model is a function of the
amount of activation a node attains. In Figure 3.7 the results of the sim¬
ulation are shown: a fragement of the network representing the texts, as
well as the two test items, Tomatoes are round and Tomatoes are red, with
theii activation values. For the “playing-with-tomatoes” text, more acti¬
vation flows to Tomatoes are red than to Tomatoes are round. On the other
hand, in the context of the painting-tomatoes” the opposite will be true.
In the context of the eating-tomatoes” text, the activation values for the
two test sentences turn out to be equal.
 Propositional representations 79

targets:

RED [TOMATO] ROUND [TOMATO]

1 31

long-term memory: \ \ II

targets:

RED [TOMATO] ROUND [TOMATO]

7 4
i r
/ /
long-term memory: i i

Figure 3.7 Part of the long-term memory network after reading the play-
ing-with-tomatoes (top) and painting-tomatoes (bottom) paragraph and two
target sentences. The numbers below the nodes in the network are long-term
memory strengths; the numbers below the targets are the activation values
they receive from the net.

A number of other studies demonstrating the context dependency of

concepts have been reviewed by Barsalou (1993), who advocates a view
related to the present one. Of particular significance is one of his own
studies that he discusses. In this experiment, subjects were asked to write
down features that define common categories. Only 44% of the features
in one subject’s description existed in another subject’s description, indi¬
cating that the definitions subjects provided were highly idiosyncratic.
Indeed, when subjects were asked to provide definitions for the same con¬
cepts on two successive occasions, their own overlap was merely 66%.
80 The theory

Thus, not only are concepts idiosyncratic, but they are also highly unsta¬
ble. However, Barsalou was able to show that only the concepts that sub¬
jects constructed in this underconstrained experimental situation were so
unpredictable. If subjects were given all the features generated by every¬
one in the experiment, they agreed very well that these were indeed fea¬
tures characteristic of the concept in question (97% agreement between
subjects, 98% within ^sqbjects). Everyone had more or less the same
knowledge base when it came to these familiar everyday concepts. Fur¬
thermore, in a more constraining context, different subjects tended to
agree much better on the features that characterize a concept. Thus, in
the experiment reported by Barsalou, intersubject agreement was 45%
when subjects defined concepts like vehicle in isolation, but rose to 70%
with even a minimally constraining context (“Taking a vacation in the
rugged mountains of Mexico”).
Barsalou’s data nicely illustrate the fact that, although concepts are
fleeting and flexible, not all is chaos, because the knowledge bases from
which these concepts are constructed are more stable and predictable,
and most of the time the context itself will be sufficiently constraining to
ensure that the concepts different people form will be similar. Neverthe¬
less, concepts are never quite the same - surely a limiting factor in com¬
munication.
The theory of meaning advocated here is not only constructivist but
also minimalist. Clearly, readers can study a text over and over again and
construct very elaborate meanings for its propositions and concepts. Lin¬
guists, philosophers, and literary critics do this all the time, and most peo¬
ple do so at least some of the time. But most of the time, in reading or con¬
versation, the process of meaning construction remains shallow, not just
because comprehenders are inherently lazy but mostly because no more is
required. A slight knowledge elaboration of a text is usually quite sufficient
for whatever action is intended. Most of the time texts do not need much
elaboration and interpretation to arrive at stable interpretations upon
which appropriate responses can be based. Long-term working memory
allows the comprehender easy elaborations and inferences whenever they
are required. It is enough for the well-informed reader to feel that the
potential for the elaboration of meaning is there - there is no need to real¬
ize it. It is of course possible to do so, and we often do so, sometimes read¬
ily, sometimes with the expenditure of considerable effort. Indeed, the
deliberate construction of meaning may extend over long periods and may
 Propositional representations 81

be a socially shared activity, as is obvious, for instance, in the case of texts

that have special cultural significance, such as the Bible.
This is not a conception of meaning that will make logicians happy.
What is true, what is a contradiction, if meaning is subjectively con¬
structed in specific contexts? Logicians found it necessary to invent their
own sense of meaning. They had to invent logic precisely because the
everyday sense of meaning is useless for precise reasoning. Logic in its
various forms is an extremely successful, well-developed, and useful sys¬
tem. But it is a system that was invented to make precise reasoning pos¬
sible, not to describe or simulate human cognition. For that we need a
very different kind of system, one that is useless for the logician or for¬
mal semanticist but that meets the needs of the researcher who is inter¬
ested in describing how humans think, comprehend, decide, and act.
If the meaning of words must be constructed in their context, the dif¬
ference between literal and metaphorical, or idiomatic, word meanings is
minimized. Both involve constructive processes, and there is no reason to
suppose that one kind of construction is necessarily prior or more difficult
than the other. Consider the following examples (from Kintsch, 1989):

(9) The cat sat on the mat.

(10) He let the cat out of the bag.

Out of context, there is not much to the understanding of (9): A proposi¬

tion SIT[CAT,ON-MAT] must be formed, and some more or less dys¬
functional associations will be activated, such as cats purr, my cat is black,
or philosophical argument. To understand (10) literally, a proposition
LET[HE,CAT,OUT-OF-BAG] must be formed, and once again, some
random associations having to do with carrying cats around in bags may
be activated. To understand (10) as an idiom, the same proposition is
formed, but this time it is embedded in a different set of associations hav¬
ing to do with betraying secrets and surprising revelations of some sort or
another. Just how the right set of associates is selected in each case - that
is, how the meaning of the phrase is actually constructed in a discourse
context - is within the purview of the construction-integration model to
be described in the following chapter. Out of context, all three construc¬
tions are quite trivial and (except in a linguistic or philosophical discus¬
sion) remain superficial. In context, they may be optionally elaborated,
depending on the particular context. However, there are no distinct lit¬
eral or nonliteral processing modes, and it takes people about equally long
82 The theory

to come up with a literal or nonliteral interpretation for such sentences

(Glucksberg, Gildea, & Bookian, 1982).

3.2.2 Emergent structures

One of the most salient differences between knowledge nets and most
other forms of knowledge representation is that knowledge nets are asso¬
ciative, hence based on perception and experience, and therefore relatively
unorganized and chaotic. Semantic nets, scripts, frames, and the like are
all attempts to make knowledge organization more orderly, more logical.
Clearly, knowledge can be organized logically; there is abundant evidence
for the psychological reality of higher-order schematic structures such as
scripts and frames. Why, then, has an associative net been chosen as the
basis for representing knowledge? And how can one account for the obvi¬
ous fact that people operate with schematic structures all the time if such
structures are not components of the basic knowledge representation?
When psychologists and AI researchers began to use scripts and frames
as knowledge units, they viewed them as preexisting structures in mem¬
ory that are retrieved upon demand and used for a wide variety of pur¬
poses. This view soon came into conflict with the available psychological
evidence as well as the computational demands of AI systems. Computa¬
tionally, scripts and frames proved to be too inflexible to serve the purposes
for which they were originally designed (Schank, 1982). Scripts have to be
applied in ever changing contexts, and various ways of fine-tuning scripts
proved to be inadequate to provide the flexibility that was needed. For
instance, attempts to fine-tune the restaurant script by distinguishing var¬
ious tracks (fancy restaurant, Chinese restaurant, fast-food restaurant)
only led to an endless proliferation of subscripts. Empirically, too, there
were problems. Although certain aspects of scripts (differences in the cen¬
trality of items in a script - Bower, Black, & Turner, 1979; the direction¬
ality of scripts - Haberlandt & Bingham, 1984) could be substantiated by
psychological experiment, others could not. Primarily, the distance
between events in a script did not behave as it should if scripts were fixed
linear structures: The time it takes to retrieve an event given another script
event as a cue does not vary proportionally to the distance between the
events (Bower et ah, 1979; Haberlandt & Bingham, 1984.)
Schank (1982) therefore modified the original script notion by intro¬
ducing smaller units called Memory Organization Packets (MOPs), out
of which scripts can be constructed in contextually more appropriate
 Propositional representations 83

ways. A different but related proposal was made by Kintseh and Mannes
(1987), who showed that scripts and related structures can be constructed
out of associative nets. Scripts are not fixed, fully elaborated, preexisting
knowledge structures that merely need to be retrieved for use. Instead,
only the raw material for constructing a script is part of the knowledge
net: a script-proposition that can function as a frame for other, contextu¬
ally appropriate, associated information in the knowledge network.
Scripts and other types of schemas are emergent structures, and because
they always are constructed in the context of their use, they are well
attuned to that context.
Kintseh and Mannes (1987) start from the observation that the retrieval
of the items of a script occurs in quite a different way from the retrieval of
the items of a conceptual category. If subjects retrieve instances of some
conceptual category, such as cars, a plot of the number of items retrieved
against time yields a scalloped, negatively accelerated curve as in Figure
3.8 (Bousfield & Sedgwick, 1944; Walker & Kintseh, 1985).
Typically, a subject produces a burst of category items in response to a
retrieval cue that he or she is using, pauses, and then produces another
burst. Such retrieval cues may be “Japanese cars,” “cars in my dormitory
parking lot,” “cars my parents owned,” “cars I have wrecked,” and the
like. Each retrieval cue works for a while, rapidly retrieving some number
of items that are closely associated with it, but then dries up. What hap¬
pens during the pause is that the subject is searching for a new retrieval
cue when the old one has become ineffective. Finally, the subject is unable
to generate new, effective retrieval cues and quits. The results of this
process are the seal loped curves, steep at first and then flattening out, as
shown in Figure 3.8.
If a subject retrieves the items of a script, such as going to a grocery
store, a smooth function increasing at a constant rate is obtained, as in
Figure 3.9 (Kintseh & Mannes, 1987). This pattern of retrieval might be
explained by assuming that a script is retrieved and the items are read off
from it at a more or less constant rate until the end is reached. However,
the protocol data analyzed by Kintseh and Mannes suggest a different
interpretation. The retrieval of items is associative and occurs in bursts,
very much as in category retrieval. The difference is that the subject does
not have to engage in the time-consuming task of generating a new
retrieval cue when an old one becomes ineffective, because the next
retrieval cue is readily available. It is provided by a script proposition that
is part of the knowledge net. Such a proposition is not a full script but
84 The theory

Figure 3.8 A representative retrieval function for one subject in a category

retrieval task. Time is plotted in seconds. From Kintsch and Mannes (1987).

TIME

Figure 3.9 Number of grocery store items retrieved by a single subject as a

function of time. Time is plotted in 30 sec. intervals. From Kintsch and
Mannes (1987).

merely a skeleton script - an ordered list of episode names. For the GRO¬
CERY-SHOPPING script this would be as shown in Figure 3.10. Each
episode name can serve as a retrieval cue. Retrieval of items associated
with the retrieval cue occurs associatively, just as in the case of category
retrieval, except that the associations among the elements that comprise
a script are often unidirectional. Going through the aisles is associated with
 Propositional representations 85

Figure 3.10

going to the checkout counter, but there is no link in the reverse direction.
Hence, the retrieval process will continue to move through the script
without backtracking. The particular items that will be retrieved in each
retrieval episode will vary in this probabilistic process and will be
strongly influenced by whatever other retrieval cues are operative in the
situation. But there will be no scallops as in category retrieval, for there is
no need to pause to search for a new retrieval cue. The next retrieval cue
is provided by the script proposition, and it is just as effective as the pre¬
vious one, so that there is no general slowing down as in Figure 3.8.
Kintsch and Mannes (1987) have described a simulation that shows
how scripts can emerge in this way from an associative knowledge base,
given basic information about the temporal sequence of the scriptural
episodes. For this purpose, they combined the model of memory retrieval
proposed by Raajmakers and Shiffrin (1981) with the processes for cate¬
gory and script retrieval that have been discussed here. They obtained the
scalloped, negatively accelerated retrieval functions typical for category
retrieval and the smooth, constant-rate functions typical for script gener¬
ation from the same associative net. Thus, they have demonstrated that
86 The theory

global structures such as schemata need not be in the mind but may
emerge from the action of certain types of control processes on a globally
unstructured, only locally linked knowledge base.

3.3 Latent semantic analysis: a vector representation

Propositional representations of text have proven their usefulness in

research for more than 20 years. Propositional representations of knowl¬
edge, as explored earlier in this chapter, may prove to be equally useful.
However, both have a weakness that is even more serious in practical appli¬
cations than in research: We cannot construct them automatically and
must rely on hand coding (even though such coding can be reasonably reli¬
able and objective). Hence, they are difficult or impossible to use in really
large applications. It is fine to analyze propositionally a brief text for an
experiment or a simulation or to construct an illustrative knowledge net to
test some empirical implication of a simulation. But one cannot proposi-
tionalize a whole textbook or all the knowledge of the student studying it.
There is, however, another way to represent meaning that is related con¬
ceptually to propositional representations and that is not subject to these
limitations. It involves a switch in thinking about propositions not as
nodes in a network but as vectors in a high-dimensional semantic space.
The meaning of a proposition or concept in the abstract is given by its
place in a knowledge net. The meaning of a proposition or concept in a
discourse context is given by its position in the network representing that
discourse, enriched with information retrieved from the knowledge net.
Thus, CAT as well as CHASE[DOG, CAT] are defined by the nodes to
which they are linked in a person’s knowledge net. There will be some
overlap between these nodes but also some differences. For example,
SCARED[CAT] will be linked only weakly to CAT but strongly to
CHASE[DOG,CAT],
In fact, the labels CAT and CHASE[DOG,CAT] are superfluous; we
could equally well denote the two nodes in the network as Px and PN. We
do not do so because we would get confused rapidly if we were to use
such a denotational system. English language labels are much easier to
remember. But what is important is the pattern of link strengths to neigh¬
boring nodes in the network.
b urthermore, the graphical notation of a node linked to surrounding
nodes in a network is not essential either. It is equivalent to a vector rep¬
resentation in which each row and column corresponds to the nodes of
 Propositional representations 87

the network and the entry in the 2th row andyth column is the strength of
the link between nodes i and j.
A concept or proposition can thus be thought of as a vector of num¬
bers, each number indicating the strength with which the concept or
proposition is linked to another concept or proposition.
What determines these numbers? Presumably; they are the end prod¬
uct of lifelong experience, of interacting with the world we live in. We
learn - by observation, by talking to others, by reading stories - that cats
are not scared normally but that they are when chased by a dog. The val¬
ues in our vectors, therefore, are the fine-tuned products of numerous
and diverse experiences that we as humans have.
There appears to be no way one could provide an artificial system
with these numbers by some sort of hand coding. The system is too com¬
plex and too huge and too subtle, and unavailable to reliable introspection.
The only way to acquire these numbers is to live a normal human life,
learning through interaction with the human environment. A machine has
no chance to learn all these numbers perfectly, because it cannot live and
act like a human - an obvious point, much belabored by philosophers.
Does that mean that we cannot build a machine that will simulate
human cognition adequately? Are we forever restricted to hand coding of
propositions, as is being done in most of the research described in this
book? Not, perhaps, if we shift our criteria somewhat. Machines cannot
act and live as humans can and hence they cannot learn from experience
as we do, but they can read. Therefore, they can learn from reading.
A machine that knows about the world only from reading surely is a far cry
from a human with real red blood and surging hormones, but there is a lot
to be learned from the written word. It is only the second-best choice, but
suppose we teach a machine what the strength values in all these concept
and proposition vectors are by experience with the written word only.
Latent Semantic Analysis (LSA) is a technique that allows us to do
something like that (Landauer & Dumais, 1997; Landauer, Foltz, &
Laham, in press). This technique uses singular value decomposition, a
mathematical generalization of factor analysis. It was originally developed
as an information retrieval technique by Deerwester, Dumais, Furnas,
Landauer, and Harshman (1990) and extended to discourse analysis and
general problems concerning learning and language by Foltz (1996) and
Landauer and Dumais (1997). The reader will have to refer to these
sources for an adequate description of the technique and technical details.
Here, I try to give only a general impression of what LSA does and then
88 The theory

discuss later some specific uses of LSA in the investigation of discourse

comprehension.
In LSA the learning process is started by recording which words occur
in the same textual contexts. It reads texts in digital from and counts the
times that words appear in each segment of a text. For example, suppose
LSA reads about 4 million words from an encyclopedia. It counts for each
section of the encyclopedia which words appear and how often. It makes
no inferences about the words it reads, so that tree and trees for LSA are
different words. For 30,000 encyclopedia paragraphs containing 35,000
different words, we end up with a huge paragraph by words matrix, with
many empty cells, because each paragraph contains only a small subset of
the words. In the examples we discuss, the results are based on an analysis
of this kind provided by Susan Dumais of Bellcore and described in Lan-
dauer and Dumais (1997) and Landauer, Foltz, and Laham (in press).
If words were the appropriate units of cognition, we could stop here
and define each word as a very long vector, the entries of the vector being
the number of times the word has appeared in each paragraph or docu¬
ment. But we know that will not be a workable solution, for the very rea¬
son for introducing propositional representations in the first place was
that words are not satisfactory units for cognition. Thus, instead of defin¬
ing words directly in terms of documents (and documents in term of
words), LSA substitutes a semantic approximation. It uses the well-
known mathematical technique of singular value decomposition to radi¬
cally reduce the dimensionality of the space. A theorem of matrix algebra
states that any square matrix, M, can be decomposed into the product of
three matrices:

M —- A * D * A'

where A and A' are matrices composed of the eigenvectors of the matrix
and D is a diagonal matrix of the eigenvalues (or singular values) of the
matrix. The theorem generalizes to the nonsquare matrices used by LSA.
The eigenvalues are ordered in terms of their magnitude or importance.
Multiplying the three matrices yields back the original matrix, M. What
is done in LSA is to throw away most of the eigenvalues (and their asso¬
ciated eigenvectors), keeping only the largest ones, say, the 300 largest
ones. Multiplying the three matrices thus reduced does not reproduce M
precisely but only approximates the original M. That turns out to be a
considerable advantage. The original matrix contains too much informa-
 Propositional representations 89

tion - all the details and accidents of word use. By discarding all this
detail, we keep only the essence of each word meaning, its pure semantic
structure, abstracted from particular situations. This constructs a seman¬
tic space of, in the encyclopedia case, 300 dimensions in which each word
and document from the original matrix can be expressed as a vector. Fur¬
thermore, new words and documents can be inserted into this space and
compared with one another and with any of the vectors originally com¬
puted. Of the various ways to compare vectors, the only one I discuss here
is closely related to correlation: A measure of relatedness between vectors
is the cosine between the vectors in the 300-dimensional space. Identical
vectors have a cosine of 1, orthogonal vectors have a cosine of 0, and
opposite vectors have a cosine of-1. For instance, tree and trees have a cos
= .85; tree and cat are essentially independent, cos = -01; cat and The dog
chased the cat yield a cos = .36.
What is to be gained by this vector representation? Unlike proposi¬
tional analysis, it is fully automatic and objective. It is computationally
not very demanding (once an original semantic space has been con¬
structed). The cosine measure of semantic relatedness is readily inter¬
pretable. Thus, it is not necessary to assign links between nodes in a net¬
work arbitrarily or by collecting empirical association data. The cosines
between words, sentences, or paragraphs are easily computed and pro¬
vide objective, empirically based measures. We need no longer guess what
the neighboring nodes of a word (or sentence) are in semantic space - we
can look it up in the LSA space. Furthermore, words and documents are
treated in exactly the same way. “Documents” for LSA are akin to expe¬
riences or episodes for a human learner. For us the meaning of a word or
proposition is determined as much by the episodes in our memory to
which it is related as by the other words to which it is related; that is,
meaning involves both semantic and episodic memory. LSA computes
relationships both between words and between documents qua episodes.
Thus, it allows us to explore this important aspect of meaning, which is
not easily done within conventional approaches.
The initial results that the developers of LSA (Landauer, Durnais,
Foltz, and their co-workers) obtained with this method impressively
demonstrate its promise: LSA appears indeed capable of capturing much
of word meanings. More specifically, Landauer and Durnais (1997) found
that the vectors for words derived from the encyclopedia analysis pre¬
dicted the correct answers to standardized vocabulary tests in which
90 The theory

students are asked to judge similarity of meaning. Latent semantic analy¬

sis simulations matched the performance of moderately competent stu¬
dents: successful foreign applicants to U.S. colleges on tests of English
language competence (the TOEFL). Landauer and Dumais also demon¬
strated that LSA learns word meanings from reading at about the same
rate as late primary school children do. Both of these LSA predictions
crucially depended on reducing the dimensionality of the semantic space
to about 300 dimensions. That is, the words themselves do not matter, but
the semantic dimensions derived from their co-occurrences do. In more
recent unpublished work by Foltz, Landauer, and Laham, LSA has been
trained on introductory psychology textbooks. Its concept representa¬
tions were then tested with the same multiple choice tests that students
took. Latent semantic analysis usually scored about 60% correct (the
chance level is 22%), somewhere near the tenth percentile of the real stu¬
dents. These results illustrate both LSA’s impressive ability to approxi¬
mate human meaning but also the substantial gap that still exists between
humans and LSA. Integrating LSA into the comprehension model
developed here might help us to close this gap and at the same time obtain
a better and more realistic model.
The vector representation of LSA is similar to the feature vectors pop¬
ular in many psychological theories - except that we do not have to define,
invent, or identify specific features. We need not interpret the values on
the 300-dimensional LSA vector (in fact, we cannot), but we can objec¬
tively and automatically represent the meaning of various verbal units in
this way and use these representations in computational models of com¬
prehension and memory. The theory of memory based on vector repre¬
sentations has been relatively well explored (Estes, 1995) and hence can
be used for modeling with LSA vectors. Latent semantic analysis needs
no parser; it treats sentences, paragraphs, and whole texts holistically,
representing each as a vector. As currently implemented, LSA has limi¬
tations that need to be explored. The very fact that it needs no parser also
means that it does not take into account syntactic information, at least in
its present form. It allows us to approximately represent global meaning
but not the analytic, formal aspects of human thought. Therefore, for
example, LSA cannot serve as a basis for computing truth values. There
may be several ways to overcome this limitation, one of which I have
begun to explore here, namely, to combine the LSA vector representation
with the construction-integration model of comprehension.
Latent semantic analysis is a young technique, and the research
 Propositional representations 91

described in this book involving LSA is only a beginning and still quite
tentative.2 Nevertheless, it opens exciting new possibilities that should be
explored by discourse researchers.

How one represents the semantic structure of a text, the mental

model that a reader forms when comprehending a text, as well
as the reader’s knowledge and experience that are needed for
comprehension, are of fundamental importance for a psycholog¬
ical processing model of comprehension. The processing model
itself is described in chapter 4 and then explored in the second
part of the book. In this chapter, two representational formats
were discussed. Both of them are language-based, that is, they
are representations at the narrative-linguistic level, but they
provide reasonably adequate approximations to other levels of
representation as well, such as perception and action-based
forms of mental representation. The first is the propositional
representation that came into use in psychology and the study of
language in the 1970s. It has been widely used since then and is
by now quite familiar. The second is a vector representation that
shares some features with propositional representations, but it is
as yet not well known and remains to be explored further.
In the use of propositional representations of texts, we look
back at a 20-year history of success. Much of what we have learned
about text comprehension and memory in these decades we could
not have discovered without a good system for representing the
meaning of texts and the structure of texts. The propositional
notation has served that purpose well. It is not an adequate nota¬
tion for logic, semantics, or some branches of linguistics, but it
was never intended for that use. It is a tool for research in the psy¬
chology of language, and much of the progress in that field has
been based on the availability of that tool.
The use of propositional networks as knowledge representa¬
tion is newer and not nearly as well established. It has several
attractive features, however. Obviously, it is good to have the same

2 See sections 5.1.2 on words meaning, 5.3 on metaphor, 6.2.3 on macrostructures, 7.4
on retrieval structures, 8.2.2 on evaluation of summaries, 9.6 on learning from text,
and 11.4 on decision processes. Readers may check (and extend on their own) these
computations by using the LSA program available at http://samiam.colorado.edu
/~lsi/Home.html
92 The theory

representational format for both the mental model a reader con¬

structs from a text and the reader’s knowledge. More important
is the great flexibility of the propositional notation. Many differ¬
ent structures can be represented in a predicate-argument for¬
mat: frames, schemas, scripts, even procedures. Thus, the propo¬
sitional notation allows us to appropriate a wide variety of mental
structures that-have been found useful in cognitive science in one
way or another, without forcing us to choose among alternatives
that all have their advantages and disadvantages. Most important,
however, propositional representations invite one to think about
meaning in a different way. The meaning of a node in a proposi¬
tional network needs to be constructed from the information the
net provides. It is an active process of sampling related, neigh¬
boring nodes in the net and of integrating them into a coherent
whole. This may be quite a superficial process or an extensive and
deep one, but it always occurs in a specific context that crucially
biases the final outcome. Thus, meaning is not fixed and ready¬
made but must be constructed anew in every new context.
The trouble with propositional nets as knowledge representa¬
tions is that, although they work well in small, illustrative exam¬
ples, it is impossible to create a large, realistic propositional
knowledge net, for the whole net would have to be hand coded,
just like a text representation. Latent Semantic Analysis may pro¬
vide a practical alternative here. It can learn what it needs to know
by reading very large amounts of texts. The information it needs
is information about the higher-order correlations in the co¬
occurrence of words. As is the case with propositions, LSA treats
meaning as a pattern of semantic relationships. Words, sentences,
and whole texts can be represented as vectors in a high-dimen¬
sional semantic space and compared with one another. In the
chapters that follow, some of the advantages of this alternative
knowledge representation are explored when it is used in con¬
junction with the processing model originally developed for
propositional representations. The preliminary results that are
available at this point suggest a great potential for LSA for psy¬
chological knowledge representation.
 4

Modeling comprehension processes

The construction—integration model

Language has been studied and analyzed for centuries. Philoso¬

phers, linguists, logicians, and others have accumulated a rich
store of knowledge about language. What has emerged, however,
is not a uniform, generally accepted theory but a rich picture full
of salient details, brilliant insights, ambiguities, and contradic¬
tions. Most of this work has focused on analyzing language as an
object, rather than on the process of language comprehension or
production. The importance of the actual process of language
comprehension has not gone unrecognized, for instance, by liter¬
ary scholars who have understood very well the role that the
process of reception plays in the appreciation of a literary work,
yet the tools for explicit modeling of comprehension processes
have not been available until quite recently. Progress in the build¬
ing of artificial intelligence systems for natural language process¬
ing as well as in the techniques for the simulation of higher-order
cognitive processes has provided the foundation upon which psy¬
chological process models of comprehension can now be built.

We comprehend a text, understand something, by building a mental

model. To do so, we must form connections between things that were pre¬
viously disparate: the ideas expressed in the text and relevant prior knowl¬
edge. Comprehension implies forming coherent wholes with Gestalt-like
qualities out of elementary perceptual and conceptual features. It is a mar¬
velous and wondrous achievement, for there are myriads of such features
ready to yield many different configurations. How are the right ones
94 The theory

selected from these all-too-rich offerings, and how do they cohere into an
understanding that is attuned to our goals and motivations as well as the
characteristics of the situation within which we find ourselves?
There is no question that this happens. But how? The traditional
answer is that order is guaranteed because the process of understanding
is under the control of a schema that guides it (Bartlett, 1932; Schank &
Abelson, 1977; Selz, 1922). A schema in this view serves as a control
structure that regulates comprehension processes in a top-down fashion.
It works, on the one hand, like a perceptual filter, in that it admits mate¬
rial consistent with itself but blocks irrelevant materials, and, on the other
hand, it serves as an inference machine, in that it fills in the gaps that are
inevitably found in the actual stimulus material.
Schema theory has taken many forms in the years past, in both AI and
psychology, and has become rather more sophisticated than the caricature
just presented suggests. Nevertheless, it has its problems. First and fore¬
most, there exist enough psychological data to question whether the top-
down guidance of comprehension is as tight as schema theory suggests. I
do not review these data here, but results are presented throughout this
book that indicate the need for conceiving of comprehension as a more
bottom-up, loosely structured process. Second, human comprehension is
incredibly flexible and context-sensitive. It is hard to see how one could
model that process with fixed control structures like schemas. It is, there¬
fore, worth exploring an alternative.
A coherent model is certainly the outcome of comprehension, but it
does not have to be the result of forcing comprehension into a pro-
crustean schema. One can conceive of comprehension as a loosely struc¬
tured, bottom-up process that is highly sensitive to context and that flex¬
ibly adjusts to shifts in the environment. Comprehension, in this view,
may in fact be quite chaotic in its early stages. It becomes orderly and well
behaved only when it reaches consciousness. Such a process can be mod¬
eled by a construction process that is only weakly controlled and proceeds
largely in an associative, bottom-up manner that is followed by a con¬
straint satisfaction process in the form of a spreading activation mecha¬
nism and that yields the coherence and order that we experience.
I have called this model the construction-integration (Cl) model. In
this model mental representations are formed by weak production rules
that yield disorderly, redundant, and even contradictory output. How-
 Modeling comprehension processes 95

ever, this output undergoes a process of integration, which results in a

well-structured mental representation. Thus, the end result is the same,
both according to schema theory and the present proposal. But the way
this structure is achieved is quite different.
According to schema theory, the earthquake schema would rule out the
chocolate meaning of mint in The earthquake destroyed all the buildings in
town except the mint; in the Cl model propositions with both meanings of
mint would be constructed, but the one that was irrelevant to the context
would be rapidly suppressed. Again, according to schema theory, a reader
might very well be reminded of river when reading The car drove over the
bridge, but not of cards, which would be blocked by the schema, whereas
such irrelevant associations should be commonplace (though fleeting)
according to the Cl model. Upon hearing The hiker was surprised to see a
bear blocking his path, a schema-based inferencer might readily infer that
the hiker was scared but hardly that bears like honey. Hence, on a naming
task, schema theory would predict a priming effect for scared but not for
honey; Cl would predict priming effects for both, as long as the target
word was presented right after bear in the sentence. In every case, the
smart rules of a schema-based theory would prevent the wrong con¬
structions from ever being formed. In contrast, the production rules in
the Cl model are weak and dumb and do not discriminate what is con¬
textually appropriate from what is not; they are just as likely to instanti¬
ate the wrong as the right meaning of a word, or to form an irrelevant as
a relevant inference. The construction of the “correct,” contextually
appropriate meaning results from the integration process that quickly
deactivates contextually inappropriate constructions.
This chapter presents the basic elements of the construction-integra¬
tion theory. After this initial sketch, the remaining chapters of this book
will elaborate on the theory in the context of a wide range of issues con¬
cerning comprehension.
The process model to be described here was first described in Kintsch
(1988; see also Kintsch & Welsch, 1991; Kintsch 1992a, 1992b). It con¬
tinues and extends the earlier model of Kintsch and van Dijk (1978; see
also van Dijk & Kintsch, 1983); van Dijk and Kintsch (1978) provided the
basic process model but did not deal with knowledge use in comprehen¬
sion; Kintsch (1988) added this feature, creating the construction-
integration model.
96 The theory

4.1 Construction and integration

To say that construction rules are weak and bottom-up and require an
integration process to function does not specify what these rules are. It
merely characterizes an architecture within which various kinds of mod¬
els can be developed. That is, to model a particular cognitive process, a set
of (weak) construction rules must be formulated that are involved in this
process. I discuss a few examples here that are important for understand¬
ing text. This discussion remains informal and relies on hand-simulation
of the rules involved. Construction rules for a number of different cog¬
nitive processes are described and evaluated empirically in the remaining
chapters of this book.

1. Rules for the construction ofpropositions. This is a large and important

class of rules. An exhaustive and explicit statement of this set of rules
would constitute the parsing component that is missing from the present
model. In practice, we proceed as described in section 3.1.
2. Rules for interconnecting the propositions in a network. As discussed
previously, three levels of connections among propositions can be distin¬
guished: Propositions may be only indirectly related, they maybe directly
related, or one proposition may be subordinated to another. Further¬
more, propositions may be negatively linked, that is, interfere with each
other. Whenever two alternative propositions are formed from the same
sentence or phrase, this contradiction is indicated by a negative link. As
an example, consider the syntactically ambiguous phrase they are flying
planes, which yields the structure

FLY[THEY,PLANES]-ISA[THEY,FLYING [PLANE]]

The negative link is assigned not because the two interpretations are log¬
ically incompatible, but because the two alternative propositions were
formed from the same sentence. Thus, the system does not have to under¬
stand the contradiction in order to assign a negative link, but it does so on
purely formal grounds. If the same string of words can be configured
once into one proposition and once into another, the two constructions
are mutually exclusive and inhibit each other.
A propositional network is thus formed. Link strengths may be chosen
to be all equal, or they may be varied according to theoretical considera¬
tions. For instance, embedding relations may be given more weight than
 Modeling comprehension processes 97

direct links, which in turn may be given more weight than indirect links.
Link strengths may also be estimated empirically from free association
data or by means of the cosine values in an LSA analysis. Different ver¬
sions of the model may thus be obtained to be evaluated empirically for
their adequacy.
3. Rules for the activation of knowledge. The basic mechanism for knowl¬
edge activation is assumed to be associative. Items in working memory
activate some neighboring nodes in the knowledge net with probabili¬
ties proportional to the strengths with which they are linked to them.
This mechanism may be modeled formally (see, for instance, Mannes &
Kintsch, 1987, where the SAM theory of memory retrieval by Raajmak-
ers & Shiffrin, 1981, is used for that purpose; likewise Doane, McNa¬
mara, Kintsch, Poison, & Clawson, 1992, or Kitajima & Poison, 1995,
where memory retrieval processes are of central importance.) Sometimes,
however, a more informal treatment is appropriate, as in many of the illus¬
trative examples discussed later. Thus, Kintsch (1994a) obtained empiri¬
cal estimates of what knowledge was likely to be retrieved but then made
the simplifying assumption that all of it actually was retrieved. Even more
informally, often only directly relevant knowledge is dealt with explicitly
in uses of the model when the complete details of knowledge retrieval do
not play a central role.
4. Rules for constructing inferences. Just as we are unable to state explic¬
itly all the rules for deriving a propositional representation from a text,
we cannot list all rules that people use for inferencing. Examples of such
rules appear frequently throughout these chapters. For instance, a simple
rule used in forming a situation model might involve the following tran¬
sitive inference:

If A above B and B above C, then A above C.

As another example, a rule for constructing a macroproposition might

take the following minitext:

(1) Jane drove to Alfalfa’s, picked up some fresh fruit, a halibut

steak, and some Italian cheese for dessert, and paid with her
credit card.

into

(2) Jane went grocery shopping.

98 The theory

Rules like (1) and (2) are used intuitively in applications of the present
model, just like the parsing rules. We are primarily concerned with the
resulting structures, not with the parsing and inference rules themselves.

Once a network of propositions has been constructed through hand

simulation of the kind of rules described here, and once these proposi¬
tions have been linked into a network in a theoretically motivated and
consistent way, a spreading activation process is used to stabilize the net¬
work. Activation is spread around the network until it stabilizes in a way
that takes account of the pattern of mutual constraints that exists among
the nodes of the network (the propositions of a text together with the
inferences and knowledge elaborations a reader has produced). The final
activation values of the nodes thus come to reflect the constraining prop¬
erties of the network as a whole.
In a network of n nodes, at time t, each node has an activation value
a -(f). The activation values of all nodes in the net at time t are given by the
vector

A(/) = (fli(f), «2(0, • ■ • a„W) “ {«,<*)}

The initial activation vector is usually taken to be A(l) = {1}. The

strength of the link between the nodes i and j in the network is n>tJ. Thus,
the total pattern of interconnections among the nodes is given by the con¬
nectivity matrix

W = {**}

Activation spreads to a node from all of its neighboring nodes, so that the
activation of a node at time t + 1 becomes

X
ah + 1) = -
max a;(r + 1)

Dividing the activation value of a node after activation has spread to it by

the maximum activation value in the net renormalizes the activation val¬
ues and keeps them from growing without bounds. In earlier versions of
the model, activation values were normalized by requiring the total acti¬
vation in a net to sum to 1 by dividing each activation score by the sum of
 Modeling comprehension processes 99

all scores. Other standard connectionist techniques could also be used for
normalization - for example, a sigmoidal transformation, as in Goldman
and Varma (1995).
The spreading activation process is continued for as many iterations as
are necessary for the activation values to settle in a stable pattern. Once
no activation value changes by more than some criterial amount, say .001
or .0001, the process is stopped. The conditions under which a spreading
activation process stabilizes, that is, approaches an asymptotic activation
vector A(°°), or simply A, are not known in general (though see Roden-
hausen, 1992). Symmetric connectivity matrices with no negative values
behave like ergodic Markov chains. That is, an asymptote exists and is
independent of the starting vector.
The values of A indicate the strength of each node in the integrated
network. Nodes that are positively connected to many other nodes in the
net will be strengthened. Nodes that have few connections or are nega¬
tively connected will wither away or become suppressed. This is a process
of constraint satisfaction: Nodes that satisfy the multiple constraints of
the network become stronger, whereas nodes that do not become weaker.
The resulting pattern of activation indicates for each node the role it plays
in the network as a whole and can be considered a measure of the strength
of that node in the mental representation.
A simple numerical example helps to form some intuitions about this
integration mechanism. Consider the network of five nodes shown in
Figure 4.1. The dashed line indicates a negative link. If we let all links
have a strength of either 1 or -1, we obtain the following connectivity
matrix, remembering that each node is linked to itself:

A B C D E
A 1 1110
B 1 10 10
C 1 0 10 1
D 1 10 1-1
E 0 0 1-11

We assume that each of the five nodes has an activation value of 1 initially:

A(l) = (1, 1, 1, 1, 1)
100 The theory

Figure 4.1

Multiplying this initial activation vector once by the connectivity matrix

yields the following values:

(4, 3, 3, 2,1)

The activation values after one iteration are obtained by dividing these
scores by their maximum:

A(2) = (1.00, .75, .75, .50, .25)

If this new activation vector is repeatedly multiplied by the connectivity

matrix, renormalizing it after each multiplication, the activation values
stabilize after nine iterations. That is, the maximum change after nine
iterations is less than .001. These final activation values for this network
are as follows:

A(9) = A(~) = (1.00, .85, .46, .85, .00)

Node E has become completely suppressed, in spite of its connection

with node C, because of the strong inhibition from node D. Nodes A, B,
and D support one another effectively, whereas C turns out to be only a
marginal element in this net.
These calculations can be readily performed with a program written
by Mross and Roberts (1992). This program for Macintosh is available on
 Modeling comprehension processes 101

request.1 It allows two kinds of input: Either proposition lists may be

used that are automatically turned into a network and its corresponding
coherence matrix by linking propositions with common arguments, or
nets of any kind that are automatically translated into matrix form can be
constructed graphically. The spread of activation can be calculated cycle
by cycle or all at once until a stopping criterion is reached. Various dis¬
play options are available, as well as a number of features that I introduce
later. The program allows readers not only to check the results reported
here, but it also provides an easy means of exploring on their own con¬
straint satisfaction networks of the type described here.

4.2 Processing cycles in text comprehension

Text representations must be built up sequentially. It is not possible

psychologically to construct and integrate a text representation for a
whole book chapter or a whole lecture. The chapter and the lecture have
to be processed word by word and sentence by sentence. As each text
segment is processed, it is immediately integrated with the rest of the text
that is currently being held in working memory. The immediate process¬
ing hypothesis, first put forward by Just and Carpenter (1987), generally
holds, at least for lower-level processes in comprehension. Occasion¬
ally, however, readers use delay strategies when dealing with potentially
ambiguous syntactic constructions (e.g., Ferstl, 1994b) or they continue
reading when constructing a situation model when they do not under¬
stand something, in the hope that the succeeding text will clarify their
problem. But in general information in a text is processed as soon as pos¬
sible. In the model this means that as each text element is processed and
a new proposition is added to the text representation, it is immediately
integrated with the text representation. Thus, integration is not a sen¬
tence wrap-up phenomenon but is performed whenever a new element is
added to the network under construction. Therefore, an ambiguous word
in a discourse context will be disambiguated within about 350 ms after
presentation - not just at the sentence end (Till, Mross, & Kintsch,
1988).

1 Send diskette to W. Kintsch, Institute of Cognitive Science, University of Colorado,

Boulder, CO, 80309-0344.
102 The theory

The integration that takes place at the sentence end has a special sta¬
tus, however. Except for very short sentences, working memory at this
point is usually loaded to capacity and must be cleared to make room for
the next sentence. Whatever has been constructed is transferred to long¬
term memory. As a consequence, except for one or two central proposi¬
tions that are retained in the focus of attention because of their presumed
relevance to further processing, all that has been constructed up to this
point in working memory is now lost from consciousness/primary mem¬
ory. However, in a normal text, this information is still readily retrievable
because the succeeding sentence most likely will contain retrieval cues
that make it accessible in long-term working memory, as discussed in
chapter 7. The active processing focuses on only the current sentence,
plus whatever information had to be retrieved from long-term memory
that is necessary for its processing. Thus, working memory is like a spot¬
light that moves across a text, sentence by sentence, constructing and
integrating a mental representation in the process. The representation
that results from this cyclical process is a coherent structure and not a
sequence of disjoint structures, each corresponding to a sentence. The
coherence is a result of the fact that propositions are kept in short-term
memory from one sentence to the next to serve as bridging material and
because earlier portions of the text as well as general knowledge items that
were retrieved during processing provide a lattice of interconnections.
The number of propositions that are carried over from one cycle to the
next is a free parameter of the model. The estimates that have been
obtained so far suggest that it is a small number - generally 1. Just which
proposition is carried over is determined by the activation values propo¬
sitions receive at the end of a processing cycle. Typically, the strongest
proposition of the previous cycle remains in the focus of attention as
attention shifts to a new sentence. This strength criterion replaces the
selection mechanisms that were introduced in earlier work, for example,
the “leading edge strategy” of Kintsch and van Dijk (1978). A full dis¬
cussion of memory use in comprehension is given in chapter 7.
As a result of this cyclical processing of a text, not all relations in a
text that can be detected by linguistic analysis are actually realized in the
mental representation. Only those relations that hold between proposi¬
tions that were together in working memory at some time during the sen¬
tence-by-sentence process of comprehension play a role in the text repre¬
sentation.
 Modeling comprehension processes 103

As is common in connectionist models, the original coherence matrix

W and the final activation vector A can be combined to give the outcome
of the comprehension process, that is, the long-term memory strength of
each element of the network and its interconnections. A new matrix M of
memory strengths m ■ can be defined such that

where w,,Ij is an element of W and a,i is the final activation value of the zth
element; the summation is over the k processing cycles that the zth ele¬
ment has participated in, k > 1.

4.3 Textbase and situation model

The comprehension process yields as its end product a mental represen¬

tation of a text, the episodic text memory. According to the theory pre¬
sented here, it consists of a network of interrelated propositions of vari¬
ous strengths, given by the long-term memory strength matrix M.
This episodic text memory is a unitary structure, but for analytic pur¬
poses it is useful to distinguish two components - the textbase and the
situation model. The textbase consists of those elements and relations
that are directly derived from the text itself. It is what would be obtained
if a patient linguist or psychologist were to translate the text into a prop¬
ositional network and then integrate this network cycle by cycle, as
described earlier, but without adding anything that is not explicitly spec¬
ified in the text. In general, this procedure yields an impoverished and
often even incoherent network. The reader must add nodes and establish
links between nodes from his or her own knowledge and experience to
make the structure coherent, to complete it, to interpret it in terms of the
reader’s prior knowledge, and last but not least to integrate it with prior
knowledge. Various sources of knowledge must be used in the construc¬
tion of situation models - knowledge about the language, about the world
in general, and about the specific communicative situation. Furthermore,
not only general knowledge enters this process but also the reader’s per¬
sonal experience. Such sources of knowledge may all be needed to com¬
plement the textual information and to transform what by itself is only an
isolated memory structure into something that relates to and is integrated
with the reader’s personal store of knowledge and experience.
104 The theory

The extent to which a reader will actually perform the work of trans¬
forming a textbase into a situation model is subject to numerous influ¬
ences. In a given comprehension episode, one influence or another may
dominate, but it is generally the case that both play a role, complement¬
ing each other.
Figure 4.2 illustrates the relationship between textbase and situation
model. The textbase can be more or less coherent and complete, and the
situation model can be more or less adequate and precise. Mental text
representations may fall anywhere in this quadrant. In the text-that-tells-
it-all, in which every detail as well as the overall structure is made per¬
fectly explicit (as far as that is possible!), the textbase is also a good situa¬
tion model and no further knowledge elaborations on the part of the
comprehender are required. Typically, however, the mental text repre¬
sentation is a mixture of text-derived and knowledge-derived informa¬
tion, not necessarily in equal parts. Extreme cases, in which either the

high
LU BB&F(1972)
Q O The entirely ex^cit text
O
/
X /
O
/
S3 /
t /
CO /
LU /
X
I- /
LL.
/
O
>- /
o /
< / M&K(1993)
3 o
o
LU
/
Cl
< low high

COHERENCE OF THE TEXTBASE

Figure 4.2 Textbase and situation model jointly determine the nature of the
mental text representation. Extreme cases are Bransford, Barclay, and Franks
(1972), where the situation model dominates at the expense of the textbase,
and Moravcsik and Kintsch (1993), where the textbase dominates at the
expense of the situation model.
 Modeling comprehension processes 105

textbase or the situation model dominates to the exclusion of the other,

are instructive. Bransford, Barclay, and Franks (1972) and Bransford and
Franks (1971) described some experiments in which they obtained no
memory for the text itself - neither the actual words and phrases of the
text (the surface structure of the text) nor its meaning (the propositional
structure). However, subjects had understood the texts very well and
were able to form stable situation models on the basis of which they could
answer questions correctly and make inferences about the text. The
experimenters produced these results by using artificial little texts that
generated a great deal of intersentence interference so that subjects
quickly forgot the text; however, subjects remembered very well the sim¬
ple situations described by the text. At the other extreme, there were
experiments such as that reported by Moravcsik and Kintsch (1993) in
which subjects remembered a text reasonably well but failed to under¬
stand it at the level of the situation model. Their texts were well written
and provided rich structural cues to the reader, so that skilled readers
could construct a good, coherent textbase that was sufficient for the
reproduction of the text. However, the texts concerned topics about
which the subjects knew very little, so that they were unable to construct
more than rudimentary situation models, which were insufficient for
tasks that demanded more than the reproduction of the text.
The textbase consists of the network of propositions that represent the
meaning of the text, as understood by a particular reader. Usually, an ide¬
alized “normal” reader is assumed, who reads everything carefully and
constructs a proposition whenever the text invites one. This is not what
real readers necessarily do, but it provides a baseline against which to
score the performance of real readers. Thus, even if a text is written in
such a way that an ideal reader could construct from it a highly coherent
textbase, such will not necessarily be the case.
Another component of the textbase, in addition to the propositional
structure, is the surface structure of the text. When the reader is read¬
ing a text or listening to a conversation, it is generally the case, that at
least some of the exact words and phrases are remembered at least for a
time. Under certain conditions, for example, in sentence recognition
experiments or in memory for actual conversations, this surface memory
may play a significant role. Thus, words, word groups (such as a noun
phrase), or a prepositional phrase unit may be represented as nodes in the
106 The theory

textbase, with links to their propositional counterparts as well as links

between them, according to the linguistic relations involved. One does
not have to stop there. For some texts, especially poetry, other relations
among words or phrases such as rhymes, rhythm, or alliteration must be
included as relations in the textbase if comprehension or memory for
such texts is to be adequately simulated (Kintsch, 1994b).
All kinds of knowledge elements may be used in the construction of a
situation model. The textbase is made coherent, for instance, by inferring
a connective between only indirectly related propositions:

(3) Jane could not find the vegetable and the fruit she was looking
for. She became upset.

Not finding something one is looking for is a common reason for becom¬
ing upset. Hence, a causal link is inferred between the two propositions,
thus forming a simple situation model.
Elaborative inferences come in many kinds:

(4) Jack missed his class because he went to play golf. He told his
teacher he was sick.

A situation model may be formed by the elaboration Jack lied.

In

(5) A turtle sat on a log. A fish swam under the log.

the situation model might be an image of a lake, a log with a turtle, and a
fish underneath. Various elaborations are implied by such an image, for
example, that the turtle is above the fish, the fish is in the water, and so on.
Situation models often involve more than filling gaps between propo¬
sitions or local elaborations. They may provide a basis for the overall
organization of the text even in cases where the text structure is not sig¬
naled explicitly in the text. Readers recognize certain text structures and
can use this knowledge to impose an organization on the text. For
instance, in a little story “Jane goes grocery shopping” readers can gen¬
erate a macrostructure for the text even if none is signaled by reference to
their grocery-shopping script, which identifies the episodes typically
involved in grocery shopping (preparation, shopping, checkout, pay/
leave). The sentences of the text can be subordinated to these episode
nodes, thus creating the macrostructure of the text via the schema-based
 Modeling comprehension processes 107

situation model. If an episode (e.g., checkout) is not mentioned explicitly

in the text, this schema can be a source for gap-filling elaborations.
A situation model is, therefore, a construction that integrates the text-
base and relevant aspects of the comprehender’s knowledge. No general
rules can be stated, because knowledge elaborations may be of many
different types, and the extent to which they occur may differ widely
among readers and occasions. How much elaboration occurs depends on
the text - whether it is self-sufficient or not - but also on the readers, their
goals, motivation, and resources available.
In some cases it is necessary to distinguish between different kinds of
situation models. Thus, in their work on how students understand and
solve word algebra problems, Nathan, Kintsch, and Young (1992) found
it useful to distinguish between the textbase, a situation model, and a
problem model. In an overtake problem, for instance, the situation model
represents the student’s general understanding of what happens when
two vehicles move along the same track at different speeds; the problem
model represents the mathematization of this situational understanding
in terms of algebraic entities and equations. Students have no difficulty
with the former but may be unable to formulate the latter. Fischer, Hen-
ninger, and Redmiles (1991) have made a similar distinction with respect
to computer systems. Users often do have a good understanding of the
tasks they need to perform at a general level but do not know how to use
their computer system for that purpose; they have an adequate situation
model but do not know how to map it into a system model. The objects
and relations in the situation model represent objects and relations in the
world that we are familiar with; the objects and relations in the problem
(or system) model represent objects and relations in a formal world, that
of algebra or UNIX. Obviously, these two models have to be in corre¬
spondence, but they need to be differentiated, because they involve quite
different processing demands.
We distinguish between textbases and situation models not because
they are somehow different mental objects. On the contrary, there is a sin¬
gle, unitary mental representation of a text. To describe this representa¬
tion and its genesis, it is useful to separate out textbases and situation
models, as well as problem models and surface structures, as our analyses
require.
To lend these discussion some concreteness, two simple examples are
108 The theory

discussed to illustrate a few aspects of the construction of textbases and

situation models.

4.3.1 Imagery as a situation model

The minitext (from Kintsch, 1994a) to be considered here is:

(6) John traveled By car from the bridge to the house on the hill. A
train passed under the bridge.

To form the textbase we assume that the two brief sentences are both
processed in a single cycle. The first sentence yields a single proposition
with a modifier; another proposition with a filled location slot is derived
from the second sentence; the two propositions are only indirectly related
(by the common argument BRIDGE), as shown in Figure 4.3. The
textbase is thus strongly constrained by the text; most or all cooperative
readers would be expected to form a propositional textbase like the one
shown in the figure. Just what sort of situation model would be formed is
much less constrained. We make the assumption here, which will be cor¬
rect for at least some readers, that an image will be formed. For the first
sentence, this image might be of a house on a hill, a bridge crossing a
river, a road connecting the bridge and the river, and a car driving from
the bridge up to the house. Note that this image does not include John,
but there is a road and a river, as well as a particular spatial relationship
between hill and river - for example, the hill might be to the left of the

TRAVEL
agent: JOHN
instrument: CAR
source: BRIDGE
goal: HOUSE
mod: ON-HILL

PASS
object: TRAIN
location: UNDER-BRIDGE

Figure 4.3
 Modeling comprehension processes 109

river, for which there is no license in the text. This image can be repre¬
sented by the four spatial relations (shown in the ovals of Figure 4.4),
which form a network together with the two textbase elements (shown in
the rectangules). The TRAVEL proposition corresponds to the
ON[CAR,ROAD] image element, which therefore links them together in
the network, and the spatial and propositional ON[HOUSE,FIILL] ele¬
ments correspond to each other.
As each proposition or imagery element is formed, it is integrated into
the network that is being formed in working memory. The status of the
network after the elements shown in Figure 4.4 have been constructed
and integrated is indicated by the activation values that are shown with
each element. The main point is that the gratuitous elaboration of includ¬
ing a river in the image receives a low but positive activation value. This
will change, as the next sentence becomes part of the network (see Figure
4.5).
We have the two propositional elements derived from the second sen¬
tence, which give rise to a new imagery element: a train under the bridge.
This is linked to the previous image of the bridge, road, and house but is
incompatible with the other element of the previous image of a river
110 The theory

under the bridge. Thus, the negative link is based on an inference that a
bridge over train tracks does not have a river underneath (neglecting the
possibility that a bridge could span both train tracks and a river). The two
UNDER nodes therefore interfere with each other and are linked by a
value of-1 in the network (indicated by a broken line). The outcomes of
the final integration cycle are shown at the top, with each node. The
image of the train under the bridge dominates the net and completely
suppresses the image of bridge and river.
This example was chosen to demonstrate two points. First, it shows
how the model can deal with images - not very elegantly, but it can do so.
Some imagery information can be presented in the way shown here but
with some loss of information - the concreteness of the image together
with some of its spatial implications. Second, it provides a good example
of the operation of weak, dumb rules. In response to the first sentence, an
image was constructed of a bridge and a river. The text said nothing about
 Modeling comprehension processes 111

a river; it was imported as a part of the image of a bridge. In the second

sentence, this image had to be corrected because a new image interfered
with it. No damage was done by the gratuitous inclusion of a river in the
budding situation model - it simply disappeared when better information
became available. Knowledge (river-under-bridge) was activated associa-
tively, that is, by a context-independent, dumb mechanism. But the
mechanism need not be any smarter, because the network representation
corrects errors easily and without requiring major repairs in the structure
that has been generated.

4.3.2 A script-based situation model

Many texts contain linguistic markers of importance and structure that

signal their macrostructure. The following example shows how a text that
lacks such signals can nevertheless be assigned a macrostructure as part
of the construction of a situation model if it fits a familiar schema. This
text is slightly longer and must be processed in more than one processing
cycle. We assume a cycle size of six elements. Once working memory con¬
tains six new elements, the model stops accepting further input as soon
as it reaches a sentence (or, in some cases, phrase) boundary. In other
words, working memory capacity is not inflexible, in that more than six
new elements may be processed if it is necessary to avoid breaking up a
complex proposition. The text is as follows:

(7) (SI) Jane went grocery shopping on Sunday afternoon. (S2) She
had to park the car far away from the store because the parking
lot was full. (S3) The aisles were as crowded as the parking lot.
(S4) The lettuce was wilted and the fruit was picked over. (S5)
She became upset. (S6) But she filled her cart with whatever she
could find. (S7) The bill was larger than expected. (S8) She car¬
ried two heavy bags across the parking lot.

The first sentence (SI) retrieves the grocery-shopping script as in Fig¬

ure 4.6. This script provides a macrostructure for the text, in that its slots
become the superordinate macropropositions. The first sentence (SI)
merely specifies the agent and time slot for this script. Because so far
fewer than six network nodes have been created, the model continues to
112 The theory

GROCERY SHOPPING

agent: PERSON

event sequence: PREPARATION

SHOPPING
CHECKOUT
PAY-LEAVE

place: GROCERY STORE

time: $

Figure 4.6

read in the second sentence (S2). This sentence is assigned to the prepa¬
ration slot of the grocery-shopping script, because preparation is associ¬
ated in the knowledge base with making a shopping list, driving to the
store, parking, and entering store. Thus, the structure that is generated
when the first two sentences are processed now contains seven nodes
(Figure 4.7). All nodes and links are assigned a value of 1. The main
proposition is GROCERY-SHOPPING, with three filled slots; the
PREPARATION slot is connected to PARK, which has two modifiers
and is linked (via an explicit BECAUSE relation) to FULL. Working
memory is loaded to capacity, and the end-of-cycle integration deter¬
mines the final activation values for these nodes, which are as shown in
Figure 4.7 before each element of the net.
The third sentence (S3) is associated with the SHOPPING slot of the
main proposition, and it also reinstates the FULL proposition from the
previous processing cycle. Because S3 is a short sentence and only three
new elements have been constructed at this point, the model goes on and
reads in S4 as part of the same cycle. It is also connected to the SHOP¬
PING episode. In addition, the PARK proposition is carried over in the
short-term memory buffer as the most highly activated proposition from
the first processing cycle. This yields the network shown in Figure 4.8.
The third processing cycle comprises S4 and S5. The key element here is
the UPSET proposition. Because other information has already been
 Modeling comprehension processes 113

B
E
C
A
U
S
E

.48
FULL[PARKING-
LOT]

Figure 4.7

processed that constitutes normal conditions for getting annoyed, the

reader may2 generalize and form a macroproposition, ANNOYED, with
the slots agent, condition, state, and consequence (the latter remains
unspecified). The full parking lot, the crowded aisles, and the picked-
over fruit all get connected to this new macroproposition as precondi¬
tions and hence are reinstated (Figure 4.9). The link between UPSET
and FILL represents the sentence connective but. Again, activation val¬
ues after the end-of-cycle integration are shown next to each proposi¬
tional element.
The last processing cycle includes the final two sentences, S7 and S8.
The ANNOY proposition is carried over as the most highly activated ele¬
ment on cycle 3 (Figure 4.10).
Thus, the processing of this text resulted in the creation of two macro¬
propositions, GROCERY-SHOPPING and ANNOYED, that serve to
organize and interrelate the text. The first macroproposition was gener-

2 This is only one possible interpretation of this text - not all readers will follow this
course.
114 The theory

Figure 4.8

ated by the fact that a script that was retrieved on the basis of the first sen¬
tence proved to provide an appropriate organizational structure for the
remainder of the text. Each sentence of the text was recognized as a pos¬
sible filler for the slots of the script that was available. This computation
involves matching each (complex) text proposition with propositions
associated with the slots of the scripts in long-term memory. The second
macroproposition involved a similar process. A proposition in the text -
UPSET — retrieved information from long-term memory (let’s call it an
ANNOY schema) about normal conditions for becoming upset in a gro¬
cery store; text propositions were found that matched this information
and hence were subordinated to the ANNOYED proposition. These two
macropropositions provide an organization for the text. They are the core
 Modeling comprehension processes 115

1.00
ANNOYED
agent: JANE
.41
condition: FULL[PARKING LOT]

state:

.41
CROWDED[AISLES]

.37 .91 .41

BECOME UPSET[JANE] PICKOVER[FRUIT]

BUT

.32
GROCERY SHOPPING
agent: JANE
.87
FILL[JANE,CART,$]
.75
event sequence:SHOPPING

.43
FIND[JANE,GROCERIES]

.17
WHATEVER

Figure 4.9
116 The theory

Figure 4.10

of the situation model; the rest of the text provides detail that fits into that
model and supports the situation model. The situation model was of
course constructed on the basis of that detail, but once the construction
process is over, it becomes the dominant element in the episodic text
structure.
Note that in all the preceeding examples one cannot make a strict dis¬
tinction between situation model and textbase. Certain propositions are
clearly text derived and hence belong to the textbase, but the way they are
organized and indeed their very strength in the overall structure depend
on the situation model that has been created of a woman getting annoyed
while grocery shopping. A reader without knowledge about shopping in
a supermarket who has not experienced the annoyances of such places
could and would not create this kind of structure. His or her memory rep¬
resentation would therefore be quite different - more like a pure textbase.
 Modeling comprehension processes 117

To calculate the long-term memory strength of the nodes in the net¬

work (section 4.2), all activation values are squared, multiplied by the
node strength in the network (which is always 1), and summed if a node
participates in more than a single cycle. These strength values are most
informative if calculated for complex propositions rather than for each
propositional element separately, except that time or location slots are
best kept separate (for purposes of scoring recall protocols). The result¬
ing strength values are shown in Figure 4.11. For SI, the strength of the
main proposition and the time element are shown separately. The other
sentences correspond to complex propositions, except sentences S2 and
S4, in which two propositions are linked by a because and and, respec¬
tively.
The strength values shown in Figure 4.11 can form the basis for vari¬
ous experimental predictions, as is illustrated in later chapters. For
instance, they could be used to predict the rank order of recall frequen¬
cies for the complex propositions in this text. Or they could be used to
form a predicted summary, which, if based on the three strongest propo¬
sitions, would be something like

S8-carry
S7-large-biIl
S6-fill-cart
S5-upset
S4-picked-fruit
S4-wilt-lettuce
S3-crowd-store
S2-full-lot
S2-parking
SI-Sunday
SI-grocery-shop
0 12 3 4

Memory Strength

Figure 4.11 Long-term memory strengths for the propositions of the gro¬
cery-shopping text.
118 The theory

(8) This woman went grocery shopping. She got upset because it
was crowded.

Indeed, a variety of theoretical summaries could be formed in this way,

with a relative likelihood associated with each one.

4.4 Earlier versions of the model

Although much has changed since Kintsch and van Dijk (1978) began
work on a process model of comprehension, the basic framework of that
model is still the same today: the propositional representation, the cycli¬
cal processing, the distinction between the micro- and macrostructure.
But the model has evolved and, indeed, changed in some important ways.
The year of the mental model was 1983. Johnson-Laird published his
influential book of that title, and van Dijk and I introduced the related
notion of situation model (van Dijk & Kintsch, 1983). This was a major
change in the theory, which shifted the emphasis away from the text itself
to the knowledge/text interaction. This emphasis has been retained and
strengthened here.
A more technical innovation made by van Dijk and Kintsch (1983) was
the use of complex propositions in contrast to atomic propositions, as in
the earlier work. However, in actual modeling work I continued to use
atomic propositions for some time thereafter, and only in the present
work have I employed complex propositions more consistently.
In 1988 the Cl architecture was developed (Kintsch, 1988). It involved
a shift from a schema-based control system to a bottom-up system in
which context effects were viewed as constraint satisfaction. This made it
possible for the first time to deal seriously with the use of knowledge and
memory in comprehension. In the present book I have tried to develop in
detail the program laid out in 1988. In particular, the theory of long-term
working memory of Ericsson and Kintsch (1995) is used to specify the
details of the use of knowledge and memory in comprehension, as detailed
in chapter 7. Thus, the new framework has significantly modified earlier
views of working memory management, the use of episodic text memory
and knowledge in comprehension, and inferencing.
In the present book I also introduce Latent Semantic Analysis (LSA)
as an alternative to propositional representations. It is consistent with the
processing theory presented here, as I have tried to show with the various
 Modeling comprehension processes 119

illustrative examples to be analyzed in Part II of the book. Propositions

have the advantage that we can interpret them easily, but they depend on
hand coding. Thus, they are not suitable for large-scale applications
because automation is impossible. The vectors of LSA, like distributed
representations in a neural net, are not as readily interpreted, but they
can be .constructed automatically and objectively and do not impose lim¬
its on the size of the texts to be analyzed. It may be advantageous for the
theorist to think in terms of propositions because of their concreteness,
but the LSA format offers advantages of scale, objectivity, and auto-
maticity that open up new horizons for the theory of comprehension.

A process model of text comprehension attempts to describe the

step-by-step processes by which written or spoken language is
transformed into a mental representation in the reader’s or lis¬
tener’s mind. The construction-integration (Cl) model assumes
that this process involves two phases: a construction phase, in
which an approximate but incoherent mental model is con¬
structed locally from the textual input and the comprehender’s
goals and knowledge, and an integration phase that is essentially
a constraint satisfaction process that rejects inappropriate local
constructions in favor of those that fit together into a coherent
whole. The construction rules in this model can be relatively
simple and robust because they have to take into account only the
local context. The global context becomes important only in the
integration phase, when the tentative, incoherent network that
has been formed by the context-free construction rules settles
into a stable state.
The Cl model differs thus from a strictly top-down, schema-
controlled process that would require more sophisticated, fine-
tuned, context-sensitive construction rules. Such rules tend to
be brittle and unreliable in the face of the complexity and vari¬
ability in the real world. Human comprehension processes do
not appear to work this way but employ simpler and more robust
local rules in conjunction with a holistic integration process.
Mental models of texts have two sources: the text itself, and
what readers contribute to it from their knowledge and experi¬
ence, after their own needs. Accordingly, it is often useful to dis¬
tinguish between a textbase, which more or less reflects the text as
120 The theory

it is, and a situation model, which includes the reader’s elabora¬

tions. Different readers on different occasions will weight these
two aspects of the text representation differently. The contrast
between textbase and situation model is one of the main themes
in this book, recurring in one form or another in just about every
chapter to come.
The presenfchapter merely introduces a framework for build¬
ing models for specific comprehension tasks. In Part II of the
book, such models are constructed and empirically evaluated for
a wide variety of tasks, ranging from simple word identification
in discourse to learning from text, action planning on the basis of
verbal instructions, and judgments based on verbal cover stories.
Part II

Models of comprehension
 5

Word identification in discourse

5.1 The construction of word meanings

Newell (1990) observed that although the simplest experimental

tasks may have the least ecological validity, they tend to provide
the best opportunities for empirically testing assumptions about
the basic architecture of a theory. Experiments on word identifi¬
cation in discourse have played this role for the construction-
integration (Cl) model. The primary claim of that model is that
construction processes are relatively independent of context and
that the contextual fine-tuning is achieved in a subsequent inte¬
gration phase. Studies of word identification allow us to test
these predictions.

How is the meaning of a word in a discourse constructed? Are there

differences when words are used literally and when they are used meta¬
phorically? How are anaphora identified, and how does the structure of a
discourse emerge from this process? Rich linguistic and psycholinguistic
results are available to significantly constrain answers to these questions.
I sketch the principal features of these data and describe simple simula¬
tions with the Cl model to show that the model provides a suitable frame¬
work for the discussion of how word meanings are constructed in dis¬
course.
The process of identifying a word - in reading or listening - involves a
complicated sequence of analyses - perceptual as well as conceptual - and
takes considerable time to complete. As we shall see later in this section,
the construction of a word meaning in a discourse context takes approxi¬
mately 350 ms. Although an eye fixation during reading lasts only around
124 Models of comprehension

200 to 250 ms, word meanings in a discourse context need more time than
that to stabilize. This means that meaning construction at the word level
extends beyond the fixation time. In general, readers attempt to interpret
whatever structure they encounter - lexical, syntactic, semantic, or dis¬
course - immediately rather than engaging in a wait-and-see strategy (for
a detailed review, see Just & Carpenter, 1987.) However, this means only
that they begin the interpretive process as soon as possible. It is not nec¬
essarily the case that each word is fully encoded while it is being fixated.
During the fixation, the perceptual construction processes are initiated
and largely completed. According to the Cl model, these do not produce
a final product - a particular word meaning - but must be integrated into
the discourse context. This process of contextual constraint satisfaction
starts immediately, as soon as anything has been constructed to be inte¬
grated, but may not be completed until sufficient local context is available
to allow the integration process to stabilize.

5.1.1 From the retina to working memory

The sequence of intermediate perceptual representations generated dur¬

ing reading has been studied extensively (see, for instance, Potter, 1983).
First, light receptors register visual information on the retina; then, in a
series of transformations, the neural information is analyzed en route to
the associative cortex. Higher levels of representation are derived from the
lower-level representations as the information at each lower level is main¬
tained briefly in associated memory buffers. What is stored in these mem¬
ory buffers are temporary representations, which are accessible to other
cognitive processes only in highly restricted ways. Some low-order buffers
can be accessed only by certain specialized processes. Their contents are
never consciously experienced, and their existence can be demonstrated
only by special experimental methods. The contents of other buffers, on
the other hand, may be transferred to the level where conscious experience
is possible; but at that point they become contents of working memory,
characterized by the same properties as other contents of working mem¬
ory. 1 hus, some results of intermediate perceptual and conceptual pro¬
cessing are only temporarily buffered and by their very nature are transi¬
tory and unconscious. Others are normally so but may become cognitive
end products in their own right and are thus consciously experienced.
Some of these intermediate buffers receive input not only from the
 Word, identification in discourse 125

sensory, perceptual, and conceptual processes but also from working

memory directly. That is, they serve as slave systems for the central exec¬
utive, as Baddeley (1986) has shown. These are the articulatory loop and
the visual-spatial scratch pad, which are capable of recycling special sorts
of information under the control of the central executive.
Thus, putting some of the notions of Potter (1983) and Baddeley
(1986) together, we arrive at the following sequence of intermediate
buffers in reading a text. The neurophysiological bases of the first three
buffers listed below, including the differentiation between space and pat¬
tern information, are relatively well understood. The remaining memory
buffers with higher levels of representation involve learned recodings of
visual stimuli.

1. Retinotopic icon. This buffer is the perceptible continuation of a sin¬

gle presented word resulting from photoreceptors and other neural
mechanisms.
2. Spatiotopic visual memory. Retinal information from successive eye
fixations is integrated at this level of processing. The printed text is rep¬
resented as a stable structure located in space.
3. Reatopic visual memory. Spatial characteristics of the retinal infor¬
mation are less relevant at this level of transformation than are the con¬
figuration of visual features and patterns of the perceptually available text
segment. Even when this text segment is removed and replaced with an
irrelevant stimulus (visual mask), information about the original text seg¬
ment is retained in this memory buffer for several seconds.
4. Conceptual buffer. At this level of representation, words and objects
are perceived and understood. Experimental results show that words can
be momentarily understood but are then lost because of interference from
other cognitive processes. Conceptual representations become conscious
and receive permanence only when transferred to working memory.
5. The articulatory loop/acoustic buffer. Even skilled, adult readers
transform visual information in reading into an acoustic form (we ignore
here the debate whether this level of representation is best understood as
acoustic or articulatory-motoric). Articulatory representations not only
are intermediate results in the sequence of cognitive processes but also
may become contents of working memory, and material from working
memory can be recycled in the articulatory loop. Thus, this buffer may
function as a slave system for working memory (Baddeley, 1986).
126 Models of comprehension

6. The visual-spatial scratch pad. This is Baddeley’s (1986) other slave

system, which has much the same function in the visual-spatial domain
as the articulatory loop in the acoustic domain.
7. Working memory. This is Baddeley’s (1986) central executive. It com¬
prises the entire cognitive work space (here I deviate from Potter, 1983)
where information about the current text segment is stored in rapidly
accessible form. Information in working memory is the cognitive end
product of all the foregoing processes but is itself complex and structured
and involves different levels of representation.

Figure 5.1 (modified from Potter, 1983) illustrates this temporal

sequence of buffers. Positron emission tomography (PET) and event-
related potential (ERP) studies indicate activity in the right posterior cor¬
tex during visual feature analysis (Posner & McCandliss, 1993), begin-

I'igure 5.1 Intermediate processing stages with their associated memory

buffers in word identification.
 Word identification in discourse 127

ning at about 100 ms. By 200 ms there is activity in the left-lateralized

visual word form area if the strings analyzed are word candidates, but not
otherwise. The time necessary to encode acoustical information can be
estimated as about 175 ms: A reader needs at least 400 ms to pronounce a
word, 225 of which appear to be needed for articulation (if a subject
knows which word is to be pronounced, the time to articulate it is 225
ms). Conceptual processing begins about 100 ms after the fixation of a
word. At 250 ms, PET and ERP results show diffuse activation in frontal
brain regions of both hemispheres. Priming studies (e.g., Swinney, 1979)
indicate that it takes about 350 ms for the meaning of a word to be fixated
in context. PET studies indicate that if subjects are given enough time,
the activity associated with semantic processing shifts from the frontal
regions to posterior regions (Posner & McCandliss, 1993).
The sequence of buffers shown in Figure 5.1 is critical to the integra¬
tion and comprehension of segments of text, such as phrases and sen¬
tences. Deficits in any one of these storage buffers may have complex
results beyond the buffer in question. Thus, phonological storage deficits
affect not only the acoustic/articulatory buffer itself but also general
learning and comprehension capacities. Baddeley, Papagno, and Valla
(1988) observed a patient with such a deficit and found a normal ability
to associate meaningful, familiar word pairs, but an almost total inability
to associate words with unfamiliar sounds (words in another language).
Apparently, the patient could process familiar sound patterns adequately
without having to maintain them in the acoustic buffer. But the patient
did not have enough time to analyze unfamiliar sound patterns without
maintaining them at least briefly in the acoustic buffer. Baddeley and Wil¬
son (1988) observed a similar patient who had no trouble identifying sin¬
gle words and comprehending brief sentences, even syntactically com¬
plex ones. This patient, however, could not understand longer sentences
because, although comprehension abilities were intact, the patient could
not maintain speech sounds long enough in the acoustic store to complete
a successful analysis. Analysis was unsuccessful, for instance, when an
early word in a sentence could not be disambiguated until late in the sen¬
tence.
Figure 5.1 represents a best guess at this time. I am not strongly com¬
mitted to this particular sequence of buffers and their time course, or to
the way the arrows between these buffers have been drawn. Further
research will change some of this detail; the important point concerns the
128 Models of comprehension

existence of such a sequence of intermediate buffers and their relation to

working memory. It is, of course, a highly significant research question to
come up with a more accurate version of Figure 5.1, but it is not my ques¬
tion. What I want to focus on here is the question of how and where
context effects enter into this picture. The Cl model makes some clear
predictions in this respect that can be tested empirically. The test case
involves the identification of lexically ambiguous homographs.

5.1.2 Homographs

Many studies in recent years have explored how lexically ambiguous

words are understood in a discourse context. Various forms of schema
theory predict that the discourse context works like a filter that selects the
contextually appropriate meaning of a homograph, so that the inappro¬
priate word sense is never activated. This theory has the great advantage
that it corresponds to our everyday experience. As readers and listeners
we encounter numerous words that have more than one potential sense,
but we usually do not become conscious of these contextually inappro¬
priate meanings. The alternative theory, suggested by the Cl model as
well as by modular conceptions of cognition (Fodor, 1983), maintains that
all meanings of a homograph are accessed initially but that the contextu¬
ally appropriate meaning wins out before any of the alternate meanings
reach consciousness. That is, the construction process itself is dumb
and does not take the word context into account; it serves as a source of
constraints, however, quickly eliminating inappropriate alternatives. This
is the exhaustive access model, which seems best to correspond to the
experimental facts.
The original demonstration of exhaustive access of homograph mean¬
ings was reported in 1979 by Swinney, who used a cross-modal lexi¬
cal decision task in his study. Subjects listened to a disambiguating text,
such as a man surreptitiously entering an office and planting a bug. Just
as the word hug was heard, a letter string appeared on a screen in front of
the subject, who had to decide by pressing an appropriate response key
whether the string was an English word or not - that is, the subject made
a “lexical decision.” Four types of letter strings were presented: non¬
words as a control, unrelated words, close associates of the contextually
appropriate meaning of the homograph (e.g., spy), or close associates of
the contextually inappropriate meaning of the homograph (e.g., spider).
 Word identification in discourse 129

The results were most interesting: The lexical decision times to words
that were associates of either sense of the homograph were both primed,
that is, they were about 40 ms shorter than the lexical decision times for
unrelated control words. In further work, it was found that if the lexical
decision task is not presented immediately after the target word but
delayed for 350 ms, the contextually inappropriate associate is no longer
primed. Swinney concluded that all meanings of a homograph were
accessed initially but that the context quickly selects the appropriate one,
so that the inappropriate one never even enters consciousness.
Following Swinney’s original demonstration, a large number of stud¬
ies of lexical access have appeared using either the lexical decision task or
simply a naming task. Swinney’s findings were replicated repeatedly, but
the opposite result was obtained in a number of experiments, too. Despite
this confusion, the issue seems to be settled today: Initial lexical access
does appear to be exhaustive and context independent, at least for homo¬
graphs with two balanced meanings. In a recent review of that literature,
Ravner, Pacht, and Duffy (1994) conclude that the most likely reason for
the repeated findings of context effects is that both the lexical decision
task and the naming task are subject to postlexical influences. Thus, what
appears to be a context effect on lexical access is really an effect of the
context on the integration of the word into the discourse context subse¬
quent to lexical access. They argue convincingly that measures of eye
movement provide a better and less biased index of lexical access, and
they report some data that let us observe both the context-independent
access to multiple meanings of homographs and the process of contextual
integration that begins the moment the multiple meanings have been
generated.
Rayner et al. (1994) measure the fixation time on homographs embed¬
ded in a disambiguating prior discourse. They find that readers fixate
longer on a homograph that has two meanings, one of which is much
more frequent than the other, when the subordinate meaning is instanti¬
ated, than on balanced homographs or on unambiguous control words
matched for length and frequency. They call this the subordinate bias
effect. Presumably, the more frequent meaning is accessed first, and the
longer initial fixation time is required to access the subordinate meaning
required by the context.
Now what happens if a biased homograph is repeated, both times in its
subordinate sense? If context can constrain access, it makes no sense for
130 Models of comprehension

the reader to initially re-access the primary meaning of the homograph

that has just been rejected. In actual fact nothing happens. Initial fixation
times for the repeated homographs are just like the initial fixation times
when the word was first read (290 ms for repeated homographs and 295
for first reading - both significantly longer than the fixation times for
neutral control words, 275 ms). Thus, there is no selective access to the
contextually appropriate word meaning when a homograph is repeated;
instead, the reader must go through the same exhaustive access process as
when the word was read for the first time. However, following this con-
text-independent access, which is reflected in the initial fixation time, the
integration process for homographs in the subordinate meaning is much
faster when the words are repeated than when they are read for the first
time: Total fixation time is 336 ms for repeated homographs, but 408 ms
when the homograph is read for the first time. Thus, the discourse con¬
text effects are there and have a clear and immediate influence on total
fixation time, though not on initial access. The process of construct¬
ing the alternate senses of a homograph is context-free - following the
dumb construction rules of the Cl model - but the contextual constraints
become effective right away to ensure the integration of the word mean¬
ing into the ongoing discourse.

5.1.3 Local and global priming in discourse: What words mean

Our own studies of word identification in discourse are in good agree¬

ment with this picture that has emerged from the literature. Kintsch and
Mross (1985) developed an all-visual lexical priming task in which sub¬
jects read a text and were interrupted at predetermined points for a lexi¬
cal decision. Their results clearly favored an exhaustive access over a
selective access model. Immediately after reading a balanced homograph
in a disambiguating context, associates to both word senses are activated,
but after 570 ms only the context-appropriate sense remains activated.
Kintsch and Mross used texts based on simple, familiar scripts. Thus,
they were able to investigate whether words that were part of the script
but not associatively related to the target word were also primed immedi¬
ately after the presentation of the target word. For instance, in a text
describing someone boarding a plane, right after the priming word
plane either an associated word - fly - or an unassociated word -gate -
that was part of the boarding a plane script was presented. No priming
 Word identification in discourse 131

was obtained for unassociated thematic words. Priming was clearly a mat¬
ter of local associations. The theme neither helped to prime unassociated
words related to it, nor did it suppress associated words unrelated to it.
Till, Mross, and Kintsch (1988) extended these results by including
not only local associates but also thematic associates as test words in
their study. An example of one of their experimental sentences was the
following:

(1) The townspeople were amazed to find that all the buildings had
collapsed except the mint.

After reading this sentence, subjects were given a lexical decision task
with either one of four test words (nonword strings were, of course, also
used in the experiment):

• money - an associate of the contextually appropriate sense of the homo¬

graph mint
• candy - an associate of the contextually inappropriate sense of the
homograph mint
• earthquake - a thematic inference
• breath - an unrelated control word

These test words were presented either 200 ms after the last word in the
sentence, 300, 400, 500, 1,000, or 1,500 ms later. Figure 5.2 shows the
amount of priming that was obtained for these test words as a function of
the stimulus onset asynchrony (SOA) as the difference between the lexi¬
cal decision time for contextually inappropriate associates minus the lex¬
ical decision time for contextually appropriate associates. For short
SOAs, the Till et al. (1988) data replicate the results of Swinney (1979)
and Kintsch and Mross (1985). Both the contextually appropriate associ¬
ates and the inappropriate associates are equally activated initially. We can
explain this in the following way. The process of word-meaning con¬
struction is completed by 350 ms: The contextually appropriate word
sense has been established, and its associate money remains activated.
Because the alternate meaning is now deactivated, no priming effect is
obtained for candy at this point.
However, globally relevant inference words corresponding to the
theme of the sentence do become activated after a certain time even in the
absence of any local associations, unlike the scriptally relevant test words
132 Models of comprehension

Inferences

Associates

Figure 5.2 Time course of priming effects for associate and inference tar¬
gets. Points represent the mean priming effect: response latencies for contex¬
tually inappropriate targets minus response latencies for contextually appro¬
priate targets. After Till et al. (1988).

in Kintsch and Mross (1985), which had no global relevance.1 Inference

activation was measured by the difference in the lexical decision times for
control words minus those for inference words. The thematic inference
was first activated at the 1,000 ms SOA. The construction of a thematic
inference apparently takes this much time. Presumably, the construction
of a thematic inference is part of the processing that goes on during the
well-known sentence wrap-up time, the extra reading time that is usually
observed at the end of a sentence (Figure 5.2).
Long, Oppy, and Seely (1994) reported an experiment that extends and
replicates Till et al. (1988) by including a 750 ms SOA. They observed
significant associative priming already at 300 ms. Evidence for topic
inferences was obtained earlier than in the Till et al. study: At the 500 ms
SOA, appropriate topic words were responded to significantly faster than
inappropriate topic words. This effect was even more pronounced at the
750 ms SOA. Thus, perhaps the best estimate we have at this time for the
fixation of word meanings is 300 to 350 ms, whereas 500 to 750 ms appear

1 E.g., ugate” is part of the boarding a plane script but may not be topically relevant in
a story about someone almost missing a plane.
 Word identification in discourse 133

to be required for topic inferences. Meaning construction is not an

instantaneous process. It takes time, and because it occurs in continuous
interaction with the discourse context, it is a dynamic process.
The theory of meaning implied by these analyses is highly construc¬
tive and contextual. The clearest way to see this is to frame the discussion
in terms of latent semantic analysis (LSA), which was introduced in sec¬
tion 3.3. In Kintsch and Mross (1985), we theorized that word identifica¬
tion involves the stages of sense activation and sense selection. Given a
homonym like mole, the different senses molex and mole2 are looked up in
a mental lexicon and then the contextually appropriate sense is selected.
Latent Semantic Analysis implies a very different scenario that is more in
agreement with the analyses presented in this chapter. A word meaning-
in LSA is a vector in a multidimensional semantic space. The vector for
mole would be somewhere in between those for molex and mole2, right in
the middle of no-man’s-land; there are no separable senses that could be
selected. Instead, when mole is used in a sentence, its meaning vector
combines with the vectors of the other words in that sentence to form a
new vector, the centroid of the word senses if all words are weighted
equally. Mole in this context now has meaning, but its meaning is no
longer separable from the context.
Whether a word is a homonym makes no difference for this argument.
A word that has only a single sense in the traditional view is represented
in LSA in the same way: as a vector in the semantic space. When it is used
in a sentence context, it contributes its share to the sentence vector, merg¬
ing with it and losing its identity.
In a master’s thesis at the University of Colorado, Steinhart (1996)
made some interesting observations. For certain types of sentences he
found no priming effects, even though the target words were strongly
associated with the final word in all the sentences according to associative
norms, which he also collected. This failure is puzzling in the traditional
view that word identification amounts to looking up the right sense of a
word but is readily explained by a contextual account in which the mean¬
ing construction is inseparable from the context. Consider first an exam¬
ple from Till et al. (1988) for which reliable priming effects are obtained:

(2) The gardener pulled the hose around to the holes in the yard.
Perhaps the water would solve his problem with the mole.

Targets: ground and face

134 Models of comprehension

Landauer and Dumais (1997) have already shown that in an LSA simula¬
tion of the Till et al. (1988) materials the cosines between the target words
and the experimental sentences mirror the obtained experimental results.
We go a step further here by incorporating their semantic relatedness esti¬
mates into a construction-integration simulation, which allows us to
model the changing time course of priming. Table 5.1 shows the cosines
obtained from LSA, that is, the strength of the semantic relation, between
the prime mole and the two target words, ground and face. Both targets are
reasonably closely related to the prime; empirically, both are strong asso¬
ciates of the prime. In separate Cl simulations, these two target words were
then linked to the words in (2) according to their cosine values. The results
of the spreading activation process are also shown in Table 5.1. Ground is
strongly activated, because it has high or moderate cosines with several
words in (2), whereas face ends up with a low activation value, because its
cosines with all of the words in (2) except mole are low.
The implications of this analysis are in agreement with the Till et al.
data: There is immediate priming for both ground and face because of
their relatedness to mole. But once mole has been integrated into the sen¬
tence context, only ground will be primed; face is no longer semantically
related to the mole-in-the-yard.
Now consider an example of the kind Steinhart used:

(3) For his birthday last year, I wanted to give my brother something
ugly. But no pet store in town would sell me a mole.

Targets: ground and face

Table 5.1. Cosines and results of spreading activation process

cosine

mole - ground .35

mole-face .24

activation

(2) + ground .49

(2) + face .06
(3) + ground .08
(3) + face .06
 Word identification in discourse 135

The two target words can be linked up with (3) in the same manner,
employing the cosine values between the targets and the words in (3) as
link strengths. However, now ground, the target related to the contextu¬
ally appropriate sense of mole, fares no better than the target related to the
inappropriate sense: All the cosines are low in both cases, as are the acti¬
vation values resulting from a Cl simulation, also shown in Table 5.1.
It does not seem that word identification is a look-up process in a men¬
tal lexicon. Otherwise, the ground-mole sense should have been selected in
(3) as well as in (2), and ground should have been primed in both cases,
because we know that it is a strong associate. On the other hand, suppose
that in word identification there are no “senses” of words to be picked out
of a lexicon. If words and sentences are represented as vectors in a seman¬
tic space, mole is a weird mixture of the ground-mole and the face-mole, by
itself priming ground as well as face. Once mole is integrated into the sen¬
tence context of (3), it is no longer the ground-mole but a new construc¬
tion, the ugly-pet-mole. The ground-mole burrows holes into the ground
and annoys gardeners; none of this is important for the ugly-pet-mole — it
is ugly and supposed to annoy my brother! Reading (3) does not select a
particular sense of mole but constructs a new meaning of mole meticu¬
lously attuned to its context.
Results in the memory literature lend support to Steinhart’s priming
data. For instance, Barclay et al. (1974) had subjects study sentences like
The man lifted the piano or The man tuned the piano. Later, two kinds of
recall cues were provided: either Pianos are heavy or Pianos make nice
sounds. The subjects had to recall the study sentences. Depending on
which form of the study sentence they had read, one or the other of these
cues was appropriate. Recall was significantly better (80%) with the
appropriate cue than with the inappropriate cue (61%). Similar results
were obtained in an experiment by Anderson and Ortony (1975), who
found that the sentence Television sets need expert repairmen was recalled
best with the cue appliance, and the sentence Television sets look nice in the
family room was recalled best with the cue furniture. Clearly, both appliance
and furniture are potential components of the meaning of television set, but
only the contextually relevant feature was actually used in comprehension.
These memory data with unambiguous words support Steinhart’s
finding that priming effects for homonyms are influenced by their con¬
textual use. However, negative results like Steinhart’s are at best sugges¬
tive and need to be replicated with appropriate controls. If they do hold
 136 Models of comprehension

up, this would constitute strong evidence for the position advocated here.
Word meanings are not something to be pulled from a mental lexicon but
are ephemeral constructions that are generated during comprehension.
In the next section, I explore in a little more detail the nature of this con¬
struction process as it is conceived by the Cl model.

5.1.4 A simulation of priming effects in discourse2

Meaning construction is a bottom-up process that starts rather indis¬

criminately (the multiple activation of word senses). Meaning construc¬
tion takes time (word meanings are fixed by 350 ms, but sentence themes
require about twice as much time). Nevertheless, the process of meaning
construction is highly interactive, with local and global factors jointly
determining its course from the very beginning. Some interesting exper¬
imental results by Schwanenflugel and White (1991) and a simulation of
these results by Kintsch (1994a) show how local and global processes
occur at the same time and illustrate the complex dynamics of a growing-
network.
The brief text presented in Table 5.2 is taken (with slight simplifica¬
tions) from a study by Schwanenflugel and White (1991). Consider how
the Cl model would simulate comprehension of this text and predict how
it would be recalled. Let us assume that we want to simulate comprehen¬
sion processes at the level of sentences or phrases, that is, in six process¬
ing cycles, as shown in Table 5.2. The first step is to construct a proposi¬
tional representation, resulting in the list of atomic propositions also
shown in the table.
Propositions PI to P13 are connected in a network by argument over¬
lap, and we could proceed cycle by cycle to calculate the outcome of the
integration process in this network. However, the construction phase of
comprehension is not yet finished, for we have not yet considered the
construction of a situation model, the construction of a macrostructure,
or associative knowledge activation processes.
In the present example, the strategic construction of a macrostruc¬
ture involves the following steps. A hypothesis is made that the first sen¬
tence contains the macropropositions, and later statements are recog¬
nized as instances of the generalization given by the first sentence. This

2 This section is adapted from Kintsch (1994a).

 Word identification in discourse 137

Table 5.2. Example of a text and proposition list

1. The bar was designed for male professionals

2. and turned away all secretaries and nurses who tried to get in.
3. It had an exclusive look to it.
4. The bar was in a nice place located in the business district in town.
5. It was perfect to take clients to the bar.
6. The bar would not serve drinks to women.

Proposition
Cycle Number Number Proposition

1 PI DESIGN[BAR,P2]
1 P2 MALEfPROFESSIONALS]
2 P3 TURN-AWAY[BAR,SECRETARIES,NURSES]

2 P4 TRY[SECRETERIES,NURSES,P5]

2 P5 ENTERfSECRETARIES,NURSES]

3 P6 LUXURIOUSfBAR]
4 P7 LOCATION[BAR]
4 P8 IN[P7,NICEPLACE]

4 P9 IN[P7,BUSINESSDISTRICT]

4 P10 IN[P7,TOWN]
5 Pll PERFECT[BAR,P12]

5 P12 TAKE[ CLIENTS,BAR]

P13 NOT-SERVE[BAR,DRINKS,WOMEN]
6

Source: From Schwanenflugel & White (1991).

strategy is plausible in the present case because in the expenment fiom

which this text was taken, many texts of the same structure were read by
the subjects.
In the simple, brief text used here, the situation model and the
macrostructure of the text are largely the same. A great deal of domain
knowledge is necessary for the construction of the macrostructure, so
that in this process a situation model would also be obtained. One has to
know what it means for a bar to be designed for male professionals in
order to recognize that turning away secretaries and taking clients to the hat
are instances of the general concept. Thus, we obtain a situation model in
the form of a bar-for-male-professionals script with five slots — the same
structure that results from the application of the macrostrategies.
Thus, we need to mark PI and P2 as macropropositions and P3, P6,
138 Models of comprehension

P7, P12 and PI3 as macrorelevant. This is done by increasing the

strengths of the links between PI and the macrorelevant propositions by
a factor of 2, as well as the strength of the links from PI and P2 to them¬
selves. The choice of doubling link strengths is arbitrary and is merely
intended to show that certain links in the net should be assigned more
strength than others. Obviously, one cannot obtain precise quantitative
predictions with such guesswork, but qualitative predictions can never¬
theless be derived. The advantage of this procedure is that it avoids the
objection that whatever success the model has is due only to skillful fine-
tuning of the parameters in a very complex model.
Finally, associative knowledge elaboration during comprehension must
be simulated. We have assumed that this occurs independently of context.
Hence, free association data obtained in a situation outside the present
discourse context for the concepts and propositions (expressed as simple
declarative sentences) of the text provide an estimate for which items
of knowledge are most likely to be activated when reading this text. In an
informal experiment, the 12 associations shown in Table 5.3 were
obtained. The 12 associates (k 1—k 12) were then added to the 13 proposi¬
tional nodes constructed in Table 5.3 and connected to their source nodes

Table 5.3. Content words from the experimental text shown in

Table 5.2 and frequent free association responses

Stimulus Item Associated Response

bar drink (kl)

design(person, building) architect (k2)
male(person) female (k3)
professional lawyer (k4), doctor (k5)
turn-away (person,from-building) door (k6)
secretary typewriter (k7)
nurse doctor (k5)
enter(person,building) door (k6)
exclusive(object) fashion (k8)
place country (k9)
business district shops(klO)
take(clients,to-bar) lunch (kl 1)
not-serve(bar,drinks,to-person) minors (kl2)
 Word identification in discourse 139

with a value of .5, to indicate that an item retrieved from long-term mem¬
ory during comprehension receives less weight than one that is explicit in
the text.
We are now ready to start the integration process. The complete net¬
work that has been constructed is shown in Figure 5.3. A cycle size of five
elements was assumed, but knowledge elaborations belonging to a text
proposition were always included in the same cycle as the text proposi¬
tion, even if more than five elements had to be processed in working
memory because of that. A buffer size of 1 was assumed, that is, the most
highly activated proposition on each cycle was carried over to the next
processing cycle. Six cycles were required to process this text under these
conditions. The long-term memory strengths of the 13 propositional
nodes in the network that resulted from these six processing cycles are
shown in Figure 5.4.3
The knowledge elaborations have been omitted from Figure 5.4
because none of them received enough strength to be experimentally sig¬
nificant. However, this does not always have to be the case. Imagine a
reader who is reminded of discrimination every time she reads the six sen¬
tences of Table 5.2. For such a reader, the knowledge elaboration discrim¬
ination would achieve considerable strength. Specifically, the discrimina¬
tion node will have a strength of .62 in the network, the eighth highest
value. Thus, under appropriate circumstances it is quite possible that
knowledge elaborations end up with greater memory strength than actu¬
ally presented items.
In the experiment of Schwanenflugel and White (1991), subjects read
texts like the one shown in Table 5.2 with an RSVP procedure, that is, one
word at a time, appearing centered on a computer screen. It was a dual
task experiment, because in addition to reading the text (and answering
questions about it later), at certain points during reading subjects made
rapid lexical decisions. The point of particular interest here concerns the
last word of sentence 6, namely, women.
We make the assumption that priming effects will occur in the lexical
decision task to the extent that the target words are activated by the con-

3 The normalization procedure used for these calculations required the total sum of
activation to remain constant in each cycle and differs therefore from the one
employed in the Mross and Roberts (1992) program.
Models of comprehension

1'igure 5.3 l he network of propositions and knowledge elaborations for the

text (1) to (6). Plain links have an initial value of 1; heavy links have an initial
value of 2; broken links have an initial value of .5.
 Word identification in discourse 141

DES[BARJP2]
MALE [PROF]
TURN[BAR,SEC,NUR]
LOC[BAR,P9]
PERF[BAR,P12]
LUX[BAR]
TAKE[CLT,BAR]
NS[BAR,DRK,WOM]
IN[P7,NICEPL]
IN[P7,BUSDIS]
IN[P7,TOWN]
TRY[SEC,NUR,P5]
ENTER[SEC,NUR]

0 1 2 3 4 5

Strength

Figure 5.4 Final memory strengths for the propositions shown in Figure
5.3.

text of reading. Suppose that subjects are reading sentences 1 to 6 but that
at the point where the last word should have been presented, a lexical
decision trial occurs with the word women. We need to calculate how
strongly this word will be activated by the memory structure for the text
that already has been constructed at this point. This structure is identical
with that shown in Figure 5.3, except that the last proposition, P13, is
incomplete and has the form NOT-SERVE[BAR, DRINKS, $] - the
recipient of the action has not yet been presented. How will this structure
activate the target word women? Women is connected with the episodic
text structure in two ways. First, it fits semantically into the slot of the
incomplete NOT-SERVE proposition and hence receives activation from
that source. However, we already have a potential slot filler for that propo¬
sition in the form of the knowledge elaboration minors, which became a
part of the net as an associate of NOT-SERVE(BAR,DRINKS,10-
PERSON). Because women and minors are mutually exclusive alternatives
for the same slot, they interfere with each other; thus, we add a link with
a negative weight of-1 between them. Hence, there are two sources of
activation for women, a positive and a negative. If we let activation spread
to it from the incomplete network through the two links just mentioned,
142 Models of comprehension

the positive is considerably stronger because it derives from the dominant

discourse topic, whereas the negative source has only a weak associate to
draw upon.
Now suppose that instead of women the word students had been pre¬
sented as a target in the lexical decision task. Obviously, students also fits
the slot of the NOT-SERVE proposition - The bar did not serve drinks to
students is a perfectly acceptable English sentence - and, therefore, students
also competes with the preexisting associate minors. Hence, it is connected
to the net in exactly the same way as women. However, something else now
changes in the net. When the NOT-SERVE proposition was completed
with women, it formed a macrorelevant statement, as in the original text.
However, the completion student, while semantically acceptable, no longer
fits into the current discourse structure. In the context of the present dis¬
course, not serving drinks to students is not an instance of bar-designed-for-
male-professionals, and hence the link between the discourse topic PI and
the proposition completed with STUDENTS no longer receives extra
weight. As a consequence, the target word students becomes less strongly
activated than women. Thus, the model predicts discourse priming, in that
items that fit into the discourse structure become more highly activated
and hence have shorter reaction times. This was precisely what Schwa-
nenflugel and White (1991) observed, from whose study the present text
was borrowed. A priming effect occurred for discourse-relevant comple¬
tions like women when compared with discourse-neutral control words
like students.
Schwanenflugel and White (1991) also used minors as a target word in
their experiment, in addition to women and students, and observed that
minors was also primed when presented at the end of sentence 6. The
priming effect for minors was more or less the same as for women. Thus,
although minors never occurs as a knowledge intrusion in subjects’ recall
of the text (see Kintsch, 1994a), and the model predicts that it should not
(the calculated LTM strength of the minors node is .001), it nevertheless
yields a significant priming effect simply because it is a locally strong
association. The model predicts this priming effect, too. The target
minors is connected to the episodic text structure in two ways. First, it is
a possible slot filler for the NOT-SERVE proposition, just as the other
two targets were, yet, unlike the other two, it does not compete with the
preexisting knowledge association minors, although it is activated from
 Word identification in discourse 143

that source too. However, as was the case for students, the NOT-SERVE
proposition completed with the object minors is no longer macrorelevant
and hence receives no special activation from the discourse topic. As a
consequence, minors ends up with an activation value not as high as that
for women but higher than the activation of students.
The Cl model therefore correctly predicts at the same time both local
priming.effects that are independent of context and priming through the
discourse context. Purely associative models would predict the former,
schema theories the latter, but what is needed to account for what actually
happens is something like the marriage of bottom-up processes and con¬
textual integration processes as postulated by the Cl theory. Actually, the
theory makes even more powerful predictions, for which, however, no
empirical tests are available as yet. As shown in Figure 5.5, it predicts that
the target minors is strongly activated from the very beginning (via the
preexisting minors that is already a node in the network), but that its acti¬
vation then decreases because it does not fit well into the discourse con¬
text. Women and students, in contrast, start out with zero activation but
then become activated from the discourse structure to different degrees.

Women

Students

-o— Minors

Figure 5.5 The settling process for the activation values of three target
words in a lexical decision task on the Cl model. The targets were presented
in place of the last word of sentence 6.
144 Models of comprehension

Presumably, the Schwanenflugel and White data points fall somewhere

near cycle 3 or 4 in the simulation results shown in Figure 5.5. A more
stringent test of the theoretical predictions would require further exper¬
iments tracing out the complete speed-accuracy trade-off curves.

5.2 Anaphora

The Latin word textus is derived from a verb meaning “to

weave.” The warp of the fabric that is a text are the anaphora:
repeated and reinstated discourse referents that are a major
source of coherence in a text. The manner in which these repeti¬
tions and reinstatements occur is governed by an intricate set of
rules that ensure that a speaker and a listener, or a writer and his
readers, agree on the intended referent and where to look for it.

5.2.1 The psycholinguistics of anaphora resolution

Once a word has been introduced in a text, it is often repeated in some

form, reference is made back to the already introduced concept by means
of an anaphor: by repeating or paraphrasing the word, by a pronoun, or
even implicitly (zero anaphora). What are the main experimental findings
about anaphora, and can they be accounted for within the framework
developed in the previous section?
Anaphora vary on a dimension of lexical content or specificity (e.g.,
Gernsbacher, 1989). Pronouns or demonstratives have little or no lexical
content of their own and rely entirely on their context to provide their
content. On the other hand, definite noun phrases and proper names have
a specific lexical content of their own. Their meaning may be influenced
and modified by the context in which they are used, just as the meaning
of other words that are used for the first time must be constructed in the
discourse context, but unlike pronouns they contribute lexical content of
their own.
As we have seen in the previous section, readers fixate longer on a
homograph that is biased toward its subordinate meaning by the text that
precedes it, compared to an unambiguous control word. This is true even
when the word has been encountered in its subordinate meaning earlier
 Word identification in discourse 145

in the same passage of a text (Rayner et al., 1994). These results suggest
that repeated words are identified in much the same bottom-up way as
words that appear for the first time in a discourse context.
There is, of course, a difference in the way new words and repeated
words are processed. Part of the process of meaning construction for the
repeated word involves retrieving the context in which the word appeared
earlier and incorporating it into the present context. This retrieval of
prior information about a repeated word occurs quite rapidly, roughly in
the same time that it takes to activate and fixate information about the
word from semantic memory. A study by Dell, McKoon, and Ratcliff
(1983) illustrates this process.
Dell et al. (1983) investigated the case in which a discourse referent is
first introduced with one word (e.g., burglar) and then referred to with
another, definite noun phrase (e.g., the criminal). One of their texts was
the following:

(4) A burglar surveyed the garage set back from the street.
Several milk bottles were piled at the curb.
The banker and her husband were on vacation.
Thej criminal slipped2 away from the streetlamp3.

At the positions labeled 1,2, and 3 in the last sentence, subjects were given
a single word recognition test with one of three test words: the prior ref¬
erent of the noun phrase (burglar), a word appearing in the same proposi¬
tion as the prior referent (garage), and a control word from a different sen¬
tence (bottles). The results are shown in Figure 5.6 as the difference in the
recognition times between the test words and the control word. Before the
anaphoric word (position 1), response times for all words were about
equal. About 250 ms after the anaphor (position 2), both the prior refer¬
ent and the word referring to a concept from the same proposition as the
prior referent were recognized significantly faster than the control word,
suggesting that both were activated in the reader’s working memory. At
position 3, the prior referent was still activated, but other concepts from
the prior context of the anaphor were no longer activated.
This pattern of results is roughly what one would expect from the Cl
model. Figure 5.7a shows the network that a reader has constructed
according to the model at position 2: It contains the two surface elements
the criminal and slipped, the corresponding proposition, a bridging infer-
146 Models of comprehension

Burglar

■O' Garage

Figure 5.6 Priming effects (recognition time for test word minus recogni¬
tion time for control word) for prior referent (burglar) and other concepts
from the same proposition {garage) at three positions in text (4).

ence linking criminal to burglar, and the reinstated proposition contain¬

ing BURGLAR from the first sentence, which includes GARAGE. The
argument of the SLIP proposition has been written as CRIMINAL/
BURGLAR to indicate that as a result of a bridging inference and the
retrieval of BURGLAR, CRIMINAL has been identified with BUR¬
GLAR. Integrating the network shown in Figure 5.7a yields an activation
value of 1.00 for CRIMINAL/BURGLAR and .54 for GARAGE.
Hence, both concepts are strongly activated at this point, compatible with
the results in Figure 5.6 for position 2.
Once the whole sentence has been read (position 3), the full network
shown in Figure 5.7b is obtained. In this network, the central concept
CRIMINAL/BURGLAR still has an activation value of 1.00, but the
proposition containing the related concept GARAGE has now become
marginal and its activation value is reduced to .18. It is reasonable to
assume that such a weak activation could not be detected by the experi¬
mental method of Dell et al. (1983).
Explicit anaphora are thus rapidly identified with their prior referent,
reactivating in the process at least temporarily other related concepts
from the prior context. It is probably safe to generalize Dell et al.’s results
to other types of explicit anaphora such as proper names, which are iden-
 Word identification in discourse 147

SLI P[CRI M/bURGL]

A
ISA[BURGL,CRIM]

SURVEY! BURGL,GARAGE]

(a)

ISA! BURGL^CRI M]

SURVEY! BURGL .GARAGE]

(b)

Figure 5.7 (a) The Cl network at position 2. (b) The Cl network at position
3. Circles are surface structure nodes, squares indicate text propositions, and
the triangle indicates a bridging inference.

tified just as rapidly (Gernsbacher, 1989). Thus, explicit anaphora are

treated much like other nonanaphoric words in a text, except that in the
construction of their meaning the context of their prior appearance is
integrated. This seems reasonable and plausible. After all, when readers
encounter a definite noun phrase or proper name, these may be used as
anaphora or they may not. Pronouns, on the other hand, are different;
readers know they are dealing with anaphora (or perhaps kataphora - a
pronoun whose referent has not yet been introduced).
What factors influence pronoun resolution and when does it occur?
148 Models of comprehension

The literature is rich but confusing. The results of many studies that vary
by only a single factor are difficult to interpret because other confound¬
ing factors were not controlled. An overall picture emerges nevertheless.

1. Speakers use pronouns to refer to entities in the focus-of-atten-

tion/short-term-memory/consciousness (Chafe, 1974; Ehrlich, 1980;
Fletcher, 1984).
2. When there are several antecedents, all of them tend to become acti¬
vated. If there are enough contextual cues to disambiguate a pronoun, the
irrelevant referents will become deactivated, but referents are not auto¬
matically identified when no individual one is sufficiently salient (Greene,
McKoon, & Ratcliff, 1992).
3. Syntax does not dominate pronoun resolution. That is, even when a
text is unambiguous (e.g., using she in the context of John and Sally),
other salient referents are still considered (Garnham, Oakhill, & Johnson-
Laird, 1982; Greene et al., 1992).
4. Referents in the discourse focus are more likely to be selected as the
antecedent of a pronoun than referents that are less discourse relevant.
This discourse bias is manifested in several ways. Recency of mention
powerfully biases antecedent selection (Carpenter & Just, 1977; Clark &
Sengul, 1979; Ehrlich & Rayner, 1983). Prior topicalization makes an
antecedent more salient (Anderson, Garrod, & Sanford, 1983; Sanford,
Moar, & Garrod, 1988). Antecedents that are part of the same scene or
the common ground between speaker and listener are favored (Anderson
et al., 1983; Gordon & Scearce, 1995; Greene, Gerrig, McKoon & Rat¬
cliff, 1994; Lesgold, Roth, & Curtis, 1979).
5. Pragmatic information is used in pronoun resolution together with
other types of information from the very beginning (Garrod, Freudenthal,
& Boyle, 1994; Tyler & Marslen-Wilson, 1982), as soon as it becomes avail¬
able (McDonald & MacWhinney, 1995).

Pronoun resolution is therefore best viewed as a multiple constraint

satisfaction problem. Pronouns are used to create the textual web by mak¬
ing the reader engage in processes that serve to strengthen the links
between its various parts. In the next section I show how these processes
are modeled by the Cl theory. The theory is particularly suited for com¬
bining many different effects to arrive at a global decision about meaning.
 Word identification in discourse 149

Hence, pronoun resolution becomes an especially interesting field of

application for the theory.

5.2.2 Anaphora resolution in the Cl model

An experiment that examines the interaction of several different factors

in anaphora resolution is that of Garrod et al. (1994). It is an eye move¬
ment experiment whose dependent variables are the total reading time for
the target sentence containing the anaphor and first-pass fixation dura¬
tions for various regions of the sentence, in particular the anaphor and
verb regions. The factors they vary are (1) whether or not the referent is
in the discourse focus; (2) the form of the anaphor (either an ambiguous
or unambiguous pronoun or an explicit anaphor - that is, a name or a def¬
inite description); and (3) the pragmatics of the verb of the target sen¬
tence that makes one or the other of two possible antecedents more suit¬
able. The results of Garrod et al. (1994) are in general agreement with the
conclusions reached in the foregoing sections — everything matters, and
everything matters from the very beginning. For instance, pragmatic
effects do not come in late, after an anaphor has been tentatively identi¬
fied on the basis of its syntactic properties; instead syntactic and prag¬
matic factors both contribute to the resolution process from the very
beginning, though some effects take longer to develop than others. Fur¬
ther details of their results are discussed in connection with the following
simulations.
One of Garrod et al.’s examples (1994) is as follows:

(5) Flying to America. Jane wasn’t enjoying the flight at all. The dry
air in the plane made her really thirsty. Just as she was about to
call her, she noticed the stewardess coming down the aisle with
the drinks trolley.

Continuation 1: Right away she ordered a large glass of Coke.

Continuation 2: Right away she poured a large glass of Coke.

What concerns us here is how readers interpret the “she” in the two con¬
tinuation sentences - as Jane or as stewardess.
To simulate these conditions in the Cl model, the comprehension of
the text prior to the test sentence must be simulated. The resulting net-
150 Models of comprehension

work is shown in Figure 5.8. It consists of the propositional textbase

(squares; to simplify things, modifiers are merely indicated as “mod”)
corresponding to (5), the situation model based upon the FLYING
schema (triangles), and the two alternative representations of continua¬
tion sentence 1 containing the pronoun “she”: Right away she ordered a
large glass of Coke (circles), which are connected by a negative link of-1
(all other links are +!.)• The nodes constructed during reading of (5),
which are in long-term memory at the time the continuation sentence is
presented, are filled in, and the test nodes are open. The parts of the
FLYING schema that are used here are the prop, the PLANE, and two
roles, the PASSENGER-JANE and STEWARDESS. SERVE is a char¬
acteristic activity for STEWARDESS, activated during reading of (5),
and ORDER is a characteristic activity for PASSENGER, activated by
the continuation sentence.
What is of interest is how the two continuation sentences are processed

mod mod

F'Sure 5 8 The network corresponding to text (5). The continuation sen¬

tence shown is Right away she ordered a large glass of Coke.
 Word identification in discourse 151

in the context of this LTM trace. Each continuation sentence was twice
represented as a main proposition with a modifier: once for JANE and
once for the STEWARE)ESS, connected with an inhibitory link. In the
order case, the test proposition was connected to three situation model
elements: the actors Jane and stewardess as possible antecedents of the
pronoun she and, for the JANE version only, to order because ordering a
drink is’what passengers are supposed to do. In the case of the pour sen¬
tence, it was linked to the two possible antecedents and to serve in the sit¬
uation model for the STEWARDESS version only. Because we are inter¬
ested in the whole course of the integration process and not just the final
outcome, the starting values of the test sentence elements were set to zeio
and those for prior test elements were set to their LTM values. Most of
the prior text was deactivated, and only the three elements directly linked
to the new input, plus the situation model header FLIGHT, were allowed
to participate in the processing of the test sentences. Figure 5.9 shows the
course of pronoun resolution in the two continuation cases. Foi the older
sentence, the interpretation shejane orders a Coke gains quickly in activa¬
tion, whereas shestewardess orders a Coke never gathers much activation at all.
Discourse focus and pragmatic acceptability combine in favor of the pas-

Figure 5.9 Activation values for the two referents Jane and stewardess in she
orders and she-pours sentences.
152 Models of comprehension

senger ordering the drink. The situation is quite different, however, for
continuation 2. Initially, it is Jane, the referent in the discourse focus, that
is more strongly activated, and only after four processing cycles does the
fact that stewardesses are more likely to pour drinks than are passengers
make itself felt. In the end, however, the model arrives at the correct,
intended interpretation in both cases. Asymptotically, shejune orders a Coke
and shestewardess pours a Coke are equally strong. Total reading times for
these sentences in Garrod et al.’s (1994) experiment were not signifi¬
cantly different.
The pattern of results is quite different if the text (5) is altered so that
it contains a steward rather than a stewardess. Now the pronoun she in the
continuation sentences uniquely identifies JANE, but discourse focus
and verb pragmatics still have their effects, as is shown in Figure 5.10.
These simulations were performed in the same way as before, except that
no link was made between shestewardpours a Coke and STEWARD because
of the incongruent pronoun. The model does arrive at the intended ref¬
erent Jane for both continuation sentences, but shejane pours a Coke never
catches up with shejane orders a Coke. There is a significant reading time
difference in this case in the data of Garrod et al. (1994).

I' igure 5.10 Activation values for the two referents Jane and steward in she
orders and she-pours sentences.
 Word identification in discourse 153

Figure 5.11 shows the results of another variation: Here the text con¬
tains steward, but the continuation sentences also contain he. A straight¬
forward, totally unambiguous case, but the discourse focus effects in the
model are powerful. Again, the model arrives at the intended interpreta¬
tion of he, namely, the STEWARD, but it takes some time, and hesteward
orders a Coke never receives much strength. The Garrod et al. (1994) data
show significant differences in reading times. The differences between
Figures 5.10 and 5.11 reflect discourse focus effects: JANE has an LTM
strength of 1.54 and STEWARD of 0.40, and the strong concept (i.e., the
one in the discourse focus, in the terminology of Garrod et al.) is better
able to violate pragmatic constraints than is the weak concept.
Figures 5.9 to 5.11 can hardly be considered decisive tests of the Gar¬
rod et al. (1994) data. It is not obvious just what feature of the simulations
would predict reading times. I have suggested that when the activations
of two continuation sentences reach the same asymptotic value, no read¬
ing time differences should be expected. Another possibility is that read¬
ing times for the continuation sentences might depend on the difference
between how strongly and how long the intended and unintended alter¬
natives were activated in the integration process. Thus, if one adds up all
activation values in these figures for each pair of alternatives and takes the

Jane-order

steward-order

Jane-pour

steward-pour

Figure 5.11 Activation values for the two referents Jane and steward in he-
orders and he-pours sentences.
154 Models, of comprehension

difference (e.g., the sum of activation values for shejane orders a Coke
minus the sum of activation values for shestewardess pours a Coke), one
should get a predictor of reading times. Indeed, these differences are cor¬
related r = .84 with the mean reading times reported by Garrod et al.
(1994). There are, however, only six means, so one cannot make any
strong claims about the goodness of fit of the model.
It is also of interest to examine with the Cl model how explicit
anaphors are identified in the context of Garrod et al.’s (1994) experi¬
ments. Consider, for instance, the case in which the continuations for text
(5) do not involve a pronoun but either a name {Jane) or a definite
description {stewardess). In this case, simulations can be performed just as
before, except that now there is only a single alternative to be considered
for each continuation sentence, either JANE or the STEWARDESS,
whatever the sentence was. Figures 5.12 and 5.13 show what happens:
Explicit anaphora are identified very rapidly, but there are still effects due
to the pragmatics of the verb {Jane orders is activated more rapidly than
Jane pours, and the reverse is true for stewardess). The effects are very
small, however, and Garrod et al.’s (1994) results show no differences.
The discourse focus effect, on the other hand, is a little bigger, as a com¬
parison of Figures 5.12 and 5.13 reveals. Again, Garrod et al. observed no
differences.
Thus, the model provides a framework within which we can simulate
anaphora resolution in considerable detail, in fact, in more detail than the
resolution power of our experimental procedures. These simulations
summarize and organize what we know about pronoun resolution from
the psycholinguistic literature as exemplified by the results of the Garrod
et al. experiments. Furthermore, the simulations could be done without
any really new theoretical machinery. It is true that some things were
done differently here than in other simulations reported in this book. In
particular, because we were interested in the time course of activation, the
test sentence was started with an initial activation value of zero, and prior
elements were given their LTM strengths as starting values. But these are
technicalities, well within the general theoretical framework. They are
not used elsewhere only because they are not needed and would simply
add irrelevant complexities.
It would have been a significant ad hoc theoretical adjustment to intro¬
duce a new concept such as discourse focus into the theory. However, the
Word identification in discourse 155

Figure 5.12 Activation values for the referent Jane in Jane-orders and Jane-
pours sentences.

Figure 5.13 Activation values for the referent stewardess in the-stewardess

orders and the stewardess-pours sentences.
156 Models of comprehension

model could do very well without it because this concept was implicit in
the Cl theory all the time. Currently activated discourse entities that have
acquired a high LTM strength in the process of comprehension are said
to be in discourse focus. Thus, one can talk about discourse focus in the
framework of the Cl theory, as done here, where the use of that term was
natural and convenient. However, discourse focus is not really a concept
needed by the Cl theory. For many theories of text comprehension, dis¬
course focus is a central, explanatory concept (e.g., Greene et al., 1994;
Grosz & Sidner, 1986; Sanford & Garrod, 1981). The Cl model is more
parsimonious in this respect, for discourse focus is simply a concept that
falls right out of its basic architecture.
There is another context in which simulations such as the ones per¬
formed here may become relevant. Within functional linguistics, there
are several studies that attempt to describe the factors that control the use
of anaphora in discourse. This is essentially correlational work: A partic¬
ular discourse characteristic is observed to be correlated with the use of,
say, zero anaphora, pronouns, or full noun phrases. The most basic of
these factors is linear distance (Givon, 1983). If few words intervene
before a concept is repeated, many languages employ zero anaphora. If
more words intervene between concept repetitions, pronouns are used,
unstressed pronouns or stressed ones in English if the distance is larger.
If the number of intervening words is high, noun phrases are normally
used. These default cases can be overridden for specific linguistic pur¬
poses. For instance, pronoun use when zero anaphora would suffice may
signal a topic shift (e.g., Fletcher, 1984; see also section 6.2.1). Thus, lin¬
ear distance is by no means the only determinant of pronoun use. Fox
(1987) shows that a theoretically motivated measure based on the distance
in a theoretical structure provided a better fit than linear distance mea¬
sures. A host of additional syntactic and semantic factors have been iden¬
tified. These include, for example, the role of the antecedent (e.g.,
whether the antecedent concept was the subject or actor, whether it was
the protagonist of a story, whether it was animate or not, and so on). It has
been shown that not only linear distance in words matters, but also
whether or not a paragraph boundary intervenes. By putting all these fac¬
tors together in a multiple correlation, good accounts of pronoun use can
be obtained.
It may be conjectured that the activation value in working memory of
 Word identification in discourse 157

a repeated concept, as calculated in the foregoing simulations, reflects all

these factors and, together with some considerations about the use of
working memory in discourse comprehension (see chapter 7), could be
used as the basis for determining the choice of anaphora in discourse. To
test this conjecture, sufficiently extensive simulations would have to be
performed. If the prediction envisaged here were successful, we would
have a highly parsimonious theory of anaphora selection. No special
mental computations would be required when a speaker or writer uses
anaphora. The same processes that generate the discourse in the first
place also determine which anaphor will be selected.

5.3 Metaphor

Metaphor, setting our mind to flying betwixt one Genus and

another, allows us to discern in a single Word more than one
Object.
Umberto Eco, The Island of the Day Before

The process of meaning construction described so far has focused on

literal meaning. Because figurative thought and figurative language are
ubiquitous, however, one cannot discuss meaning construction without
taking into account nonliteral meaning. It is no longer the case that
metaphors are considered a peculiarity of poetic language that might as
well be disregarded; rather, we have come to realize that figurative lan¬
guage exists everywhere we look (e.g., Gibbs, 1994), as does figurative
thought (e.g., Lakoff, 1987).
How do people understand figurative language, metaphors in particu¬
lar? Metaphors consist of some kind of comparison by means of which
features are transferred to the metaphor topic that are not normally asso¬
ciated with it. Thus, in Sermons are sleeping pills a low-salience feature ol
sermon is pushed into the foreground, whereas in Atoms are little solar sys¬
tems certain properties of solar systems are assigned to atoms. How does
the comprehender know to make such a feature transfer, rather than
include sermon in the class of sleeping pill and atom in the class of solar
systems? The classical answer has been that the comprehender recognizes
that the literal interpretation of the sentence is impossible - sermons have
158 Models of comprehension

properties that do not allow them to be included in the class of small

white pills. Once the incoherence of the metaphorical sentence has been
recognized, the comprehender attempts to find a metaphorical compari¬
son. Early theories of metaphor in linguistics, philosophy, and psy¬
chology (including Kintsch, 1974) were all of this kind. Psychological
research results quickly showed that these theories were wrong. If com-
prehenders first interpret a metaphorical statement literally, recognize its
incoherence, and then get to work on a nonliteral interpretation, the
metaphor comprehension must involve more processing, require more
resources, and take more time than literal comprehension. A number of
studies have shown convincingly that this is not the case (for a review see
Cacciari & Glucksberg, 1994). Metaphors are as easy to comprehend as
literal sentences. Sometimes a literal interpretation is easier to arrive at,
but just as often the metaphorical reading is more readily attained.
It is probably fair to say that the view is commonly accepted today
that there exist no essential processing differences between metaphors
and literal sentences. As long as the semantic interpretation of a sentence
is conceived as a process of looking up the appropriate word meanings in
a mental lexicon and then computing the total meaning from the con¬
catenation of the word meanings, metaphors have to be regarded as
something abnormal. If, however, there is not that much to be looked up,
and most of the meaning construction occurs in context and with the
particular material at hand, there is no reason literal constructions
should always be the default and metaphorical interpretations be the last
resort.
The proof of this claim requires actually showing that the process of
meaning construction proceeds along much the same line for metaphors
as for literal sentences. Three simple simulations show that this can be the
case for the Cl model. The examples chosen are simple indeed and there¬
fore possibly misleading (not the general theory; just that these examples
are so trivial), but they are easy to understand and serve to illustrate the
general principle involved.
First, consider the metaphor The theory is a laser beam, which predi¬
cates something about laser beams to theory. But what is being predicated?
One way to find out is to construct a simulation with some degree of
objectivity by using LSA as a guide (see section 3.3). Latent semantic
analysis can indicate what some of the strong neighboring nodes are of
 Word identification in discourse 159

Table 5.4. Four neighbors 0/ theory (top four items) and four neighbors of
laser beam (bottom four items) and their cosines in the LSA space

Cosine with Theory Cosine with Laser Beam

relativity .64 .14

bang .64 .04
propose .60 .02
explain .60 .08

light .09 .68

coherent .06 .65
illuminates .01 .44
O
m

polarizer .60
f

both theory and laser beam, and it tells us how strong these links are. With
this information we can construct a Cl model network, integrate it, and
obtain some predictions as to the meaning of the metaphorical statement.
Specifically, the four nearest neighbors of theory in the encyclopedia-
based LSA space (excluding proper names, morphologically related
words, and function words) are the first four nodes shown in the first col¬
umn of Table 5.4 together with their link strengths (the cosine between
the respective vectors in the LSA space). The table also shows four
neighbors of laser beam - the last four items in column 1 - and the cosines
between all these vectors.
A network consisting of the text nodes THEORY, LASER-BEAM,
and ISA[THEORY,LASERBEAM] and the eight knowledge elabora¬
tions in Table 5.4, with all their pairwise cosine values as link strengths,
yields the activation values shown in Figure 5.14 when integrated.
Although I have no data to evaluate these predictions, they seem intu¬
itively acceptable. The dominant features of theory in isolation are deem-
phasized, and several features derived from laser beam become relatively
strong: coherent, light, and illuminates. However, note that this feature
transfer is quite crude. Features like polarizer, which seem intuitively
irrelevant, also are introduced at a fairly high level of strength.
As another example, consider the famous opening line of Carl Sand¬
burg’s poem: The fog comes on little cat feet. Again, I selected four neigh-
160 Models of comprehension

Activation
■

Figure 5.14 Activation strength of nodes related to theory in isolation and to

the metaphorical statement The theory is a laser beam.
Activation

Figure 5.15 Activation strength of nodes related to The fog comes in isolation
and to the metaphorical statement The fog comes on little cat feet.
 Word identification in discourse 161

bors of The fog comes from the LSA space as well as four neighbors of on
little cat feet. The text propositions

(6) PI COME[FOG]
P2 LITTLE[CAT[FEET]]
P3 ON[Pl,P2]

are integrated together with the eight semantic neighbors that have been
retrieved by LSA, with the cosine values from the LSA serving as link
strengths. Figure 5.15 shows the results. Dominant features of The fog
comes, such as drizzle, become less central to the meaning of the whole
sentence, and some new features are introduced, such as stalk and kitten.
Note that The fog comes already was somewhat soft, and although this node
is even more strongly related to little cat feet, its absolute activation did
not increase much, but its relative weight in the pattern that constitutes
the meaning of text (6) certainly has increased.
Note that if I had used the sentence The theory is abstract, the process
of meaning construction would have been essentially the same, though
quite nonmetaphorical. Some of the things we associate with theory
would be deemphasized, whereas some features of abstract would trans¬
fer, creating a new concept.
My next example makes this equivalence between metaphorical and
literal processes explicit. Consider The old rock had become brittle with age,
uttered about a former professor and the wall of a medieval monastery,
respectively (adapted from Gibbs, 1994):

(7a) John was in for a surprise when he visited his former professor
after many years. The old rock had become brittle with age.
(7b) John was in for a surprise when he touched the wall of the
monastery tower. The old rock had become brittle with age.

The inference ISA[FORMER[PROFESSOR], OLD[ROCK]] is re¬

quired in (7a) and PART-OF[MONASTER\ [WALL], OLD[ROCK]] is
required to understand (7b). Both are presumably made on formal
grounds: The definite determiner the signals a prior referent, and the wall
and the professor are the nearest candidates. Sentence (7a) is metaphoric
and (7b) is literal. Yet the process of meaning construction is the same for
both. To construct an illustrative example, I have assumed that each noun
phrase retrieves from long-term memory four of its most strongly asso¬
ciated neighbors. To reduce the arbitrariness of the example, I used LSA
162 Models of comprehension

Table 5.5. Four strongly associated neighbors of key

noun phrases in examples (7a) and (7b)

former professor: university

emeritus
faculty
*Nicolaas Bloembergen
wall of monastery tower: vaulted
buttresses
masonry
^Westminster Abbey
old rock: old
monolith
volcanic
*Dome of the Rock

to determine what these neighbors are and how strongly they are associ¬
ated to the noun phrases in the (7a) and (7b) sentences. Thus, I assumed
that the items shown in Table 5.5 will be retrieved; in each case I selected
three word concepts, that is, words that have a high cosine between their
vector and the eliciting noun phrase in the LSA space, and one document
(the starred item) - the LSA analog of a memory episode.
Latent semantic analysis knows the world only through reading docu¬
ments, whereas people have real experiences, episodes they remember and
use to understand and interpret language. Nicolaas Bloembergen, I sup¬
pose, must be a famous old professor to whom an encyclopedia article was
dedicated, and when LSA read the phrase old professor it was reminded of
that article, much as a human might be reminded of some old professor or
of a particular encounter with a member of that curious species.
If we construct a network linking the two noun phrases in (7a) and (7b)
in the same way as in Figure 5.14 and integrate these networks, the acti¬
vation patterns shown in Figures 5.16 and 5.17 are obtained. Whether
metaphoric or not, the results are much the same: Features of old rock are
transferred to wall as well as to professor, and some of the original features
of these terms are muted, though by no means effaced. The wall reminds
us now more of the Dome of the Rock than Westminster Abbey and becomes
old and more like a monolith; it is less likely to be composed of masonry.
The professor, too, ages and assumes some features of the Dome of the Rock
 Word identification in discourse 163

Wall

—O. Old rock

Figure 5.16 The activation of eight nodes linked to wall of monastery tower
alone and linked to the wall of the monastery tower is an old rock. The first four
items have been retrieved from the neighborhood of wall of monastery tower,
and the last four items have been retrieved from the neighborhood of old rock.

and a monolith; his university associations become less prominent. Previ¬

ously characterized only quite blandly by his role, this simple metaphor
creates a better specified individual characterization for the professor. Pre¬
sumably this is the reason people use figurative language. It allows them
to say things compactly and effectively that would be cumbersome and
perhaps impossible to say otherwise. But the process involved is no dif¬
ferent from the nonmetaphoric example.
If there is no processing difference between metaphoric and non¬
metaphoric language use, how can people tell what is a metaphor and
what is not?4 Judgments of metaphoricity might be a matter of semantic
distance: If two quite unrelated noun phrases are put together in a sen¬
tence, we call it nonliteral. However, the semantic relatedness between
the noun phrases in (7a) and (7b), as measured by the cosine between

4 Actually, there may be less agreement than linguists think concerning what a
metaphor is. I often have trouble convincing my class that “the stock market crashed"
is a metaphor.
164 Models of comprehension

Professor

-o. Old rock

Figure 5.17 The activation of eight nodes linked to old professor alone and
linked to the old pro fessor is an old rock. The first four items have been
retrieved from the neighborhood of old professor; and the last four items have
been retrieved from the neighborhood of old rock.

their vector in the LSA space is .07 and .04, respectively. Old rock is just
as unrelated to wall of the monastery tower as to former professor. Perhaps
judging whether a sentence or phrase is or is not a metaphor involves
more analysis and is not directly based on semantic distance. Perhaps the
aesthetic pleasure one derives from reading figurative or poetic language
is also based on such a postcomprehension analysis. For instance, a reader
might detect a category violation in the case of the simple IS-A meta¬
phors we have discussed here: We say (7a) is a metaphor because we know
that rocks and professors belong to different conceptual categories, and
we take pleasure in understanding the sentence nevertheless.
The illustrations of metaphor comprehension described here show
that it is feasible to use the Cl model together with semantic relatedness
data obtained from LSA to simulate metaphor comprehension. However,
many aspects of metaphor comprehension and usage have not been
addressed here (such as asymmetries in metaphorical comparisons). To
demonstrate the general validity of the points made here will require a
 Word identification in discourse 165

great deal of further work with representative examples of sentences and

corresponding human judgment data.

The construction of meaning does not stop with the construc¬

tion of word meanings, but it must start there. Constructing the
meaning of a word is a highly contextualized process, so much
.so that it becomes misleading to talk about constructing word
meanings. The process might be better described as one of con¬
structing the meaning of phrases, sentences, or even larger text
units of which the word is a part. The meaning of the word,
then, is intertwined with the meaning of these larger units and
indeed difficult to separate from them. “A rose is a rose is a rose”
is false - it is a somewhat different rose in every new context.
Because the contexts in which the rose will appear are not arbi¬
trary but are interrelated, not all these roses will be totally dif¬
ferent but will share overlapping sets of features. But these
shared features are not a common core meaning, the essence of
the rose, but are much more widely variable, depending upon
the use that is made of the word.
The Cl model describes this process of meaning construction
as starting with a bottom-up activation of a vague and nonspe¬
cific potential meaning and the gradual formation of a specific
meaning in the process of integrating the word into larger sen¬
tence and text units. In the case of a content word encountered
for the first time in a discourse context, information linked to the
word in long-term memory, semantic as well as personal-
episodic, is instantiated in working memory and participates in
the integration process. The result of the integration process is
a coherent structure into which the word meaning has become
embedded. Anaphora of all kinds — specific repeated words as
well as pronouns — are treated in much the same way, except that
prior information in the text plays a stronger role. Various poten¬
tial interpretations are interpreted in parallel, with the one that
fits best into the discourse context winning out in the integration
process.
The distinction between literal and nonliteral language is not
well defined in such a framework. The process of meaning con¬
struction is one of constraint satisfaction and contextual elabora-
Models of comprehension

tion that is basically the same whether a word is used literally or

not. It is possible to simulate how literal as well as nonliteral word
meanings are constructed in a discourse context at least for some
simple cases. The empirical data on word identification, anaphora
resolution, and metaphor interpretation are in good agreement
with these simulations, providing a reasonably firm, though still
incomplete, basis for the further exploration of meaning con¬
struction in discourse, beyond the word level.
 6

Textbases and situation models

A parser in the Cl model should take text as its input and generate as its
output a network of propositions that then could be used as a foundation
for the further modeling of comprehension processes. We do not have
such a parser and must do with hand coding as an unsatisfactory (but
workable) substitute. Nevertheless, there are at least the beginnings of
some research on parsing processes within the Cl framework, which I
discuss in the first section of this chapter. The main focus of the Cl model
has always been on how textbases and situation models are put together,
once the elementary propositions that are their building blocks have been
constructed. Thus, I review recent research on the formation of mac¬
rostructures and the role they play in comprehension. This leads to
the general topic of inferences, which is discussed in some detail here,
in part because there has been a great deal of controversy in this field
in recent years and in part because of their acknowledged central role in
comprehension. Inferences (though that term, I argue, is misleading) are
involved in both the formation of the textbase and the construction of the
situation model. In particular, I discuss the construction of spatial situa¬
tion models because of the special challenges this topic provides for a
propositional theory.
Although most research on text comprehension has been done with
narrative or descriptive texts, the principles of comprehension are very
general. How they might be applied to literary texts is explored in section
6.5. A good argument can be made that the comprehension of literary
texts is not different at the level of the basic architecture from the com¬
prehension of the kinds of texts we typically study in our laboratories,
although it demands special strategies and knowledge. However, all text
genres require domain-specific strategies and knowledge.
 168 Models of comprehension

6.1 Parsing

When the Cl model was first presented in Kintsch (1988), I suggested

that it might provide a plausible account of parsing processes. Relatively
general and robust parsing rules might be used, and much of the burden
of arriving at an unambiguous parse could be shifted to the contextual
integration process. The only researcher who has followed up these ideas
is Ferstl (1994a, 1994b). I therefore describe her work in some detail here.
The basic idea is simple. Instead of sophisticated, context-sensitive
parsing operators, simple operators are employed that use little informa¬
tion but are easy to apply. Similar to the process of establishing word
meanings, the parsing operators generate a network full of contradic¬
tions, for whenever there are choices, they yield both alternatives. How¬
ever, these alternatives inhibit each other, and during the integration
process the right alternative should win out because it fits better into the
given context. I illustrated this model of parsing with the garden-path
sentence

(1) The linguists knew the solution of the problem would not be
easy.

T he parser would create a network containing two mutually inhibitory

nodes, as shown in Figure 6.1. Following standard parsing heuris¬
tics (“minimal attachment”), it creates the proposition KNOW
[LINGUISTS,SOLUTION], but it cannot complete the parse. The
WOULD-NOT-BE-EASY must be attached to SOLUTION, but that
contradicts the proposition just created. The alternative is to treat
SOLU FION not as the object of KNOW but as the beginning of a new
phrase. This way, the NOT-EASY can be readily incorporated. The
resulting structure, of course, contains contradictions that must be
resolved by the integration process. The final activation values displayed
in Figure 6.1 show that the integration process does not do a clean job,
assuming that all links have strength either 1 or -1. Although the correct
alternative is more activated than the incorrect one, the troublesome
NOT-EASY??? still retains some activation. But perhaps this account is
not so fai off and actually describes the mental representation of a gar-
den-path sentence.
1 he model can generate a more unambiguous interpretation of exam¬
ple (1) if we allow changes in the link strength. Suppose a comprehender
is not satisfied with the result of the integration process in Figure 6.1 and
 Textbases and situation models 169

LINGUISTS 0.3848
KNOW [LING,SOLUTION] 0.3247
KNOW [LING ,$] 0.4137
SOLUTION 1.0000
NOT-EASY[SOLUTION] 0.7344
NOT-EASY???
OF-PROBLEM
0.3496
0.5189
SK
LINGUISTS

Figure 6.1 The network constructed to process the linguist-sentence (1).

All links are either 1 (solid lines) or -1 (broken lines). Final activation values
after the integration phase are displayed on the left.

decides that ungrammatical nodes like NOT-EASY??? are unacceptable

and that the constructions responsible for their creation should be more
strongly inhibited than was the case in Figure 6.1. In Figure 6.2, the same
net is shown, except that the link between NOT-EASY??? and KNOW
[LINGUISTS,SOLUTION] has been given a value of-2. In this case,
the offending proposition is now completely deactivated, but the NO T-
EASY node itself retains considerable activation as a sort of memory of
past trouble.1
How could the model make such a decision to increase an inhibitory
link? One way would be to calculate a statistic that indicates how well the
results of integration satisfy the initial constraints in the network. Such a
statistic is the harmony statistic of Smolensky (1986; see also Britton &
Eisenhart, 1993). Harmony is a function of the products of the initial and
final link and node strengths in the net. These products are large when
the initial and final strength value agree, that is, when the final solution
respects the initial constraints. In the case of figure 6.1, the harmony is
quite low, with a value of .18. It is conceivable that a harmony monitor
would reject such a solution and make changes in the network designed

1 It remains to be seen whether this is an empirically correct prediction; according to

Gernsbacher (1993), rejected nodes are suppressed, that is, their activation is nega¬
tive or below baseline!
170 Models of comprehension

LINGUISTS 0.3302
KNOW [LING .SOLUTION] 0.0000
KNOW[LING,$] 0.6249
SOLUTION 1.0000
NOT-EASY[SOLUTION] 0.8549 _
NOT-EASY??? 0.5246 AA
OF-PROBLEM 0.5248 LINGUISTS

Figure 6.2 The same network as in Figure 6.1, except that the broken link
in boldface has a strength of —2. Final activation values after the integration
phase are displayed on the left.

to provide a better solution with a higher harmony value. Indeed, the har¬
mony for the network shown in Figure 6.2 increases to a value of .30.
Ferstl (1994a) discusses a parsing problem that parallels (1). She is
concerned with another well-known garden-path sentence:

(2) The horse raced past the barn fell.

Again, a network is constructed representing RACE[F!ORSE] and

PAST-BARN, but the FALL??? node cannot be syntactically connected
to this network, as shown in Figure 6.3. The alternative RACE[SOME-
ONE,FIORSE] and FALL[HORSE], on the other hand, provides no
problem. The integration process suppresses the incorrect constructions
entirely, as shown in Figure 6.3. The difficulty in this case appears to be
in constructing the correct alternative in the first place, for informal
observations suggest that many undergraduate students have trouble with
this construction.
What is interesting about these examples is that the interpretation pro¬
ceeds using only syntactic cues. This is not generally the case, for seman¬
tic and pragmatic cues usually play a role as well. Consider the following
sentence pair, from an experiment byTaraban and McClelland (1988):

(3) The janitor cleaned the room with the broom.

(4) The janitor cleaned the room with the window.
 Textbases and situation models 171

fall?? 0.0000 fall??

RACE [HORSE] 0.0000
LOC[R[H], PAST-BARN] 0.0000
HORSE 0.8236
BARN 0.2567
FALL [HORSE] 0.8236
RACE [SOMEONE .HORSE] 1 .0000
LOC[R[S,H],PAST-BARN] 0.5677

10C[ R[S.H] .PAST- BARN]

FALL! HORSE]

Figure 6.3 The network constructed to process the sentence The horse raced
past the ham fell. Final activation values after the integration phase are shown
on the left. Positive links are indicated by solid lines, negative links by broken
lines. After Ferstl (1994a).

These sentences have the propositional representations CLEANjJANI-

TOR,ROOM,BROOM] and CLEANQANITOR, ROOM] and WITH
[ROOM,WINDOW], respectively. Out of context, subjects prefer the
“broom” sentence over the “window” sentence, but they correctly parse
both sentences (Taraban & McClelland, 1988). Ferstl (1994a) shows that
the Cl model can find the correct interpretation of these sentences with¬
out explicit representation of the verb’s thematic roles or selection
restrictions. All it needs are a few associative knowledge elaborations. The
network the model constructs for (3), according to Ferstl (1994a) is
shown in Figure 6.4.

WINDOWS 0.0731
ROOM 0.3877
CLEAN [JAN .ROOM .BROOM] 1.0000
WITH [ROOM .BROOM] 0.0000
CLEAN [JAN .ROOM] 0.0000
JANITOR 0.3729
CLEANING 0.4589
BROOM 0.4590

Figure 6.4 The network constructed to process the sentence The janitor
cleaned the room with the broom. Final activation values after the integiation
phase are shown on the left. Positive links are indicated by solid lines, negative
links by broken lines. After F stl (1994a).
 172 Models of comprehension

To represent the verb attachment bias of subjects in Figure 6.4, the self
activation of the preferred CLEANQANITOR,ROOM,BROOM] node
was increased, as was the inhibitory link between the two alternatives.
ROOM was associated with WINDOW, and CLEANING was associated
with BROOM, and JANITOR (association strengths were assumed to be
.5). As a result of these minimal knowledge elaborations, the model cor¬
rectly computes the verb attachment for BROOM in (3) and (a network
analogous to Figure 6~4) the noun attachment for WINDOW in (4).
A final example from Ferstl (1994a) illustrates how the Cl model takes
account of the discourse context in parsing a sentence. It has already been
mentioned that the noun attachment required by (4) is less preferred than
the verb attachment. This bias can be overcome by an appropriate dis¬
course context. For instance, if the sentence

(5) There was a room with plants and a room with windows.

precedes (4), the nonpreferred noun-attachment becomes easier. In con¬

trast, if the sentence

(6) There was a lounge with plants and a room with windows.

precedes (4), this effect is not obtained. Altmann and Steedman (1988)
have called this the principle of referential support: Example (5) has the
effect of introducing a room with windows into the discourse, to distin¬
guish it from the room with plants, whereas the window modifier is redun¬
dant in (6).
Ferstl (1994a) simulated this effect by first integrating (5) and carrying
over in the short-term memory buffer the two most highly activated
ptopositions lor the second processing cycle, which involved the integra¬
tion of The janitor cleaned the room with the window. The propositions car¬
ried over were WITH[ROOMbPLANT] and WITH[ROOM2,WIN¬
DOW], Similarly, (6) adds WITH[LOUNGE, PLANT] to the network.
In either case, whether (5) or (6) preceded it, the noun attachment was
successfully made (i.e., the proposition with the instrument interpreta¬
tion was correctly deactivated). However, the time course of integration
was quite different, depending upon whether the discourse context was
(5) or (6). Ferstl’s results (1994a) are shown in Figure 6.5. In the context
when only one room was mentioned, the preferred (but incorrect) instru¬
ment interpretation for window was slightly more activated for the first 14
iterations and did not become fully deactivated until after 27 iterations.
In contrast, when two rooms were mentioned so that a discourse object
 Textbases and situation models 173

CLEA[JANTrOR,ROOM,BROOM]

O. CLEAN[JANITOR,ROOM,WINDOWS]

Figure 6.5 The time course of activation for the two propositions
CLEANQANITOR,ROOM,BROOM] and
CLEAN[JANITOR,ROOM,WINDOWS]. After Ferstl (1994a).

room-with-windows had already been established, the instrument propo¬

sition weakens much earlier and is fully deactivated alter 20 iterations.
Similarly, the correct room-mth-mndow proposition increases in strength
more rapidly when it receives referential support than when it does not.
Ferstl (1994a, 1994b) also discusses other experiments in the literature
in which results were obtained that were seemingly contradictory to those
of Altmann and Steedman (1988), such as an experiment by Ferreira and
Clifton (1986). She points out that the negative results of Ferreira and
Clifton could be expected according to the Cl model, because the mate¬
rials used in their experiment were written in such a way that the lefei-
ential information was no longer in the discourse focus when the target
sentence was read. Hence, the WITH[ROOM2,WINDOW] proposition,
although it had been constructed earlier, was not carried over in the
buffer and did not participate in the processing of the target sentence,
and therefore could not have had an effect.
Ferstl’s simulations and her discussion of the literature are interesting
and provoking. Looking at the parsing literature through the eyes of the
Cl model can be instructive. However, a really systematic treatment of
174 Models of comprehension

parsing within the Cl framework is yet to come.2 Such a task may need to
wait until we have acquired a more extensive database on the parsing
strategies people actually use than is available now. Richer empirical stud¬
ies of parsing strategies, data as in Ferstl (1994b), would substantially
help our progress theoretically.

6.2 Macrostructure formation

According to the model of text processing of Kintsch and van Dijk (1978)
and van Dijk and Kintsch (1983), the formation of a macrostructure is an
integral part of normal text comprehension. It does not occur merely in
response to special task demands, such as instructions to summarize the
text, but is an automatic component of the process of comprehension that
cannot be separated from it.
That readers are able to recognize topic sentences that are expressed in
a text has been shown many times with a variety of procedures, including
reading times, think aloud protocols, and importance ratings (e.g., by
Kieras, 1980). Similarly, it has been shown that readers can produce ade¬
quate summaries of simple texts on demand (e.g., Kintsch & Kozminsky,
1977). However, there also exist good experimental data to support the
stronger claim of the theory that macrostructure formation occurs as an
integral part of comprehension.
If subjects read a text as part of a word recognition experiment, there
is no reason to think that they would intentionally engage in macropro¬
cessing if such processing were an optional, strategic part of comprehen¬
sion. Thus, if one can show that readers under these conditions form
macrostructures anyway, this would provide evidence for the automatic
nature of macroprocesses. Guindon and Kintsch (1984) produced such
evidence in a study that relied on the recognition priming method of Rat¬
cliff and McKoon (1978).
Ratcliff and McKoon had shown that recognition priming can provide
a very sensitive and at the same time nonintrusive measure for the analy¬
sis of the memory structures generated during reading. Their method
was exceedingly simple. They gave subjects sentences to read, followed by
a word-recognition test. They looked at the speed with which words from

- I he present approach has some similarities to the probabilistic parsers that have
recently been developed within computational linguistics (Charniak, 1993; Jurafsky
1196), but the compatibility of these approaches needs to be investigated further.
 Textbases and situation models 175

the sentences were recognized as a function of the word that preceded the
test word on the recognition test: whether or not the preceding word was
from the same sentence as the target word, and whether or not it was from
the same proposition. In the first case, they observed a priming effect of
110 ms over different-sentences controls. When the preceding word and
the target word not only came from the same sentence but also from the
same proposition, an additional small, but statistically significant priming
effect of 20 ms was obtained.
Guindon and Kintsch (1984) used this procedure to study macropro¬
cessing in comprehension. In experiment 1, their subjects read para¬
graphs with an initial topic sentence. For instance, a text describing the
training of decathloners might start with the topic sentence A decathloner
develops a well-rounded athletic body and contain as one of the sen¬
tences in the body of the paragraph the statement A decathloner also
builds up strong hands. In what we called macropairs, both the preced¬
ing word and the target word came from a topic sentence (e.g., develop,
body); in micropairs, both words came from a detail sentence (e.g., build,
hand)\ in control pairs, words that appeared in the text but in different
sentences succeeded each other. Hence, the correct response to all target
words was “yes.” The priming effect was computed as the difference
between the recognition time for control target words minus the recogni¬
tion time for macro- and microtarget words. Figure 6.6 (left panel) shows
that the priming effect was substantially larger for macro words than for
micro words.3
Similar results were obtained in experiment 2 of Guindon and Kintsch
(1984), except that in this experiment the paragraphs were shown with¬
out the topic sentences, so that macropropositions had to be inferred
rather than constructed on the basis of an actual topic sentence. Control
word pairs were used (words that had not appeared in the paragraphs), as
well as word pairs thematically related to the paragraph but that did not
function as macrowords. Thus, we compare the recognition times for
body as a macroword and feet as a thematically related nonmacroword with
a thematically unrelated control word. Because the subject had not seen
any of these words before, the correct response in all cases was uno” in
this experiment. The results are shown in the right panel of Figure 6.6: A

3 In a recognition experiment reported by Albrecht and O’Brien (1991), similar results

were obtained: macrorelevant, central concepts were more rapidly recognized than
peripheral concepts.
 176 Models of comprehension

Exp. 1 Exp. 2
Target Word

Figure 6.6 Priming effects in word recognition for macro words and non¬
macro words. After Guindon and Kintsch (1984).

stronger (negative) priming effect was obtained for macrowords than for
thematically related nonmacrowords.
figure 6.6 provides evidence that readers react to macro words differ¬
ently than to nonmacrowords, even though this was only a word recogni¬
tion task. Readers did not even have to understand what they read. How¬
ever, just as they could not help understanding the texts they read, they
could not help generating a macrostructure for these texts. Macrostruc-
tuie foi mation indeed appears to be an automatic and integral part of
reading comprehension.
A stud\ by Lorch, Lorch, and Mathews (1985) adds further support to
this conclusion by means ol a different experimental methodology. Lorch
and his colleagues constructed texts that were hierarchically organized
into major and minor subcategories. They compared reading times for
topic sentences after a major and a minor topic shift and found that after
a major shift topic sentences required 300 ms more reading time than
after a minor shift. However, there were no shift effects on sentence read¬
ing times when the sentences were presented in a disorganized fashion,
thus preventing readers from generating an orderly macrostructure. It is
interesting that these effects occurred when subjects were reading with¬
out special instructions. When the subjects were informed that they
would have to outline the text, these effects became even larger. Thus,
 Textbases and situation models 177

macroprocessing occurs during normal reading but can be enhanced

strategically when there are appropriate task demands.
The generation of macropropositions can be considered as some kind
of inference - an inference that does not add information to the text but
that reduces information (Kintsch, 1993). In selecting a macroproposi¬
tion, micropropositions are deleted, and in forming a generalization or
construction, several micropropositions are replaced by an appropriate
macroproposition. Information is reduced in all these cases, as a summary
replaces lower-level detail. In the studies reviewed here, this reduction
process was largely automatic because they dealt with familiar domains for
which the appropriate generalizations and constructions are readily avail¬
able to the comprehenders. For instance, if we are told that

(7) John bought a ticket, went to the airport in Denver, got on a

plane, ordered some drinks and dozed off, and finally got off the
plane in Chicago,

the construction of the macroproposition John flew to Chicago is fully

automatic because the information we were given is connected in a stable
retrieval structure (see chapter 7 for more detail) to flying in long-term
memory. Hence, John flew to Chicago does not have to be inferred by some
special inference procedure but becomes available automatically in long¬
term working memory. It is in these situations, when such retrieval struc¬
tures are available in long-term memory, that macropropositions become
automatically available during comprehension. As a consequence, they do
not have to be formed at all; the process of textbase construction simply
makes them potentially available. However, in unfamiliar domains, where
the comprehender lacks retrieval structures, macrostructure formation
for texts cannot be automatic and is not an inherent component of com¬
prehension. In fact, it may not take place at all, or if a macrostructure is
formed, it requires controlled processing - explicit reasoning procedures
as well as memory search for potentially relevant information. This dis¬
tinction between automatic and controlled processes in inferencing is dis¬
cussed more fully in section 6.3.

6.2.1 Signaling the macrostructure

Mross (1989), as part of a dissertation at the University of Colorado, used

an experimental design that allowed him to study another feature of
macroprocessing: the signals by means of which macropropositions are
178 Models of comprehension

linguistically marked in a text. He had subjects read texts that consisted

of three subtopics. Each of the three subtopics was introduced by an
explicit topic statement, as in the following example:

(8) One form of business organization is that of sole proprietorship.

There are no legal requirements for this form of organization.
(Marked text)

In the implicit version of the texts these topic statements were omitted.

(9) There are no legal requirements for the sole proprietorship form
of organization. (Unmarked text)

In his first experiment, Mross used an item recognition task to study the
speed and accuracy with which words from topic statements and words
from detail sentences were recognized. These recognition probes were
presented at two points during the reading of each paragraph. Macro¬
probes (sole proprietorship) were recognized more accurately (93% vs.
76%) than microprobes (legal requirements) and were responded to faster
than microprobes (487 ms vs. 524 ms, respectively), in agreement with
Guindon and Kintsch’s (1984) results.
In another experiment (experiment 3), Mross (1989) recorded reading
times for both topic sentences and detail sentences in the marked as well
as the unmarked texts. These results are shown in Figure 6.7. Topic sen¬
tences were read more slowly than detail sentences, replicating the results

Marked

Unmarked

Topic Detail

Figure 6. / Reading time per syllable in ms for topic and detail sentences in
the marked and unmarked condition. After Mross (1989).
 Textbases and situation models 179

of Lorch et al. (1985). In addition, we observe an interesting interaction

in Figure 6.7: topic sentences are read much more slowly when they are
not marked linguistically. Readers apparently need extra time to figure
out the topical status of these sentences when they are not given a lin¬
guistic cue.
Mross also asked his subjects to summarize the texts they read, and
he scored these summaries for completeness. It turned out that the sub¬
jects who read the marked texts wrote significantly better summaries
than the subjects who read the unmarked texts (5.6 subtopic statements
included in the summary for marked texts vs. 4.6 subtopic statements for
unmarked texts). Linguistic marking not only allowed readers to process
the texts faster but also helped them to write better summaries.
Topic marking thus appears to be a powerful means for facilitating
macroprocessing. Mross used summary sentences for this purpose, but
headings and previews work equally well for this purpose (e.g., Lorch,
Lorch, & Inman, 1993). Indeed, the language makes available a variety of
more subtle means to signal whether we are continuing with a topic or
shifting to a new one, and speakers and writers use these means liberally.
One technique used to signal a topic shift is to overspecify anaphora. We
have seen (section 5.2.1) that as long as a concept is available in working
memory, writers tend to refer to it by zero anaphora or pronoun. Full
noun phrases are typically used only when the concept must be reinstated
into working memory. Violating this convention signals a topic shift. If
one refers to a concept with a full noun phrase when a less specific refer¬
ence would have sufficed, a topic shift is signaled thereby. Linguists have
shown this to be the case in the texts they have analyzed (e.g., Fox, 1984;
Linde, 1979), and several experimental demonstrations of this effect also
exist. Fletcher (1984) gave subjects sentence pairs to read, such as the fol¬
lowing:

(10) la. Peter intended to go bowling last night.

lb. Peter intended to go bowling with Sam last night.
lc. Sam intended to go bowling with Peter last night.
2. Peter broke his leg.

and asked them to rewrite the second sentence to make it sound more
natural. His striking results are shown in Figure 6.8. For the highly
coherent (la) - (2) pair, subjects almost always chose either zero anaphora
or a pronoun. For the (lc) - (2) pair, which involves a clear topic shift, a
180 Models of comprehension

Low Coherence

O. Medium Coherence

----O— High Coherence

Figure 6.8 The choice of zero anaphora, pronouns, or noun phrases as a

function of the coherence between sentence pairs. After Fletcher (1984).

full noun phrase was the overwhelming choice, and intermediate results
were observed for the pair with medium coherence.
Fletcher’s result were corroborated by Vonk, Hustinx, and Simons
(1992), who had subjects provide verbal descriptions of picture stories
that did or did not involve a topic shift. Because there was only a single
actor in these picture stories, no linguistic ambiguity was involved. Nev¬
ertheless, topic shifts in the pictures were signaled linguistically by a full
noun phrase in 88% of all continuations, which contrasts with 83% pro¬
noun choices when no topic shift was involved.
The experiments reviewed here clearly show that macroprocesses are
real. They are as much part of discourse comprehension as micropro¬
cesses are. These experiments also allow us to glimpse how readers form
macrostructures. In particular, we note that natural language has avail¬
able rich means, including syntactic ones, to help the reader to generate
macrostructures. What the experiments discussed here do not make clear
is the enormous importance of macroprocesses for text understanding:
how easy a text is, how well readers understand it, how well they can
remember it, what they learn from it - all this is strongly dependent on a
successful macrostructure. Local comprehension problems may be a nui¬
sance, but problems at the macro level tend to be a disaster.
 Textbases and situation models 181

6.2.2 Dominance of the macrostructure: Contradictions in the text

A well-organized macrostructure is crucial for understanding and

remembering a text. Texts that are locally coherent but that contain global
inconsistencies are read more slowly and remembered less well (Albrecht,
O’Brien, Mason, & Myers, 1995; Albrecht & O’Brien, 1993). But too
much of a good thing can make readers misunderstand and misremem-
ber a text. An excessive emphasis on the macrostructure can lead to seri¬
ous distortions of the meaning of a text. Otero and Kintsch (1992) have
shown how certain pathological interpretations of texts containing an
outright contradiction can arise as a consequence of such an overempha¬
sis on the macrostructure of the text.
It is well known that readers often do not notice even direct contradic¬
tions in a text. Just how frequent such mistakes are depends on the read¬
ers and the task demands generated by the experimental situation, but the
basic phenomenon has been observed in a number of studies under a vari¬
ety of conditions. In the data of Otero and Kintsch (1992), for example,
40.3% of the contradictions in a text were not noticed by the reader,
either during reading or afterward. This is a striking result, because the
texts used in this experiment were brief and simple (though scientific
descriptions, rather than stories), and the contradictions were blatant, as
in the following example:

(11) Superconductivity. Superconductivity is the disappearance of

resistance to the flow of electric current. Until now it has been
obtained only by cooling certain materials to low temper¬
atures near absolute zero. That made its technical application
very difficult. Many laboratories now are trying to produce
superconducting alloys. Many materials with this property, with
immediate technical applicability, have recently been discovered.
Until now, superconductivity has been achieved by con¬
siderably increasing the temperature of certain materi¬
als.4

The two contradictory sentences (in boldface type) are separated by only
two sentences; all sentences in the paragraph are on the topic of super¬
conductivity; the contradiction is explicit (cooling vs. increasing the tem¬
perature). The readers were asked to report any comprehension problems

4 The actual texts used in the experiment were written in Spanish.

182 Models of comprehension

while reading and later recalled the paragraph. They were not warned of
contradictions in the texts (four of the six experimental texts contained a
contradiction). Most of the failures (82%) to notice a contradiction in a
text were of three types:

1. Subjects reproduced the definition of superconductivity but did not

mention in their recall protocols either of the contradictory sentences.
2. Subjects mentioned that superconductivity is achieved by cooling but
said nothing about the contradictory sentence.
3. Subjects mentioned both cooling and heating but explained away the
contradiction by some unwarranted, fanciful inference, such as “Up to
now superconductivity was achieved by cooling, but now it can be
achieved by heating certain materials.”

Thus, we have a high incidence of errors, and errors that are highly sys¬
tematic rather than random - a phenomenon that cries for an explanation.
To simulate comprehension of the superconductivity paragraph, the
text (11) was represented as a list of propositions. Because subjects read a
series of texts that all had the same structure (definitional sentence plus
three elaborative sentences), it is highly likely that they would have
considered SUPERCONDUCTIVITY and IS[SUPERCONDUCT-
IVITY,DISAPPEARANCE] as macropropositions. As an additional
macroproposition, we assumed that subjects would use either OF [DIS-
APPAEARANCE,RESISTANCE] or OBTAIN [SUPERCODUCTIV-
1 IT [COOL [TEMPERATURE]], corresponding to the type 1 and type
2 readers. The network contains the two contradictory propositions
OBTAIN [SUPERCODUCTIVITY[COOL [TEMPERATURE]] and
OBTAIN[SUPERCONDUCTIVITY,INCREASE[TEMPERATURE,
OF MATERIAL]], which are linked by an inhibitory link of strength —1.
All other links have strength +1.
Integrating the propositinal network thus generated in cycles yields
the long-term memory strength values for the two contradictory propo¬
sitions shown in Figure 6.9 (“normal”). The strength values are unequal,
but both propositions have substantial strength, and one would expect
that the contradiction would be noted.
T.jpe 1 and type 2 errors can be created in the model by increasing
the weights of the macropropositions in the network. If we increase
the strength of the links between macropropositions from 1 to a value of
10, the model makes type 1 and type 2 errors, depending on which
macrostructure we assume. If OF[DISPPEARANCE, RESISTANCE] is
 Tex lb uses and situation models 183

COOL

INCREASE

Figure 6.9 Strength values for the two contradictory propositions for nor¬
mal readers and three types of nondetectors.

included in the macrostructure, the definitional component becomes

so strong that both contradictory propositions receive only negligible
activation. Hence, they most likely would be overlooked, as type 1 readers
overlooked them in the experiment. On the other hand, if OBTAIN
[SUPERCODUCTIVITY[COOL [TEMPERATURE]] becomes part
of the macrostructure, the first proposition becomes so strong that it
inhibits the contradictory proposition later in the text, which is what hap¬
pened with type 2 readers.
Some readers did not report a contradiction because they explained it
away by an unwarranted inference, usually by assuming that one way of
generating superconductivity was used in the past but that now there is a
second way. We simulate these type 3 readers as normal readers who detect
the contradiction but add another proposition, FROM-NOW-ON
[OBTAIN[SUPERCONDUCTIVITY[HOT[TEMPRETAURE]]]. As
Figure 6.9 shows, both contradictory propositions receive high strengths
in this case, but there is no contradiction to report any more because
UNTIL-NOW has been arbitrarily replaced by FROM-NOW-ON.
Note that these type 3 readers employ a strategy for resolving the con¬
tradiction that is applicable in many cases, because contradictions are
often resolved by noting a difference in time or place. The present text,
however, does not warrant the use of this strategy.
Otero and Kintsch (1992) do not claim that every failure to note a con-
184 Models of comprehension

tradiction in a text is the result of an overemphasis on macroprocesses.

However, under the conditions of their experiment, this is a reasonable
hypothesis, and, as has been shown, it explains their data in some detail.
No new model had to be constructed for this purpose. All we did was to
assign an extra weight to an important and integral component of normal
comprehension processes. Comprehension of anything but the shortest
texts strongly depends on the successful construction of an adequate
macrostructure for the text. The macropropositions must be emphasized
in normal comprehension and are recalled better than the microprop¬
ositions of a text. However, if that emphasis is exaggerated, the model
becomes pathological and starts missing information in the text in ways
that parallel what real subjects do in certain experimental situations.

6.2.3 Macrostructures as vectors in the LSA space

How are macrostructures generated? According to the theory of van Dijk

(1980), macrostructures are derived from a text through the applica¬
tion of macrorules: Less important portions of the text may be deleted,
instances may be generalized, and summaries of events may be constructed.
A difficulty with this theory is that the macrorules are easy to describe
and readily illustrated by examples but cannot be automated. Given
a paragraph, it is not possible to say what its macrostructure, or gist, is.
We can suggest plausible macropropositions for it and show how these
are derivable via van Dijk’s macrorules, but these rules do not specify a
unique solution.
Latent semantic analysis may provide a useful alternative, in that it
permits us to represent macrostructures as vectors in the semantic space,
in the same way as words or sentences are represented. A macrostructure
in the LSA representation is thus the vector corresponding to a para¬
graph, and a higher-order macrostructure is simply the vector of several
paragraphs or the whole text. Thus, macrostructures in LSA are easily
computable and are unique. They are not, however, intuitively plausible
sentences oi phrases, such as those generated by the macrorules, but an
uninterpreted vector in a high-dimensional semantic space that we can
know only by its neighbors. For some purposes this may be quite suffi¬
cient, however, and in fact advantageous, because we now have the capa¬
bility to calculate the semantic distance of this macrovector with other
vectors of potential interest, as the following example illustrates.
 Textbases and situation models 185

Consider the following ministory, a text T, consisting of a setting (the

first paragraph, ^1) and an event (the second paragraph, ^2):

(12) T ^1 SI John was driving his new car on a lonely country

road.
52 The air was warm and full of the smell of spring
flowers.
53 He hit a hole in the road and a spring broke.
54 John lost control and the car hit a tree.

Latent semantic analysis lets us represent as vectors the content words in

SI to S4 as well as the four sentences themselves. Thus, we can compute
the semantic relatedness between each word in a sentence and the sen¬
tence as a whole - the cosine between the respective vectors. Figure 6.10
shows the results.5 What is interesting here is that some words have a big
effect on the sentence meaning (high cosine between word and sentence
vector, e.g., car in SI), whereas others have very little effect (word mean¬
ing and sentence meaning are only weakly related, e.g., lonely in SI). Do
ratings by readers of the contribution that each word makes to a sentence
reflect these LSA relations? Furthermore, one can also compute a cosine
between a word in a sentence and the other words in the sentence, indi¬
cating how interrelated the words of a sentence are. If these values are
high, are sentences faster or easier to comprehend than when these val¬
ues are low?
Different sentences make different contributions to a paragraph, and
the two paragraphs make different contributions to the text as a whole.
Do readers judge the importance of sentences to a paragraph, or para¬
graphs to a text as a whole, in the same way as does LSA? The necessary
experiments to answer these questions are quite straightforward. If the
answers are positive, there appear to be numerous theoretical as well as
practical implications.
The paragraph representations 1)1 and in Figure 6.10 are uninter¬
preted vectors without an English-language counterpart. Using macro-

5 The fact that these results are based on a semantic space derived from reading an
encyclopedia must be borne in mind. An adult human semantic space is based on
much more experience with words, and with words from different kinds of texts —
encyclopedias focus on that part of human knowledge that is not common sense,
whereas common sense knowledge is what is most involved in understanding simple
texts like the ones analyzed here.
186 Models of comprehension

Figure 6.10 Cosines between words and sentences, sentences and para¬
graphs, and paragraphs and the whole text.

rules, one might generate the following summary sentences for Tjl, ^[2,
and the whole text, T, respectively:

(12) S (T[l) An automobile ride in the spring.

S (^[2) An automobile accident.
S (T) An automobile ride in the spring ends in an accident.

Obviously, however, there are other plausible alternatives. However, we

can easily test whether these suggested summary sentences or phrases
adequately express their respective macronodes. For instance, the two
summary sentences, S(^|l) and S(^|2), for the two paragraphs suggested
 Textbases and situation models 187

have cosines with the ^]1 and ^2 vectors of .39 and .42, respectively, and
the cosine between the summary sentence S(T) and the whole text vector
T is also .42. These are fairly high values (compare them to the cosines
between words and sentences in Figure 6.10), indicating that we have
constructed good summaries. This technique may prove useful for eval¬
uating summaries objectively and automatically.6
The cosine values between words and sentences, sentences and para¬
graphs, and paragraphs and the whole text shown in Figure 6.10 can be
used to implement a Cl-model simulation of the text together with its
macrostructure. We add to the atomic propositions of the text the com¬
plex propositions SI to S4 corresponding to the four sentences, and a
macrostructure consisting of Tfl and TJ2 for the two paragraphs, and T
for the text as a whole, and link them with the cosine values obtained
from the LSA. The resulting network is shown in Figure 6.11. Sentence
2 fills the time-circumstance slot of DRIVE; DRIVE provides the setting
for the events described by S3 and S4; the propositions of S3 and S4 are
all causally linked. The microstructure consists of the atomic text propo¬
sitions and the four complex propositions corresponding to sentences SI
to S4. The macrostructure consists of the two paragraph vectors ^1 and
T[2 (PI and P2), and the whole-text vector T. Links between atomic
propositions have a value of 1; all other links have values equal to the
strength of their relationship as assessed by LSA and shown in Figure
6.10. The numbers after each label are the long-term memory strength
values computed from the Cl simulation.
Comprehension of this network was simulated in four cycles, each
cycle comprising one sentence, including its superordinate macronodes.
For each cycle, the most strongly activated node from the previous cycle
was held over in the short-term memory buffer. The results, shown in
terms of long-term memory strengths in Figure 6.11, are interesting.
The strongest nodes are the DRIVE proposition from the first sentence
and sentence vectors S2, S3, and S4. On the other hand, the paragraph
representations ^1 and ^2 and the whole-text vector T end up with rela¬
tively low strengths.
Does including the macrostructure in this way make a difference for
the model? It certainly does, for if we simulate comprehension of this text
without the complex propositions and macronodes in Figure 6.11, the
strengths values obtained for the atomic text propositions correlate only

6 Some empirical results are reported in section 8.2.2.

188 Models of comprehension

Figure 6.11 A simulation of the comprehension of text (12).

r = .50 with the values obtained by including the macrostructure. Which

version is better can be determined only by a systematic comparison of
such model predictions with recall and summary data for a sufficiently
large number of texts. Furthermore, although it seems reasonable that
the link weights of the macronodes should be proportional to their
cosines in the LSA analysis, setting them equal to those values as was
done here is probably an oversimplification. Just how the micro- and
macrostructures should be weighted relative to each other remains an
issue to be explored.

6.3 Inferences

Text comprehension depends as much on the reader and the pragmatic

situation as on the text itself. What readers bring to the text, their goals
 Textbases and situation models 189

and prior experience, has been studied under the label of inferencing in
text comprehension. In particular, the process of knowledge use that
every reader must engage in to properly understand a text has been char¬
acterized as making inferences in text comprehension. This is a regret¬
table terminology that has caused a great deal of confusion, for much of
what we call inferencing has very little to do with real inferences.

6.3.1 Classification of inferences

A distinction should be made between problem-solving processes when

there are premises from which some conclusions are drawn (not neces¬
sarily by the rules of logic) that may be justly called inferences, and
knowledge retrieval processes in which a gap in the text is bridged by
some piece of preexisting knowledge that has been retrieved (Kintsch,
1993). Both inferences proper and knowledge retrieval may be either
automatic (and usually unconscious) or controlled (and usually conscious
and strategic). This classification results in Table 6.1.
The classification in Table 6.1 is based on a proposal by Guthke (1991)
and is described in detail in Kintsch (1993). Retrieval adds preexisting-
information to a text from long-term memory. Generation, in contrast,
produces new information by deriving it from information in the text by
some inference procedure. Thus, although the term inference is suitable
for information generation processes, it is a misnomer for retrieval
processes.
A prototypical example for cell A, the automatic retrieval process that

Table 6.1. A classification system for inferences in text comprehension. After

Kintsch (1993)

Retrieval Generation

A C

Automatic Processes Bridging inferences, Transitive inferences

associative elaborations in a familiar domain

B D

Controlled Processes Search for bridging Logical inferences

knowledge
190 Models of comprehension

enriches the information in a text, would be the activation of with a ham¬

mer by John nailed down a hoard, or cars have doors by A car stopped. The
door opened.1 In both cases sufficient retrieval cues for the information
retrieved exist in short-term memory and are linked with pertinent infor¬
mation in long-term memory. Under such conditions a long-term work¬
ing memory forms, which means that the linked information in long¬
term memory becomes readily available in working memory (see Ericsson
6 Kintsch, 1995, and the discussion in chapter 7 for details on the oper¬
ation of long-term working memory). In the terminology of Ericsson and
Kintsch (1995), a retrieval structure exists that links the cues in short¬
term working memory to particular contents of long-term memory,
thereby expanding the capacity of working memory. Such knowledge use
is automatic, rapid, and places no demands on cognitive resources. An
alternative model for this kind of knowledge retrieval is the resonance
theory of Myers (Myers, O’Brien, Albrecht, & Mason, 1994), a well-
developed mathematical model employing a resonance analogy for mem¬
ory retrieval. According to this model, cues in short-term memory pro¬
duce a resonance in long-term memory, so that the resonating items
become available for further processing in working memory. Thus, either
via retrieval structures or resonance, relevant, strongly related items in
long-term memory become potential parts of working memory, creating
a long-term working memory that is much richer than short-term work¬
ing memory, which is severely restricted in capacity. Indeed, it is only this
long-term working memory that makes discourse comprehension (or,
indeed, any other expert performance) possible. Smooth, efficient func¬
tioning would be impossible if we had no way of expanding the capacity
of working memory beyond the rigid limits of short-term memory. These
issues are discussed in greater detail in chapter 7; the point here is that
making long-term memory contents available via retrieval structures or
resonance should hardly be called “inference.”
In cell B of Table 6.1 are cases in which automatic retrieval is not pos¬
sible. That is, the cues present in short-term memory do not retrieve rel¬
evant information that bridge whatever gap exists in the text. An extended
search of memory is required to yield the needed information. A memory
search is a strategic, controlled, resource-demanding process in which

7 I am not claiming that this sort of inference does or does not occur “naturally” dur¬
ing comprehension. I am simply classifying inferences.
 Textbases and situation models 191

the cues available in short-term memory are used to retrieve other likely
cues from long-term memory that in turn are capable of retrieving what
is needed. Consider the following:

(13) Danny wanted a new bike. He worked as a waiter.

Purely automatic, associative elaboration might not retrieve the causal

chain from want-bike to buy-bike to money to work. However, a directed
search for causal connections between the two sentences would easily
generate these by no means obscure links. In all probability, genre-spe¬
cific strategies exist to guide such search processes. In a story, one would
look for causal links. In a legal argument, one routinely looks for contra¬
dictions. In an algebra word problem, algebra formulas guide the search.
The difficulty of such procedures and the resource demands they make
vary widely.
Retrieval processes merely access information available in long-term
memory, either automatically or by a resource-demanding search. Gen¬
eration processes actually compute new information on the basis of the
text and relevant background information in long-term memory. They,
too, may be either automatic or controlled.
On the one extreme, there are the fully automatic generation proce¬
dures (cell C of Table 6.1). For instance, given the sentence

(14) Three turtles rested on a floating log, and a fish swam beneath
them.

the statement The turtles are above the fish is immediately available to a
reader. Indeed, readers often are unable to distinguish whether they were
explicitly told this information or not (e.g., Bransford, Barclay, & Franks,
1972). Note, however, that this is not merely a question of knowledge
retrieval as in doors are parts oficars. The statement the turtles are above the
fish is not something that already exists in long-term memory and is
now retrieved, but it is generated during the comprehension process. The
reason it is so highly available in the reader’s working memory is pre¬
sumably that the fish-log-and-turtle scene is encoded as an image, and
this mental image constitutes a highly effective retrieval structure that
provides ready access to all its parts - not just the verbal expression used
in its construction.
Here is an instructive case in which the decision to represent mental
192 Models of comprehension

representations by propositions can easily get us into trouble. In a propo¬

sitional system, it would appear that some inference rule, for instance,
transitivity, would be needed to derive that the turtles are above the fish.
Instead, this information is given directly by the image that serves as the
situation-model representation of the sentence in question. Indeed, at
this level of representation there is no difference between explicit and
implicit statements. A difference exists only at the level of the textbase
and surface representation, which, however, may not always be effective
(as in the experiments of Bransford et ah, 1972, in which subjects could
not distinguish between explicit and implicit statements, given study and
test sentences as in the example discussed here). These facts do not argue
against propositional representations. Instead, they argue that proposi¬
tional representations for imagery must be constructed with great care, so
that the salient aspects of the image are indeed represented in the propo¬
sitional network.
However, what happens in cell C of Table 6.1 should hardly be called
an inference either. It is simply a case in which, because of the analog
nature of the mental representation involved, more information is gener¬
ated in forming a situation model than was explicit in the text. The term
inference really should be reserved for cell D of Table 6.1. This is the
domain of deductive reasoning and extends far beyond text comprehen¬
sion, though deductive reasoning undoubtedly plays an important role in
text comprehension, too. Explicit reasoning comes into play when com¬
prehension proper breaks down. When the network does not integrate
and the gaps in the text cannot be bridged any other way, then reasoning
is called for as the ultimate repair procedure.
Inferences (real inferences, as in cell D) require specific inference pro¬
cedures. It is a matter of considerable controversy in psychology what
these inference operations are — whether inference proceeds by rule
(Rips, 1994) or mental model (Johnson-Laird, Byrne, & Schaecken,
1992). The issue is beyond the scope of this book, but the stand taken in
chapter 2 on multiple levels of mental representation implies a definite
position in this controversy. Inferences in domains where the basic rep¬
resentation is an action or perceptual representation - that is, analog
rather than linguistic or abstract - must involve operations on mental
models. Inferences in truly symbolic, abstract domains must be by rule.
Inferences in the linguistic domain, where the representation is at the
narrative level, may be based on mental models (to the extent that lan-
 Textbases and situation models 193

guage is embodied, as discussed in chapter 2) but also could involve

purely verbal inference rules.8

6.3.2 Inference generation during discourse comprehension

The literature on inferences in discourse comprehension is for the most

part not concerned with cell D of Table 6.1. Indeed, it concentrates heav¬
ily on cell A, the processes that are the least like inference, according to
the argument presented here.9 A major focus of the recent research has
been on the question of to what extent inferences are made during nor¬
mal comprehension. On the one hand, it is clear that if the readers of a
story are asked to make inferences and are given sufficient time and
incentive, there is almost no limit to what they will produce (Graesser,
1981). On the other hand, there is good evidence that much of the time,
and in particular in many psychology experiments, readers are lazy and
get away with a minimum of work (e.g., Foertsch & Gernsbacher, 1994).
McKoon and Ratcliff (1992; 1995) have elaborated on the latter position
as the minimalist hypothesis, which holds that the only inferences read¬
ers normally make are bridging inferences required for the maintenance
of local coherence and knowledge elaboration in which there are strong
preexisting multiple associations. Many text researchers (e.g., Graesser
& Kreuz, 1993; Graesser, Singer, & Trabasso, 1994; Singer, Graesser,
& Trabasso, 1994), however, feel that this minimalist position underesti¬
mates the amount of inferencing that occurs during normal reading and
would at least add inferences that are necessary for global coherence to
the list (superordinate goal inferences, thematic inferences, and character
emotional reactions). Although this controversy has contributed a great
deal to our understanding of the role of inferences in text comprehen¬
sion, it has also shown that the question concerning which inferences are
necessary for, and are normally made during, text comprehension has no
simple answer. Text characteristics (much of the research is based on sto¬
ries, mostly ministories), task demands, and individual differences among
readers create a complex, though orderly picture.

8 Good experimental evidence exists on both sides of the rule versus model contro¬
versy; whether it can be aligned along levels of representation, as suggested here, is
another question.
9 Nevertheless, I shall drop the quotation marks from “inference” after this sentence
in deference to the commonly accepted terminology.
194 Models of comprehension

Trabasso and Suh (1993) have combined discourse analysis, talk-aloud

procedures, and experimental measures, such as recognition priming,
reading times, coherence ratings and story recall, to show that their read¬
ers did make causal inferences in reading a story and that these inferences
could be predicted by their analysis.
In an illuminating series of studies O’Brien and his co-workers have
shown that causal inferences in story understanding should best be
regarded as a passive operation that makes available background and
causal antecedents via a resonance-like mechanism (or what I would call
a retrieval structure). Such a process contributes to the coherence of the
text representation (Garrod, O’Brien, Morris, & Rayner, 1990) but is not
predictive. Readers refrain from prediction unless there is absolutely no
chance of being discomfirmed (O’Brien, Shank, Myers, & Rayner, 1988).
Global automatic goal inferences occur only under limited conditions
(Albrecht, O’Brien, Mason, & Myers, 1995), probably because such
inferences are as risky as predictions — they are frequently discomfirmed
as the later text reveals a different goal. When global goal inferences
occur, resonance describes what happens better than the notion of infer¬
ence does. Through resonance, related parts of a text are connected
because of preexisting retrieval structures. In contrast, the construction
of a full mental model with rich causal connections appears rather as a
nonautomatic, controlled process (Albrecht & O’Brien, 1995; O’Brien,
1995).
How much time and resources the reader has strongly determine the
amount of inferencing that occurs. Magliano, Baggett, Johnson, and
Graesser (1993), using a lexical decision task, found that causal antece¬
dent inferences were not made when texts were presented with an RSVP
procedure at a 250 ms rate but were made when the presentation rate was
400 ms. Long, Golding, and Graesser (1992) found that superordinate
goal inferences linking various episodes of a story (but not subordinated
goal inferences) were made by readers when they were given lots of time.
But with a rapid presentation rate, only good comprehenders made such
inferences, and there was no evidence for goal inferences by poor com¬
prehenders (Long & Golding, 1993).
Readers are much more likely to make antecedent causal inferences
than consequent causal inferences (e.g., Magliano et al., 1993). For in¬
stance, readers of The clouds gathered quickly, and it became ominously dark.
The downpour lasted only 10 minutes infer the causal antecedent the clouds
 Textbases and situation models 195

caused the rain. But given The clouds gathered quickly, and it became omi¬
nously dark, they do not infer the consequent the clouds caused ram. This
finding that antecedents, though not consequent causal inferences, are
made in text comprehension is readily accounted for by the Cl model.
Suppose a text describes a situation that is a common cause of some
event, and then asserts that this event occurred, without mentioning an
explicit causal connection between the antecedent and the event. Preex¬
isting retrieval structures causally link the antecedent and the event in the
reader’s memory; the causal link is activated and is likely to become a per¬
manent part of the reader’s episodic text memory because it connects two
highly activated nodes in the memory structure.
The situation is different for the consequent inferences. The same
retrieval structures that made available the causal antecedent will make
available the causal consequent, too. But at that point in the reading
process, the consequent is a dangling node in the episodic text structure
because it is connected to nothing else in the network but the antecedent.
Therefore, the consequent will not receive much activation in the inte¬
gration process and will be excluded from episodic memory. Thus, The
clouds gathered quickly, and it became ominously dark might make available
the clouds caused rain, but if nothing else in the text connects to rain, this
node will become quickly deactivated in the network. When in a later
processing cycle other information becomes available that could have
linked up with rain, that node is most likely lost from working memory.
Hence, although the retrieval structures in the reader’s long-term mem¬
ory make available both antecedent and consequent information, only the
former is likely to survive the integration process and become a stable
component of the reader’s text memory.
According to Table 6.1, it should make a great deal of difference
whether bridging inferences occur in a familiar domain or in an unfamil¬
iar domain. In the former case, preexisting retrieval structures make
available the information that fills the gap. The process of retrieving
information from long-term memory via a retrieval structure takes about
400 ms, as is shown in chapter 7. Hence, reading times for sentence pairs
for which a bridging inference is required should be increased by at most
400 ms in comparison with sentence pairs for which this information has
been explicitly stated. In fact, 400 ms is an upper limit, because the
retrieval from long-term working memory most likely occurs at least
partly in parallel with other ongoing reading processes. Noordman and
196 Models of comprehension

Vonk (1992) provide some data that allow us to test this prediction. In one
of their texts about a woman looking for an apartment to rent, the follow¬
ing sentence pair occurred:

(15) The room was large, but one was not allowed to make music in
the room.

In the explicit condition,J:his was preceded by

(16) The girl wanted to play music in her room.

Sentence (16) was missing in the implicit condition. Noordman and

Vonk’s (1992) results are shown in Table 6.2. Statistically reliable differ¬
ences are marked by an asterisk (*), nonsignificant differences are indi¬
cated by ns. Reading times for the hut phrase in (15) were slightly but sig¬
nificantly longer in the implicit condition than in the explicit condition,
as predicted by the retrieval structure hypothesis, but no differences in
verification times were found between conditions.
The results contrast with another experiment with a similar design in
which the texts were in an unfamiliar domain. The sentence pair requir¬
ing a bridging inference in this study is exemplified by

(17) Connors used Kevlar sails because he expected little wind.

1 he text is about a sailboat race, and the required inference, which is pro¬
vided in the explicit condition, is

(18) Kevlar sails are advantageous when the weather is calm.

Presumably, the subjects in this experiment did not know this fact about
Kevlar sails before reading this text. The results are also shown in Table
6.2. Reading times for the because phrase in (17) are equal in the explicit
and implicit conditions; that is, subjects wasted no time on figuring out
the obscure relationship between the Kevlar sails and the wind while
reading this sentence. On the other hand, when forced to make the infer¬
ence in a sentence verification task, they were considerably faster to ver-
lfy (18) in the explicit than in the implicit condition. Noordman andVonk
(1992) also report a version of this experiment in which the same texts on
economics were used in the familiar and unfamiliar condition, but the
readers were either novice students in economics or expert economists.
The results are analogous to those in Table 6.2.
The but and because in (15) and (17) play a crucial role in this experi¬
ment. Keenan and I have reported an analogous experiment with familiar
 Textbases and situation models 197

Table 6.2. Phrase reading times for the second phrase in (15)
(familiar text) and (17) (unfamiliar text) and inference verifi¬
cation times in ms for explicit and implicit conditions. After
Noordman and Vonk (1992)

Reading Times Verification Times

Familiar
Explicit 1,473 2,169
* ns

Implicit 1,536 2,198

Unfamiliar

Explicit 3,910 3,179

*
ns
Implicit 3,934 3,512

materials, such as those in examples (15) and (16) but without linguistic
signals like but and because (Kintsch & Keenan, 1973). Our results were
completely different: Reading times for explicit and implicit sentences
were not significantly different, but verification times were 400 ms longer
in the implicit condition (these results are discussed in section 7.3.2).
However, I do not believe that these results contradict Noordman and
Vonk (1992). By using a sentence connective such as but or because, Noord¬
man and Vonk invite their subjects to find a connection between the two
phrases. When that is easy, because the bridging inference involves no
more than a retrieval from long-term working memory, subjects comply,
resulting in longer reading times for the to-be-linked phrase but verifica¬
tion times that are the same as for explicit sentences. When that is diffi¬
cult, because the domain is unfamiliar and suitable retrieval structures are
lacking, readers do not construct the link they are invited to form. Hence,
they have to do it in the verification phase of the experiment. In the exper¬
iment by Keenan and me, the domain was familiar enough and readers
could have retrieved the missing link readily enough, but being lazy read¬
ers and lacking a linguistic cue specifically requesting a bridging infer¬
ence, they did not make the effort. This added 400 ms to the verification
time, because the linking proposition had to be retrieved at that point from
long-term working memory. Thus, the question “Are inferences made
during reading?” needs to be approached with great care: Background
198 Models of comprehension

knowledge, task demands, and linguistic cues in the text all interact in
determining what will happen in a particular case.

6.3.3 Time course for constructing knowledge-based inferences

Of considerable interest is the time course of constructing knowledge-

based inferences in text comprehension. We know that it takes about 300
to 350 ms for word meanings to become fixed in a discourse context (sec¬
tion 5.1). Inferences require more time. In Till et al. (1988; see also sec¬
tion 5.1.3) no evidence for topic inferences was obtained at an SOA of 500
ms, but topic inferences were clearly made at an SOA of 1,000 ms (there
were no data points in between). In contrast, Magliano et al. (1993) found
that antecedent causal inferences required an SOA of only 400 ms. Long
et al. (1994), in a study modeled after Till et al. (1988), have used SOAs
of 200, 300, 400, 500, 750, and 1,000 ms. Associative effects are fully
apparent in their data already at 300 ms. Topic effects develop gradually:
they are already apparent at 500 ms but increase in strength up to 750 ms.
Because different materials and conditions were used in all these studies,
the differences in the results are not surprising.10 It seems that sentence-
level inferences require from 400 to 750 ms, depending on experimental
conditions. Thus, sentence meanings take roughly twice as long as word
meanings to fixate.

6.3.4 The construction of situation models

Much recent research has been concerned with the construction of situ¬
ation models (e.g., Glenberg, Kruley, & Langston, 1994; Glenberg &
Langston, 1992; Graesser & Zwaan, 1995; Zwaan, Langston, & Graesser,
1995; Mani & Johnson-Laird, 1982; Trabasso & Suh, 1993; van den
Broek, Risden, Fletcher, &Thurlow, 1996). It should be clear from every¬
thing that has been said so far that there is not a single type of situation
model and not a single process for the construction of such models. Sit¬
uation models are a form of inference by definition, and Table 6.1 is as rel¬
evant for situation models as it is for any other inference in discourse

10 Because poor readers do not always make inferences even when more skillful readers
do, some of these estimates (in particular, the Till study) may be unduly inflated by
averaging results over skilled and less skilled readers.
 Textbases and situation models 199

comprehension. That is, situation models may vary widely in their char¬
acter. In the simplest case, their construction is automatic. Relevant
information is furnished by existing retrieval structures, as in the exam¬
ples given for cell A in Table 6.1. Or it may be available simply as a con¬
sequence of a particular form of representation, such as imagery in (14).
Such situation-model inferences do not add new propositions to the
memory representation of the text but simply make available information
in long-term memory via retrieval structures, or information that is
implicit in the mental representation, such as an image (see Fincher-
Kiefer, 1993, and Perfetti, 1993, for similar suggestions). On the other
hand, situation models can be much more complex and result from
extended, resource-demanding, controlled processes. All kinds of repre¬
sentations and constructions may be involved. The process may be shared
by a social group or even by a whole culture and extend over prolonged
periods. Text interpretation is not confined to the laboratory. At least
some experimental methods for studying more complex cases are being
developed today. Trabasso and Suh’s three-pronged method that com¬
bines theoretical analysis, verbal protocols, and experimental procedures
is a promising development in this respect, as is the simple but very effec¬
tive “landscape” method of van den Broek et al. (1996). However, for the
most part current research on situation models prudently sticks to rather
simple cases, as I do in the next section, where the formation of spatial sit¬
uation models is discussed, a topic that presents a particular challenge to
a propositional theory such as the Cl model.

6.4 Spatial situation models

For certain kinds of texts, full understanding requires the construction of

a spatial situation model. If readers form only a textbase, they can achieve
only a superficial understanding of such texts, sufficient for reproductive
recall and recognition but not for reconstructive recall and inferences.
The latter require a workable model of the spatial situation implied by the
text. This distinction between text-based behaviors and behaviors based
on situation models was demonstrated in experiment 1 of Perrig and
Kintsch (1985), who wrote two texts describing the spatial layout of a
small town. Each text had 24 sentences, six of which contained nonloca¬
tive information. The same town was described in both texts, but in one
case a route description was used to present the information and in the
200 Models of comprehension

other the same information was provided in the form of a survey descrip¬
tion. Thus, where one text might say

(19) After two blocks, turn right on 6th Street to reach the school
house.

the other might read

(20) Sixth Street crosses Main Street two blocks north. The school
house is on 6th Street just west of its intersection with Main
Street.

Each subject read either the route or the survey text. Reading rate was
controlled by the experimenter. Half of the subjects read the text once,
and half were given three presentations of the text. After reading, subjects
were given a free recall test and a sentence verification test, with true old
sentences, paraphrases, and inferences, and false distractor sentences.
Reproductive memory, tested either by free recall or verification of sen¬
tences actually in the text (verbatim or not), was quite good. Across both
texts, subjects recalled 22% of the propositions after one reading and 40%
after three readings. Similarly, the d' for the verification of old test sen¬
tences was 1.63 after one presentation and 2.36 after three presentations,
attesting to the fact that these subjects had been able to form an adequate
textbase. However, the subjects had not been able to form a situation
model that would have allowed them to verify the inference questions cor¬
rectly. The d! value for inference sentences was a mere .20 after one pre¬
sentation and rose to only .85 after three presentations. Thus, we have here
a good textbase enabling normal reproductive memory but a weak situa¬
tion model that yielded poor performance on inference sentences.
In their experiment 2, Perrig and Kintsch (1985) used briefer texts (14
sentences), describing a simpler geographic layout, so that readers would
be better able to form a usable situation model. Furthermore, subjects
were allowed to study the text at their leisure. Again, both a route and a
survey version of the text were used. These changes in the experiment
had the desired effects. Overall, reproductive memory remained at a high
level, as in experiment 1: The average recall was 40% of the text propo¬
sitions, and the / for old test sentences was 3.01. However, unlike in
experiment 1, the verification of spatial inference sentences was also
good, with an average / of 1.79. Thus, with a simpler task and more time,
these subjects were able to form an adequate situation model that enabled
them to verify inference sentences fairly accurately.
 Textbases and situation models 201

Although in both experiments the route text led to better performance

than the survey text, readers formed routelike situation models when
they read the route text and maplike situation models when they read the
survey text. Thus, when subjects who read one version of the text were
asked to verify inference sentences in the same version, their perfor¬
mance was better (d' = 2.11) than when they were unexpectedly asked to
verify sentences in the alternate version (/ = 1.46). We also observed an
interesting interaction, in the sense that female subjects showed a prefer¬
ence for the route version of the text, whereas male subjects preferred the
survey version. However, because this interaction was not replicated in
other studies following up this line of research (Taylor & Tversky, 1992),
there is no need to discuss it here.
Perrig and Kintsch (1985) demonstrate the importance of distinguish¬
ing between the textbase and situation model in studies of text compre¬
hension. For this reason, the study is of central importance for the theory
presented here, but it did not include an actual simulation of their data.
Such a simulation was performed for an interesting set of data involving
a spatial situation model by Haenggi, Kintsch, and Gernsbacher (1995).
Haenggi, Kintsch, and Gernsbacher (1995) replicated a well-known
experiment by Morrow, Greenspan, and Bower (1987) in which the inci¬
dental formation of a spatial situation model during story comprehension
was investigated. Perrig and Kintsch (1985) used texts that had only one
goal: to describe the spatial layout of a town. Morrow et ah, in contrast,
wrote texts that told a story about a character moving around in a partic¬
ular spatial setting that was only incidental to the action described.
Hence, there was no strong task demand to form a spatial situation
model. Nevertheless, using a clever experimental procedure, Morrow et
al. showed that their subjects indeed were forming and updating a spatial
model. A number of later studies show that this outcome is by no means
necessarily the case. Whether or not subjects spontaneously form spatial
models depends on a number of subtle factors, but that is not the issue
here. It is sufficient for us that readers under the conditions of the Mor¬
row et al. experiment do form such models.
Haenggi et al. (1995), in their replication of Morrow et al. (1987), had
subjects first memorize a floor plan of a castle, as shown in Figure 6.12.
The castle had four rooms, which could be reached via certain doorways.
In each room were four distinct objects, such as the carafe and the rug in
the dining room. After the subjects had thoroughly memorized this floor
plan, they read a brief story in the course of which the main character
202 Models of comprehension

Figure 6.12 Floor plan of the castle that subjects had to memorize for the
spatial inference task in experiment 1. From Haenggi et al. (1995).

moved through the rooms of the castle. At four points during the story
two objects were presented on the computer screen, and the subjects had
to decide whether the objects were in the same room or in different
rooms. For instance, just after the main character walked into the ball¬
room, CARAFE-RUG was presented as the test pair, requiring a “same”
response. In this case the objects were in a different room than the actor.
On another test both objects were in the same room as the actor, and there
were tests where the objects belonged to different rooms. The main result
of the experiment, just as in Morrow et al., was that the decision time for
“same” responses depended on whether the objects were in the same
room as the actor in the story or in another room. In experiment 1, for
instance, the mean response time for “same” responses was 2.108 s when
the objects were in the same room as the actor, and 2.782 s when they
were in a different room. Readers seem to focus on the location of the
actor, so that objects close to that location become more available than
objects farther away. Obviously this is a dynamic process, for the readers
must update their situation model throughout the story.
 Textbases and situation models 203

In simulating these results, we face a challenging problem because the

spatial layout must be translated into a propositional representation in
such a way that the relevant spatial relations are preserved. Figure 6.12
shows that this can be done in a perfectly simple and straightforward way.
The castle has four rooms, some of which are connected by doors, and
each room contains four distinct objects. Comprehension is then simu¬
lated sentence by sentence in the usual way, and any time an object or
location is mentioned in the story, a link is formed between the knowledge
representation and the text representation. Figure 6.13 shows the net¬
work that is being generated when the sentence.

(21) Penelope walked from the arsenal to the ballroom

is processed. Forming a spatial situation model means that location infer¬

ences are made, in this case IS[PENELOPE,IN-BALLROOM]. Because

Figure 6.13 The associative network of propositions generated to simulate

the integration of text- and knowledge-based information in experiment 1.
Text-based propositions are underlined, and the frame designates the area of
knowledge-based propositions in the network. From Haenggi et al. (1995).
204 Models of comprehension

we wanted to weigh knowledge less strongly than information explicitly

mentioned in the text, all link weights among knowledge propositions
were given a value of 0.5, and the links among text propositions were
given a value of 1. Starting values of 0 and 1 were assumed for knowledge
and text propositions, respectively. The statistic of interest was the aver¬
age final activation values for all test pairs that were presented at various
points in the story. These values were higher when both objects were in
the same room as the actor than when they were in a different room (.25
versus .12, respectively, for experiment 1).
Haenggi et al. (1995) performed two variations on this basic experi¬
mental design (using a linear sequence of rooms, for instance, rather than
a circular layout as in Figure 6.12). Under those conditions, the reaction
times to objects differed depending on which room they were in. For the
most part, the pattern of activation produced by the simulation faithfully
reflected these differences. Thus, the simulation not only accounted for
the main results of these experiments but also duplicated some of the
detailed patterns of the data produced by idiosyncratic features of the
experimental materials, the particular layouts, and the stories used.
The Haenggi et al. (1995) simulations show that it is possible to simu¬
late the construction of spatial situation models with a propositional net¬
work. It may not always be as easy as it was here to preserve the crucial
spatial information in a propositional network as the spatial information
becomes richer and more complex, but we have shown that the model can
handle spatial imagery, at least in simple cases.

6.5 Literary texts11

Der Gedanke geht nach alien Richtungen sofort immer weiter,

die Einfalle wachsen an alien Seiten auseinander heraus, das
Resultat ist ein ungegliederter, amorpher Komplex. Im exakten
Denken nun wird er durch das Ziel der Arbeit, die Beschriinkung
auf das Beweisbare, die Trennung in Wahrscheinliches und
Gewisses usw., kurz durch die aus dem Gegenstand kommenden
methodischen Forderungen verschniirt, begrenzt, artikuliert.
Diese Auslese fehlt hier. An ihre Stelle tritt die durch die Bilder,
den Stil, die Stimmung des Ganzen.

11 This section is based on Kintsch (1994b).

 Textbases and situation models 205

The thought immediately proceeds in all directions; remindings

branch out from each other on all sides; the result is an unstruc¬
tured, amorphous complex. In rigorous thought, this complex
is reined in, delimited, articulated through the goal of the task,
through being restricted to what is provable, through separation
into what is probable, what is certain, and so on, in short, through
the systematic demands that emanate from the object itself. This
selection process is missing here [in artistic thinking]. It is
replaced by selection through the images, style and mood of the
whole.
Robert Musil,
Diaries (August 13, 1910)

Studies of discourse processing, such as the ones presented here, have

traditionally employed brief artificial texts that truly deserve to be called
trivial. The reasons we restrict ourselves to such texts should be obvious:
Their complexity is all we can handle, both in the theoretical analysis and
in most of our experiments. Long natural texts are not only difficult to
simulate with the means at our disposal, but they also provide the subject
with too many opportunities for misunderstandings, slips of attention,
and a sheer unwillingness to cooperate, which makes illusory the predic¬
tions derived from the model of an ideal reader that assumes that every¬
thing in the text is processed equally completely and equally perfectly.
Thus, there are compelling reasons for the choice of such texts. The
practice does not introduce distortions that are too serious (except when
text genres are confused - a miniature description of Margie watering her
flowers is not a story), but it does raise questions about the generality of
the theory. Apart from the practical difficulties involved in working with
very long naturalistic texts, are there in principle limitations that restrict
the theory to simple stories, essays, and descriptions? In particular, how
would the theory handle literary language?
My hypothesis is that the comprehension processes, the basic strate¬
gies, the role of knowledge and experience, as well as the memory prod¬
ucts generated, are the same for literary texts as for the simple narratives
and descriptive texts we have used in our research. That is not to say that
there is no difference, but the difference is in the “what,” not the “how.”
Literary language presents a novel and powerful set of constraints not
present in everyday texts, but these new and different constraints are
processed in the same way as other more familiar constraints.
206 Models of comprehension

Although there are only a few experimental studies of the comprehen¬

sion of literary texts and how they are remembered, it seems certain that
the surface structure of such texts plays a much bigger role in determin¬
ing comprehension and memory than for nonliterary texts. Long texts are
typically reconstructed rather than reproduced faithfully from memory.
In the kinds of texts we have studied earlier, this reconstruction was
based on the situation model. Properties of the surface structure or even
the textbase play a negligible role in the reconstruction process. They are
secondary in comprehension, an arbitrary vehicle for the construction of
a situation model that then becomes the basis of remembering Not so for
literary texts. Surface structure and textbase are carefully coordinated
with the situation model. What words are used, how sentences are
formed, the precise pattern of semantic relationships in the textbase are
not arbitrary but rather are calculated by the author to produce particu¬
lar effects. They function as a set of nonsemantic and nonpragmatic for¬
mal constraints that can play a crucial role in the reader’s integration
process. Not only syntactic constraints are important here but also poetic
forms such as verse, rhyme, alliteration, and so on, which establish for¬
mal relations among elements in the surface structure that affect com¬
prehension and memory. Similarly, the textbase is more than the chaff to
be disregarded as the situation model is extracted from it. It is calculated
to produce remindings that are essential to the effect of the text. In some
literary texts, the situation model itself may be trivial and unimportant.
In others, there may be several coordinated or competing situation mod¬
els that the reader is invited to construct. In the quotation that I used as
a preface for this section, Musil sketches a picture of comprehension that
is not too far removed from the construction-integration model. For sci¬
entific texts, the bottom-up proliferation of ideas generated by a text is
controlled and ordered by the constraints imposed by scientific analysis -
logical argument, scientific theory, and methodology. For literary texts,
textual relations, images, and style, as well as emotional reactions, the
mood of the whole, serve as the effective constraints instead.

6.5.1 Versification as a source of constraints

As a demonstration of the feasibility of this approach the reader by now

surely expects something like a simulation of one of Shakespeare’s son¬
nets. Alas, all I have to offer is a nursery rhyme! It is not only that I do not
 Textbases and situation models 207

dare to be more ambitious; there is also no psychological research to fall

back on in the case of the sonnets. However, there are some rather useful
studies of nursery rhymes.
It often happens in such cases that rhyme and rhythm completely over¬
shadow meaning. As a child, Eileen Kintsch knew the following nonsense
rhyme:

(22) Mersidotes and dosidotes

and little lambs eat ivy.
Diddely-divey-do,
Wouldn’t you?

Only later, after she had acquired more vocabulary and learned to read,
she found out that it wasn’t a nonsense rhyme at all:

(23) Mares eat oats and does eat oats,

and little lambs eat ivy.
Kids’ll eat ivy too,
Wouldn’t you?

The rhyme in this case had overpowered the semantics. It is not an iso¬
lated instance. In comprehending and remembering rhymes, the poetic
language is at least as essential a factor as the meaning. How can the Cl
model deal with such observations?
Kelly and Rubin (1988) have collected a large number of observations
about the following nursery rhyme, which therefore is an ideal object for
our analysis:

(24) Eenie, meenie, minie, mo.

Catch a tiger by the toe.
When he hollers let him go.
Eenie, meenie, minie, mo.

Figure 6.14 shows an analysis of the first two lines of this nursery rhyme.
The surface structure has been subdivided into two parts, the phonology
and the versification. The first text element - (1) in Figure 6.14 - is iden¬
tified phonologically as /ee/. It is linked to the second element /nie/, as
is /ie/, with /mee/, and so on. At the level of versification /ee/ is
not directly linked to /nie/ but with the verse markers VOWEL[l],
BEGINNING-OF-LINE[l], RHYME[1,3], and RHYTHM[1,3,5,7,
STRESSED]. The same element, therefore, enters into a different set of
 Versification Semantics
Phonology

Figure 6.14 1 he network generated by the phonological, versification, and

semantic relations for the first two lines of Eenie, Meeme, Mime, Mo.
 Textbases and situation models 209

relationships, depending on whether we analyze its phonological or verse

properties. This is true for all other elements, too. For instance, text ele¬
ment (2) is related at the phonological level to element (1) as predecessor
and (3) as successor, but at the versification level (2) is related to
RHYME[2,4,6] and RHYTHM[2,4,6, UNSTRESSED], The nonsense
syllables (1) to (7), which are not interpretable semantically, do not enter
into semantic relations with other elements. The text elements (8) to (14),
however, are linked via the relations among the propositions derived from
them.
We have, therefore, in Figure 6.14 three sets of overlapping relations:
at the phonological level a simple string of successive text elements, at the
versification level a rich and complex network, and at the semantic level,
which again is rather impoverished. It is obvious that in such a network
the versification relations will dominate the integration process and not
the semantics. (There are no situation models to speak of with rhymes
like these.)
The Cl model, equipped with the proper construction rules for the ver¬
sification level, should construct precisely such a network as in Figure
6.14. In the integration phase, activation flows through this network until
it settles into a stable pattern (for simplicity, I have assumed all links and
all starting values to be 1). Figure 6.15 shows the results of this integra¬
tion process. The activation of the text elements from Figure 6.14 is shown
in the bottom part of Figure 6.15. The stressed syllables of the first line,
as well as toe, have the highest activation values, and a and the have the low¬
est activation values. The activation of the verse markers is shown in the
middle part of Figure 6.15. On the average, the activation is highest in this
part of the network. The semantic elements - the propositions - end up
with very little activation, as seen in top part of the illustration.
The model thus predicts that the versification relations in this rhyme
should be dominant in memory. This is precisely what Kelly and Rubin
(1988) observed. Historically, this popular rhyme has changed in many
ways, but rhythm and rhyme have remained constant. One finds eena,
dena, dina, do but never eeme, meenie, diney, mo.

6.5.2 Multilevel situation models

For most genres that have been studied so far (descriptive texts, stories,
manuals), a single situation model is assigned to a text. The author’s
210 Models of comprehension

Figure 6.15 The activation values for the network shown in Figure 6.14.

intent is to communicate this model to the reader via the text. How effec¬
tive a textbook or a manual is depends on how well the author succeeds
with this communication. If competent readers construct the wrong
model, the author has failed. Not so for literature. It is almost always the
case that the author intends the reader to construct situation models at
more than one level: At one level there may be an action sequence, at
another a commentary about the social condition, at yet another a moral
message, and so on. It is also not necessary or even possible for the author
to define the intended situation model precisely and uniquely, for what
 Textbases and situation models 211

readers make of a text depends on their personal experiences and the way
they read the text.
As yet, only very little research in psychology and cognitive science has
been concerned with the comprehension processes involving multilevel
situation models. Story grammars, for instance, have focused primarily
on the action sequence in a story, with the moral of the story as a sort of
afterthought (Mandler & Johnson, 1977). On the other hand, there have
been attempts in the context of artificial intelligence to describe the con¬
stellation of actions that can be considered as “betrayal,” or “seduction”
(Lehnert, 1981). Multiple situation models are more familiar in a differ¬
ent context — the study of word arithmetic problems and the use of com¬
puters, for instance, as discussed later this book. There we distinguish
between the situational understanding of a word problem and the formal
problem model, its mathematization (chapter 10), or between the task
model that a computer user might have and the device model that is nec¬
essary to perform this task on the machine in question (chapter 11).
There has also been some work on the role that personal experiences play
in interpreting new situations, although not in a literary context (such as
“reminding” in Schank, 1982, or the work on case-based reasoning by
Kolodner, 1993).
There are no simulations of how multilevel situation models are
formed for literary texts. The general theory for such a task exists, and
there are no compelling reasons why it could not be undertaken. Such a
project would have to be conceived as an interdisciplinary enterprise, for
although the principles of comprehension are presumably the same for
literary texts as for other genres, only literary experts could tell us what
the content of the multilevel situation models ought to be. The cognitive
scientist can construct the situation model, but the content of the model
and the strategies used in its construction are domain specific. This is no
different from other cases - for example, when we must ask a computer
expert what the device model should be, or when we consult a biologist
about the biology content of an instructional text.

6.5.3 Comprehension strategies for literary language

Do the strategies used for comprehending literary text differ from those
used in other genres? This issue has received little attention among cog¬
nitive psychologists, but there are some interesting studies on how bal-
ladeers in primitive societies without a written record reproduce their
212 Models of comprehension

songs from memory or, rather, reconstruct them (Lord, 1960; Wallace &
Rubin, 1988). Strategies similar to those these singers use to reconstruct
their songs are employed by educated persons in comprehending literary
texts. We thus can gain an impression of the rather complex strategies
that must be acquired before literary texts can be appreciated.
Lord (1960) describes a well-known Serbo-Croatian singer of ballads
who was able to repeat a 2,294-line ballad he had never heard before after
listening to it once. His reproduction was 6,313 lines long and contained
many embellishments, character descriptions, and an emotional depth
that were lacking in the somewhat naive original. The strategies that make
such a feat possible are reasonably well understood (Lord, 1960; Rubin,
1988).
The songs of the balladeers — in Homer’s day as well as in our own
(Wallace & Rubin, 1988) - largely consist of metric formulas (90% in the
Odyssey) that can be used whenever needed, irrespective of the semantics.
Thus, we find “brilliant Odysseus,” “resourceful Odysseus,” or “long-
suffering, brilliant Odysseus,” depending on whether the verse requires
two syllables, three, or more. In composing the song, this metric vocabu¬
lary is organized around particular topics. A large variety of scripts is
used for this purpose. For instance, the topic “council meeting” is used
over and over again in just as stereotyped a manner as the restaurant
script in AI. This includes various subscripts, such as the arrival of the
hero, with horse and weapons always being described in the same order.12
Equipped with this vocabulary and his scripts, the balladeer is able to
concentrate entirely on the action and characters of the story - not unlike
musical improvisation, especially in jazz.
The balladeer employs a variety of strategies that enable him to regen¬
erate songs of a certain kind. (There is no claim to originality here; the
balladeer insists that he merely reproduces faithfully what he has heard.)
These generative strategies are employed effortlessly, automatically, and
unconsciously. Thus, they must be well rehearsed before they can be
used. Furthermore, the singer’s memory must contain a rich store of sim¬
ilar songs as well as related usable information and episodes from every¬
day life. According to the findings of Ericsson and Chase (1982), we can
safely assume that this material is organized in memory in certain ways.
Thus, what we have here is a form of expert memory, which, as is well

12 These examples are quoted after Wallace and Rubin (1988).

 Textbases and situation models 213

known, enables one to achieve a great deal but which can be acquired only
with much effort and over an extended period (e.g., Ericsson, 1985).
Comprehending literary texts makes similar demands. The reader
must be able to employ the required encoding strategies automatically
and must have readily available the knowledge basis needed for the use of
these strategies. Everyone can read and remember a trite newspaper
story. We are all experts in this regard, the required linguistic encoding
strategies and the necessary domain specific knowledge having been prac¬
ticed over a whole lifetime (Ericsson & Kintsch, 1995, chap. 7). Literary
texts are different. Unfamiliar linguistic strategies are needed for com¬
prehension, beginning with verse and rhyme, all the way to schemes for
the organizational principles of the novel. Information about the author
and the time and social context in which the text was generated may be
relevant. Remindings and reminiscences of related texts may be funda¬
mental for the interpretation of the text. I am suggesting, therefore, that
the comprehension of literary texts should be regarded in the same way
as any other expert performance. Peak performance, accordingly, would
probably require a decade of intensive study. Of course, a literary text can
be enjoyed at some level even without that expertise, but deep under¬
standing is reserved for the expert.
Returning to the question posed at the beginning of this section - Is
the comprehension of literary texts different from that of nonliterary
texts? - the answer must be “yes” and “no.” Yes, because literary texts
demand specific encoding strategies and specific knowledge that do not
play a role in comprehending nonliterary texts. Specifically, the encoding
strategies for literary language are different from those employed for
everyday language, and specific domain knowledge is required to under¬
stand literary texts. No, because the psychological processes involved are
the same in both cases: The “what” is different, but the “how” is the
same. A simulation of a literary text more complex than Eenie, meenie,
minie, mo is quite conceivable, but it would be unwise to attempt it until
the comprehension theory can be expanded beyond the realm of the
purely cognitive. Without a computational account of the emotional reac¬
tions (chapter 11) and aesthetic experiences that play a central role in
understanding literary text, such a simulation would be unsatisfactory.

The macrostructure of a text — intuitively, its gist — plays a major

role in memory and comprehension, especially for longer texts.
214 Models of comprehension

Texts differ in the extent to which this structure is made explicit.

Several experimental studies are discussed here. The first is con¬
cerned with the question whether the formation of a macrostruc¬
ture is an optional process in text comprehension, or whether it
is an integral, automatic part of understanding. We provide evi¬
dence that macrostructures are formed automatically even in the
absence of task demands. In a second study, we investigate the
linguistic markers that authors employ to signal the macrostruc¬
ture of a text. We also show that the macrostructure of a text may
dominate the comprehension process to such an extent that
material contradictory to what the reader takes to be the text’s
macrostructure is simply ignored.
Inferences are involved in the construction of situation mod¬
els. I argue that different types of “inferences” in text compre¬
hension must be distinguished. On the one hand, a distinction
exists between information that is retrieved from long-term
memory and information that is newly constructed by some sort
of inference rule. On the other hand, a distinction must be made
whether the process is automatic or controlled. Thus, we have
automatic retrieval (as in most of the classical bridging infer¬
ences), automatic generation (as in a transitive spatial inference),
controlled retrieval (e.g., when a memory search is required for a
bridging inference), and controlled generation (real inferences
based on some specific rule, including the rules of logic).
Two special cases of situation models are explored in greater
detail, both through simulations and experimentation. One is
spatial inference: keeping track of the location and movements of
an actor in a story. It is shown that the Cl model does this as a
normal by-product of comprehension, without having to postu¬
late special strategies or processes. The second case we have
investigated is the comprehension of literary texts, with a simu¬
lation of a simple example in which various literary constraints
(such as rhyme) play a significant role.
 7

The role of working memory

in comprehension

The differences between how much people remember in one sit¬

uation and how little in another can be dramatic. I listen to a child
tell a friend what happened at a party she attended a few days ago.
Now, I wasn’t there and cannot vouch for the accuracy of every¬
thing she said, but the impression she gives is that she remem¬
bers a lot about who said and did and wore what. And she had fun
when she was there, she wasn’t memorizing anything. Labora¬
tory experiments on memory provide a stark contrast to this
scene. Years ago, when psychologists did paired-associate exper¬
iments with nonsense syllables as stimuli, being an experimenter
was very boring; you sat there and watched helplessly as your
subject tried over and over again to reach criterion on that list.
When finally the list was reproduced correctly twice in a row, the
experimenter was just as grateful as the student that the torture
was over. No wonder theorists of verbal learning were obsessed
with the idea of interference!

Memory for text is usually quite good, and often very good. There is an
issue here, because if the same memory system is used for remembering
text as for learning the nonsense syllables and word lists we study in the
laboratory, the discrepancy between the good text memory and the poor
list memory needs to be explained.
The span of immediate memory is plus or minus seven items. From a
list of 30 random words, college students recall about 12 to 14 after one
reading, adults somewhat fewer. It takes about one hour to memorize 100
random words. How can people live with such a terrible memory?
They can’t and they don’t. The memory demands of many cognitive
216 Models of comprehension

tasks exceed the memory capacities we observe in the typical laboratory

experiment by a large margin, but memory outside certain kinds of labo¬
ratory situations is really quite good. Analyses of complex cognitive tasks,
such as playing chess or making medical diagnoses, or, indeed, text com¬
prehension, reveal significant memory demands. For instance, in van
Dijk and Kintsch (1983, p. 347) we presented a figure that summarized
the contents of working memory assumed to be involved in text compre¬
hension. These ranged from sensory features, linguistic expressions,
propositional structures, and situation models to control structures,
goals, lexical knowledge,Trames, and schemata, as well as episodic mem¬
ory traces of prior text and context. Because all these demonstrably play
a role in text comprehension, they must somehow be available in working
memory during comprehension. This amount of material is nowhere to
be fitted into a classical short-term working memory of seven chunks.
Analyses of other problem-solving tasks have come to the same conclu¬
sion (e.g., Newell, 1990).
Furthermore, it is clearly not the case that memory for text is as poor
as studies of list learning suggest. After one reading of a story, free recall
is quite good, and cued recall is excellent. Similarly, cued recall of lec¬
tures, conversations, newspapers, and so on is generally good and is ade¬
quate to support us in our daily life. Recall of a news story in the labora¬
tory was comparable to incidental recall when the news item was read in
the course of everyday activities (Singer, 1982).
Thus, the contrast between the extremely limited memory found in
the laboratory and memory in everyday situations, as well as certain
expert activities such as playing chess, solving physics problems, or mak¬
ing medical diagnoses is striking, indeed. What does it mean? Some
researchers have concluded from these observations that the laboratory
study of memory is ecologically invalid and has produced misleading and
useless results. Jenkins suggested that possibility as early as 1974 in his
influential article “Remember that old theory of memory? Well forget it!”
and he has had many followers since. I propose a very different solution
to this dilemma. There is nothing wrong with the findings from studies
of list learning and the theories of memory based on them; all they need
is an extension to everyday and expert activities. This extension is pro¬
vided by the theory of long-term working memory of Ericsson and
Kintsch (1995). It accepts memory theory as it is and shows how it can be
extended to account for the good or excellent memory found in certain
 The role of working memory in comprehension 217

activities by means of concepts developed within the framework of the

classical theory.

7.1 Working memory and skilled memory

Long-term memory (LTM) is everything a person knows and remem¬

bers: episodic memory, semantic memory, as well as declarative and pro¬
cedural knowledge.1 Just because a person remembers something or
knows something does not guarantee that this memory or knowledge par¬
ticipates in a given cognitive process at a given time, even though it would
be relevant for that process. The analogy with a computer may help to
understand the issue: A great deal of information is stored in various
forms accessible to the computer, but this information does not affect
processing unless it is retrieved and installed in the computer’s central
processor. The human central processor is called working memory
(WM). To affect a cognitive process, items in long-term memory must be
retrieved and inserted into working memory. Working memory is thus
the active part of long-term memory, as shown in Figure 7.1a. There are
other names for the active part of long-term memory - short-term mem¬
ory (STM), the focus of attention, and consciousness. At least for present
purposes these names refer to roughly the same phenomena. Psycholo¬
gists were therefore quick to identify working memory with short-term
working memory. Baddeley, our foremost authority on working memory,
defines working memory as “the temporary storage of information that is
being processed in any range of cognitive tasks” (Baddeley, 1986, p. 43).
This short-term working memory (ST-WM) is insufficient to account
for the role of memory in cognitive processes. This problem is discussed
in considerable detail in Ericsson and Kintsch (1995), and I only summa¬
rize briefly the arguments presented there. The capacity of STM is
severely limited. Miller (1956) suggested that STM contains as many
as 7 ± 2 chunks, but the best current estimate of STM capacity is about 4
chunks (Broadbent, 1975). This capacity is sufficient to account for
the phenomena of list learning in the laboratory, which explains why
researchers were satisfied for so long with the equation of WM and SI M.

1 The distinctions between these different types of memory, such as episodic and
semantic memory, are irrelevant for present purposes, although they may be impor¬
tant in other contexts.
218 Models of comprehension

Figure 7.1 (a) The classical theory of working memory: ST-WM is the acti¬
vated portion of LTM, containing no more than 4 to 7 chunks, (b) The long¬
term working memory theory of Ericsson and Kintsch: LT-WM is the part
of LTM (gray circles) linked to the chunks in STM (circles) by retrieval
structures.

It is only when we study expert memory, inside and outside the laboratory,
that this severe capacity constraint becomes intolerable. I have already
mentioned the many things a discourse comprehender must maintain in
working memory in order to succeed at comprehension. Expert perfor¬
mance in chess, mental calculation, scientific problem solving, and med¬
ical diagnosis makes similar demands on working memory. Furthermore,
 The role of working memory in comprehension 219

in all these tasks one can show that working memory involves a long¬
term component, not some sort of expanded temporary storage in a
super-STM. This is implied by the repeated findings of insensitivity to
interruption, good incidental long-term recall, and resistance to interfer¬
ence. Apparently, these experts have learned to use parts of their long¬
term memory as working memory - the long-term working memory, or
LT-WM.
Figure 7.1b illustrates how LT-WM functions. The items available in
the capacity-limited STM serve as retrieval cues for those parts of LTM
that are connected to them by retrieval structures (as explained in section
7.3). A single retrieval operation, using one of the cues in STM, makes
available that subset of LTM memory that is linked to the cue by such a
retrieval structure. A retrieval from LTM by a cue in STM requires about
300 to 400 ms (see Ericsson & Kintsch, 1995). Hence, the amount of
information in working memory consists of two sets of items: those
already in ST-WM, which are accessible very rapidly though not instan¬
taneously (Sternberg, 1969), and those reachable by a retrieval structure
in about 400 ms. Whereas the capacity of ST-WM is strictly limited, that
of LT-WM is constrained only by the extent and nature of the retrieval
structures that can be accessed via the contents of STM.
Retrieval structures were first studied in detail by Chase and Ericsson
(1982), who investigated a subject who was able to increase his digit span
to more than 90 digits after extended training. Later, many other subjects
were taught to extend their digit span to 30 or more items. Such feats
were achieved by developing and automating efficient encoding strategies
to store digits in LTM. This requires a large body of relevant knowledge,
for the encoder must be able to perceive familiar patterns in the digit
sequences that are to be memorized and to associate these patterns with
retrieval cues. Schemata retrieved from LTM must be used to further
organize these retrieval cues into stable retrieval structures that will sup¬
port the quick and reliable recall of the digit sequence to be learned, for
instance, S. F., Chase and Ericsson’s original subject, was a runner who
used his extensive knowledge about running times as an encoding
scheme. Typically, he formed four-digit chunks by associating the digits
with some remembered numerical fact about running — for example,
3,596 might be coded as 3 m 59.6 s, or just under a four-minute mile. He
then used spatial relations to encode these chunks into higher-order
groups, forming a hierarchical retrieval scheme, as shown in the upper
part of Figure 7.2. Having formed such a structure, S. F. was able to recall
220 Models of comprehension

Figure 7.2 A retrieval structure for digits i to n stored in long-term working

memory. On top is the hierarchical organization of retrieval cues associated
with groups of digits. On the bottom are the knowledge-based associations
relating units of encoded information with patterns and schemas. After Erics¬
son and Kintsch (1995).

the corresponding digits at any location in this structure that the experi¬
menter requested.
Subject S. F. further strengthened this retrieval scheme through elab-
orative encoding. He not only associated digits with a retrieval cue but
used higher-order relations among the encodings he had generated to
organize his retrieval structure. He might, for instance, have noted that
several successive chunks were all related to running times for a mile,
thereby forming a supergroup. The resulting retrieval structure thus was
organized both hierarchically and elaboratively, as shown in Figure 7.2.
There are several important things to keep in mind about retrieval
structures such as that shown in Figure 7.2. Only a very rich knowledge
base enables one to form such structures, because many different patterns
and associations must be readily available for encoding. Suitable knowl¬
edge bases exist only in areas where a person has reached a high level of
 The role of working memory in comprehension 221

expertise in a special expert domain, or in everyday life situations where

everyone is an expert. Knowledge alone, however, is not sufficient; one
also needs well-developed and well-practiced encoding strategies. In the
case of text comprehension, these encoding strategies are the normal
operations engaged in during comprehension, but in other instances, for
example, a waiter memorizing dinner orders, they must be specifically
learned and practiced. A great deal of practice is required for encoding to
be successful, given the real-time constraints of most tasks: Only rapid,
reliable, automatic operations will be effective. Thus, LT-WM is severely
constrained and does not come easily. It is restricted to situations in
which people have true expertise - knowledge as well as encoding skills.
Long-term working memory provides the basis for reconciling the poor
performance in traditional laboratory tasks and the very good memory in
some everyday situations and expert domains. If people have the knowl¬
edge and the encoding skills in a domain, they are able to use their LT-WM
to extend working memory by building retrieval structures that allow
them to access LTM. For text comprehension, the skills that build
retrieval structures are the same skills that are involved in comprehension
in the first place, and hence, unlike domain-specific encoding strategies,
do not have to be acquired specifically. Thus, in domains in which a reader
has a good knowledge background, LT-WM can be used. To the extent
that knowledge is lacking, LT-WM is unavailable. As a consequence, both
comprehension and memory should be affected. Stories, for instance, are
about human goals and human actions, something all of us are very famil¬
iar with. Scientific or other technical texts, on the other hand, require
domain knowledge not shared by everyone. Therefore, stories are better
remembered than technical texts (see Kintsch, Kozminsky, Streby, Mc-
Koon & Keenan, 1975, for a comparison of narratives and history texts).

7.2 Long-term working memory in text comprehension

In the traditional memory laboratory, conditions were intentionally

arranged so that the role of knowledge was minimized. Subjects therefore
had to rely on their ST-WM exclusively. Thus, memorizing became a dif¬
ficult, strenuous task, and its outcome was unreliable. Subjects in these
experiments had no chance to concentrate their efforts on important
information. They could not be selective, because the experimenter had
set up the experiment so that every item was equally important. I ext
222 Models of comprehension

comprehension is different. It is rapid and, if the text is well written and

the domain familiar, effortless. Immediate cued recall is virtually perfect.
And the reader knows which aspects of the text to concentrate on and
which can be neglected, so that important information is much more
likely to be encoded than insignificant details (the “levels” effect in recall,
as reported by Kintsch, 1974; Meyer, 1975; and many others).
Retrieval from long-term memory varies, depending on whether or not
LT-WM is involved. In the traditional experiment, retrieval is slow and
unreliable. The principle of encoding specificity severely restricts the
effectiveness of retrieval cues because only cues encoded with the mater¬
ial to be learned at the time of learning will be effective. Proactive inhibi¬
tion limits the retrieval of stimuli that are similar. In domains where LT-
WM can be used, retrieval is rapid, reliable, and flexible. It is flexible not
because the encoding specificity principle has lost its validity, but because
the material that is encoded in a retrieval structure is richly linked to var¬
ious regions of LTM, from which it can be successfully retrieved. Proac¬
tive inhibition is either absent or greatly reduced because of this use of
elaborate retrieval structures.
Ericsson and Kintsch (1995) discuss these differences between ST-
WM and LT-WM in detail, and they also describe a number of different
theoretical predictions that flow from this distinction and that can be
tested empirically. I focus here on one of these predictions.
A particularly striking instance of the theory of LT-WM making dif¬
ferent predictions from the classical theory concerns the effect of inter¬
rupting reading comprehension. Glanzer and his colleagues (Fischer &
Glanzer, 1986; Glanzer, Dorfman, & Kaplan, 1981; Glanzer, Fischer, &
Dorfman, 1984) have reported a series of experiments in which reading
was interrupted after each sentence of a text by various activities and for
various lengths of time. As a prototypical example, consider the study
described in Fable 7.1. After each sentence of a text, the subject read
another unrelated sentence. After the eight-sentence text had been read,
with or without interruptions, the subject was given a number of com¬
prehension questions. The results were quite striking: The interruptions
had no effect whatever on comprehension. Subjects answered correctly as
many questions about the interrupted text as they did when they read the
text without any interruptions. The only effect was an increase of about
400 ms in the reading time per sentence after an interruption.
These results are difficult to interpret within the classical theory of
 The role of working memory in comprehension 223

Table 7.1. Example of a text used in the interruption experiment by Glanzer et at.
(1981). Sentences 1, 3, 5, . . . comprise a connected discourse; the interspersed sentences
2 ,4, 6, . . . are unrelated

51 James Watts’ first steam engine was completed after many years of thought
and labor.
52 Honeybees gather nectar from flowers of a particular kind during spring.
53 Even then much work was needed to make it practical.
54 Democratic principles are not always observed in the developing countries of Asia.
55 He supplied enough money to make the machine a success.
56 Water management in the arid states of the Southwest is affected by legal issues.
57 Mine owners came from all over to see the engine work.
Etc.

working memory, for reading an unrelated sentence surely must wipe out
any traces of the prior text from the reader’s STM. Resource-consuming
reinstatement searches should be required, and comprehension should
be badly degraded. The theory of LT-WM, however, readily accounts for
the observed results. The next sentence of a text following an interrup¬
tion provides the cues in STM that can retrieve the LTM trace of the
previous text from LT-WM. The mental structure that the reader has
created in the process of comprehending the text itself functions as a
retrieval structure. Hence, with every new sentence, the reader gains
access to the previous memory trace of the text in LTM, at the cost of a
single retrieval from LT—WM, which takes about 400 ms.
Many versions of the Glanzer experiment exist, with longer and
shorter interruption intervals and different interrupting activities such as
doing simple arithmetic problems. A full discussion of their results can be
found in Ericsson and Kintsch (1995). All these results are in agreement
with the interpretation in terms of LT-WM offered here. Interrupting a
text does not interfere with comprehension because LT-WM allows the
reader reliable access to the LTM trace of the prior text with a single
retrieval operation. Similar observations have been reported concerning
the interruption of other highly skilled cognitive operations, such as play¬
ing a chess game.
Text comprehension is structure building. To comprehend a text
means forming a mental structure that represents the meaning and mes¬
sage of the text. Different theories of text comprehension (e.g., Gerns-
bacher, 1990; Just & Carpenter, 1987; van Dijk & Kintsch, 1983) differ on
224 Models of comprehension

the precise characteristics of that structure, but they agree on the central
issue of structure building. Any one of these theories could therefore in
principle serve as a theory of LT—WM. Whatever mental structure is
incidentally generated in the process of comprehension also serves as a
retrieval structure, thereby generating LT—WM. In the next section I
explore in somewhat more detail how the text representations generated
by the Cl model function as retrieval structures.

7.3 Retrieval structures in the Cl model

Memory plays two distinct roles in text comprehension. First, texts must
be comprehended sequentially, one sentence after another; therefore,
when focusing on one sentence, memory is needed to provide access to
the prior text. Second, text representations are not based solely on the
text but require significant contributions from the comprehender’s long¬
term memory. Hence, relevant structures and facts must be continuously
retrieved from long-term memory during comprehension.

7.3.1 Episodic text memory during comprehension

Comprehension implies the creation of a coherent mental representation

of a text. The reader can focus on only a small portion of the text at any
time — typically, a sentence - and to establish coherence must frequently
reinstate in the focus of attention items from the episodic text memory
under construction. Normally, texts are written in such a way that a sen¬
tence contains cues (e.g., overlapping word concepts or explicit connec¬
tives) that link it with previous sentences in the text. These cues provide
access to the episodic text structure in LT-WM. Figure 7.3 illustrates
how retrieval structures function in text comprehension. The elements of
the structure, the propositions derived from the text, are connected in a
hierarchical macrostructure. The text propositions become linked in with
various LTM structures - schemas, frames, scripts, associations, or
episodic memory units, as shown in the lower part of Figure 7.3. The end
result, the episodic text memory, is a tightly interconnected structure, in
part through already established links in LTM as the text becomes inte¬
grated into the reader’s knowledge and experience, and in part through
novel links - the links among propositions generated in the construction
of the micro- and macrostructure. Note the similarity between Figures
7.2 and 7.3. The episodic text representation that is formed in the Cl
 The role of working memory in comprehension 225

Figure 7.3 A retrieval structure for a text (the propositions P1-P6) stored in
long-term working memory. On top is the hierarchical macrostructure of the
text. On the bottom are the knowledge-based associations relating units of
encoded information with scripts, frames, and schemas.

model, the textbase/situation model, has all the formal properties of a

retrieval structure as described in skilled-memory theory. Our claim is
that it also functions as one.
When a new proposition is formed as a sentence is being read, it
becomes part of the textbase/situation model structure. It is linked to
other propositions in that structure to which it is related in various ways.
In the microstructure, propositions may be linked because they share a
common argument, because of temporal, spatial, or causal relations, or
because of a common schema; further links are generated by the macro¬
structure, because groups of micropropositions are subordinated under
a common macroproposition. Thus, a newly constructed proposition
is associated with other propositions in working memory, some of which
are already linked to propositions in LTM. It thereby becomes capable
of retrieving propositions that have been generated previously in the
226 Models of comprehension

process of comprehension but are now part of the network stored in

LTM. An LT-WM has been created.
Givon (1995) has pointed out that the syntax of a sentence serves as
processing instructions to the comprehender for the formation of a dis¬
course representation. For instance, different forms of anaphora indicate
to the comprehender in which part of memory a referent is likely to be
found. In English, one can refer to a prior referent via zero anaphora, a
pronoun, or a full noun phrase. Zero anaphora or an unstressed pronoun
is used when the writer (or speaker) expects the referent still to be avail¬
able in ST-WM. When the referent is more distant but still retrievable via
LT-WM, a stressed pronoun is used. When the referent was last men¬
tioned in some other, not directly related part of the text, requiring a
more complex LTM retrieval, it is reintroduced by a full noun phrase.
Consider the following example:

(1) John hit a deer driving [0] late at night on a forest road. It was dark
and at first he [Punstressed] did not realize how much damage had
been done. However, the mechanic at the garage said that the bill
would be very large. The insurance agent took a long time to
inspect the damage, and he [Pstressed] got quite impatient. In the
meantime in the forest, three vultures circled the dead deer [NP].

In the first sentence, zero anaphora is used for John as the agent of dri¬
ving. In the second sentence, John is referred to with an unstressed pro¬
noun. In both cases John is presumably still available in the reader’s focus
of attention. The stressed pronoun in the fourth sentence signals a refer¬
ent in LT-WM, namely John. When the deer is mentioned in the last sen¬
tence, it is reintroduced with a full, definite noun phrase, the dead deer.
Needless to say, there is a great deal more to say about the use of anaphora
in English (e.g., Chafe, 1974; Greene et ah, 1994; see also section 5.2).
Besides their other roles, anaphora also serve as directives for the use of
memory. They are by no means the only processing signals. Givon (1995)
has discussed other signals of this kind. His evidence is based on the lin¬
guistic analysis of texts that shows that anaphora are indeed employed in
the way he claims. For example, the distance between a referent and its
anaphora is generally short for zero anaphora, moderate for stressed pro¬
nouns, and long for full noun phrases. Such data are very important and
strongly support his analyses. It is to be hoped that psycholinguists will
also find ways to test these claims with their methods.
 The role of working memory in comprehension 227

7.3.2 The activation of knowledge during comprehension

In the foregoing example, the retrieval structure was constructed by the

comprehender on the spot. Next we turn to retrieval structures that pre¬
exist in long-term memory and that are used during comprehension in
building new structures.
Figure 7.4 shows a simple sentence and some putative knowledge that
would have been available for activation during comprehension. Actually,
all the LTM items to which the various elements of the sentence are
linked should be on this list. However, I have selected only a few exam¬
ples, including some weird, highly idiosyncratic, and context-inappropri¬
ate ones, just to illustrate the context-independent nature of the process.
For hank a context-appropriate and a context-inappropriate associate are
shown; robbery is related to a compound cue that comprises all the ele¬
ments of the sentence. If all the elements shown were actually retrieved
and took part in the integration process, the ones related only to a single
item would become deactivated, and robbery would become the strongest
node in the network because of its centrality.
None of these associations are actually retrieved in the normal process
of comprehension. They and the many others not shown in Figure 7.4 are
all available and could be retrieved because they all lie within the reader’s
LT-WM. But they are not retrieved and do not enter the reader’s

TEXT: ACTIVATED KNOWLEDGE:

Two masked-* Mardi gras

gunmen-* cowboy
made their getaway
with $100,000
from the
First National
Bank ^_-» river

money

robbery

Figure 7.4 Long-term working memory: some links between the words of a
sentence and items in LTM.
228 Models of comprehension

STM/consciousness/focus of attention unless there is some reason for

it. If a reason exists, they can be accessed within the time that a Lr-WM
retrieval requires, that is, 400 ms. One reason for retrieval is, for instance,
that the reader is participating in a priming experiment and is shown
either river or money after the word hunk in a sentence of a lexical deci¬
sion task. The word hank is embedded in a LT-WM retrieval structure
that has been generated through a lifetime experience with the use of that
word. It includes both river and money. 1 hus, both are accessible right
after the presentation of the word, before integration has taken place.
Hence, when either river or money appears as a test word, it has a small
advantage over unrelated control words that are not part of the reader s
LT-WM. To recognize a test word as an English word, the corresponding
lexical node must be sufficiently activated. This activation process is
strengthened and speeded up when the lexical node in question is already
linked to a structure in the reader’s STM, which is the case for river and
money right after bank appears in the example given.
If the lexical decision test is delayed for at least 350 ms after bank, the
word has become integrated into the discourse context and its context-
appropriate meaning has become fixed, whereas its context-inappropriate
meaning has become deactivated. Hence, associates such as river that
belong to the context-inappropriate sense of hank are no longer accessi¬
ble via the reader’s LT-WM and are therefore no longer primed.
It takes even longer (at least 750 ms; the time estimates given here are
based on Till et al., 1988) for the test word robbery to be primed. For that
to happen, the meaning of all the words in the sentence must be fixed,
that is, a complete sentence representation must have been formed. Only
such a compound cue, rather than a single word of the sentence, is likely
to retrieve robbery, which does not occur in the sentence. Thus, it takes
about 750 ms to infer the sentence topic in this case. Note, however, just
what “infer” means here: not that robbery has become an element of
STM/consciousness but that the sentence representation that has been
constructed is linked by LT-WM to that concept. If it is introduced into
STM (by being used as a test word in a lexical decision task, as in our
example, or because of other task demands, such as the question “What
is the topic of the sentence?”), it is readily available from LT-WM.
The theory of LT-WM requires us to reinterpret many other observa¬
tions that have been labeled traditionally as inferences. As another exam-
 The role of working memory in comprehension 229

pie, consider the bridging inferences studied by Kintsch and Keenan

(1973). Subjects received such texts to read as the following:

(2) Implicit
A burning cigarette was carelessly discarded. The fire destroyed
many acres of virgin forest.

(3) Explicit
A burning cigarette was carelessly discarded. It caused a fire that
destroyed many acres of virgin forest.

Immediately afterward subjects were shown the following test sentence

for verification:

(4) The burning cigarette caused the fire.

On the immediate test, verification times for explicit test sentences were
400 ms faster on the average than for implicit test sentences. In the case
of the explicit sentences, all three propositions CAUSE [CIGA¬
RETTE,FIRE], DISCARD[BURNING[CIGARETTE]] and [DE-
STROY[FIRE,FOREST]] were active in STM, and hence the test sen¬
tence could be verified quickly. In the case of the implicit sentence, only
DISCARD[BURNING[CIGARETTE]] and [DESTROY[FIRE,FOR¬
EST]] were in STM, but because both are linked associatively to
CAUSE[CIGARETTE,FIRE], this proposition could be retrieved from
LT-WM within about 400 ms, which added that much to the verification
time for the implicit test sentence.
It is misleading, however, to talk about a bridging inference in this case,
because that would imply that by some inference operation the proposi¬
tion CAUSE[CIGARETTE,FIRE] was inserted into the reader’s work¬
ing memory independently of the test statement. What happened,
instead, was that as a normal consequence of comprehension, the bridg¬
ing proposition became part of the reader’s LT-WM, and when the test
statement required it, it was readily accessible.
Given the right knowledge about what happens when burning ciga¬
rettes are thrown away in dry forests, comprehension assured that this
linking knowledge was available when needed. Without the right knowl¬
edge, this bridging cannot happen. Consider the following sentence pair:
230 Models of comprehension

(5) An abnormally low amount of hydrocele was found. The sper¬

matic cord was quite dry.

For most readers, hydrocele and spermatic cord retrieve nothing. A real
bridging inference is required to establish the coherence of such a text. A
reader might reason that the spermatic cord is some sort of place where
hydrocele, which seems to be a kind of fluid, is located. This is deliberate,
conscious inferencing, reflected in the reader’s verbal protocol, unlike the
automatic knowledge access that occurs in a familiar domain.
This is not the place for a comprehensive discussion of either the prim¬
ing data or of topical inferences. (For another discussion of these issues,
see chapters 5 and 6, respectively). My goal here is merely to show how
the theory of LT-WM is relevant to the interpretation of these phenom¬
ena.
We have already discussed the interruption studies of Glanzer and his
colleagues (e.g\, Glanzer et ah, 1981). Their studies constitute strong
empirical evidence for the operation of LT-WM. In comprehending easy
texts in familiar domains, readers construct coherent, orderly text repre¬
sentations, tied into their LTM, which serve as efficient retrieval struc¬
tures, almost obviating the effects of the interruption. If the textbase
formed were incoherent, however, no effective retrieval structure would
be available, and interruptions should have a deleterious effect. McNa¬
mara and Kintsch (1996) tried to arrange for such conditions. Instead of
the simple story Glanzer used, we chose difficult scientific, technical
texts in domains with which our subjects were unfamiliar. Furthermore,
instead of interrupting after each sentence of a text, we interrupted in the
middle of each sentence. A subject read the last half of the prior sentence
plus the first half of the present sentence and was then interrupted by an
unrelated sentence. Thus, at the time of interruption whatever represen¬
tation a subject had formed was necessarily incomplete, so it would be
less likely to function as a retrieval structure.
The results of this modified interruption procedure are shown in Fig¬
ure 7.5, together with the results of a replication of Glanzer’s work with
familiar texts and interruptions after complete sentences. In neither case
was there an effect on comprehension as measured by the subjects’ free
recall. Subjects recalled as much with midsentence interruptions as with
end-of-sentence interruptions as with no interruptions at all. However,
 The role of working memory in comprehension 231

Figure 7.5 The difference in reading times per sentence between normal
reading and interrupted reading for familiar texts with end-of-sentence inter¬
ruptions and for unfamiliar texts with midsentence interruptions. After
McNamara and Kintsch (1996).

there was an effect on reading times. The interpolation effect shown in

Figure 7.5 is the additional reading time per sentence that subjects needed
when the text was interrupted by unrelated sentences. For familiar text
with end-of-sentence interruptions, we replicated Glanzer almost pre¬
cisely - reading times were 400 ms longer on the average, indicating LT-
WM retrieval. For midsentence interruptions with difficult texts, how¬
ever, reading times per sentence increased by about 1,400 ms. As expected,
our procedure made it impossible for the subjects to use their LT-WM,
but they were able to retrieve the previous sentence context from LTM in
spite of that, using whatever cues, temporal as well as content, they had
available. The time required for reinstating the previous sentence context
under these conditions agrees well with other estimates of retrieval times
from LTM (Ericsson & Kintsch, 1995).
A final illustration of the role of LT-WM in discourse comprehension
concerns the consequences of forming different kinds of text representa¬
tions. On the one hand, readers might form a representation that is based
primarily on the meaning and rhetorical structure of a text, with minimal
links to LTM. On the other hand, a text representation may reflect the
232 Models of comprehension

actual text and its organization only sketchily, whereas LTM structures
are used to organize and elaborate the textual material. Text representa¬
tions are always a mixture of textbase and situation model, but one or the
other component may predominate. In other words, either the top or bot¬
tom half of the structure shown in Figure 7.3 is emphasized. Further¬
more, depending on their knowledge base, different readers may form
quite different situation models.
This argument can be elaborated by considering in more detail the role
of LT-WM in the process of comprehension.

Textbase dominance. Textbase dominance typically occurs when a reader

lacks the background knowledge necessary for a full understanding of the
text. Each word of the text (or, in any case, most words) is understood,
that is, it is linked to its lexical node and the associative neighborhood of
that node, thus generating a LT-WM. But this LT-WM consists of many-
different unconnected islands in the LTM structure, because the reader
does not have the knowledge necessary to form associative links between
the separate words of the text. Thus, LT-WM cannot play a major role.
Any item retrieved from such a fragmented LT-WM will become deacti¬
vated because it is connected to the text structure merely by a single link.
Hence, recall will be reproductive and remain close to the text structure.
Textbase dominance may also occur when a reader has relevant back¬
ground knowledge but does not use it during comprehension. Passive
readers are not rare, and to ensure learning from text, such readers have
to be jolted out of their passivity and induced to assume a more active
comprehension strategy. These issues are further discussed in chapter 9.
It is not clear, however, what it means in terms of LT-WM theory to say
“a reader does not use his knowledge during comprehension.” If the
knowledge is there, it ought to be automatically linked to the text, form¬
ing an LT-WM. Surely, the elicitation of river and money by bank is not
under strategic control, but perhaps the use of the whole sentence as a
compound cue (to retrieve robbery in Figure 7.5) is under strategic con¬
trol. Thus, good readers, who infer sentence topics (Long et al., 1994),
may use compound retrieval strategies, whereas poor readers do not.
Alternatively, it might be a question of the strength of the associations:
LTM items are included in LT-WM only if they are connected to an ele¬
ment of STM with a sufficiently high strength. The difference between
good and pool readers in that case would be that the former have not only
 The role of working memory in comprehension 233

more but also stronger links in their LTM structure. Poor readers may
need some sort of help to bring their text-LTM associations above
threshold strength so that an effective LT-WM is formed.

Situation model dominance. At the other extreme, a person with rich back¬
ground knowledge forms a very different LT-WM. The text is not only
linked to (mostly useless) single-word associates but also to rich and
extensively interconnected larger knowledge structures that connect dif-
ferent.text elements. Thus, activation of a schema or script might connect
a large number of text propositions in different parts of a text. When such
a structure is retrieved from LT-WM, it tends to assume a dominating
position in the integration process because of its centrality in the net¬
work. High-knowledge persons, therefore, will tend to reconstruct a text
when trying to recall it in terms of these general knowledge structures
and will reproduce correspondingly fewer text propositions and less of
the text structure.
These different outcomes can be observed in studies in which persons
at different levels of expertise recall medical case records and diagnoses
(Groen & Patel, 1988; Schmidt & Boshuizen, 1993). Interns behave like
readers who mostly depend on the textbase and have only a weak situa¬
tion model. Thus, they reproduce parts of the text without clearly differ¬
entiating essential and inessential information and make few knowledge
intrusions. The recall of doctors with several years of practice, in con¬
trast, clearly reflects their situation model. It is very rich, and much mate¬
rial is included that was not in the case record but was inferred in the
process of arriving at a diagnosis. In particular, this includes general med¬
ical and physiological knowledge that is relevant to then understanding of
these cases. Interns do not include such material in their recall protocols
either because they do not have the requisite knowledge or because,
although they learned these things in medical school, their knowledge is
too weak and insufficiently situated to be used in diagnosis.
Interestingly, the real experts, doctors with many years of practice,
produced much briefer and quite different recall protocols than the doc¬
tors with less experience. They too relied on their LT-WM and lecon-
structed rather than reproduced the protocols, but because their situation
model was very different from that of less experienced doctors, their LT-
WM and what they recalled were different, too. Expert doctors gave brief
recalls concerned with the essential symptoms and diagnosis of a case.
234 Models of comprehension

This information was not embedded in general medical knowledge, sug¬

gesting that these experts no longer used that knowledge in understand¬
ing a case. Instead, their rich experience gave them an alternative basis for
organizing and understanding this material. Their extensive direct expe¬
rience provided them with the requisite LTM structures, and they no
longer needed their school knowledge as a crutch. One can be quite con¬
fident that their medical knowledge was still intact; however, the links
between observations and medical knowledge had become rather weak,
and their representations were dominated by their much stronger links
between observation and experience.

7.3.3 The short-term memory buffer

In the text-processing theory of Kintsch and van Dijk (1978), the STM
buffer played a crucial role. It was the only way a coherent text represen¬
tation could be constructed. According to that theory, a few elements, or
at least one, from one processing cycle must be maintained in the STM
buffer to be reprocessed together with the elements of the next cycle, in
the hope that some link (argument overlap) will be found so that a coher¬
ent text representation can be formed. The buffer was the bridge in the
model between processing cycles that permitted the formation of a
coherent mental representation of the text, which had to be processed
sentence by sentence. The capacity of the buffer became an important
issue, though most estimates suggested that it remained relatively con¬
stant at one or at most two propositions.
This picture must now be modified in some important respects. First,
the Glanzer data discussed earlier (Glanzer et ah, 1981) clearly demon¬
strate that comprehension is possible without the use of an STM buffer.
The interruption procedure surely wipes out the contents of STM - yet
comprehension is unaffected, except for a slight increase in reading
times. Second, the theory of LT-WM provides an alternative mechanism
for the construction of a coherent text representation. Links between text
elements from different processing cycles are formed via LT-WM.
Thus, both empirically and theoretically, comprehension can do with¬
out an STM buffer. However, there is good evidence (Fletcher, 1981;
Fletcher & Bloom, 1988) that an STM buffer is used in normal uninter¬
rupted comprehension. One may question why an STM buffer is used
when comprehension obviously can proceed without it, but just because
a mechanism is redundant does not mean that it is useless. The STM
 The role of working memory in comprehension 235

buffer, for instance, allows for the formation of a more coherent textbase.
Propositions maintained in the buffer can be linked with the propositions
derived from the next sentence on the basis of purely textual relations,
whereas LT-WM links must be based on prior knowledge. High-knowl-
edge readers are probably not as dependent on the buffer, but interfer¬
ence with the STM buffer might have stronger effects for low-knowledge
readers.

7.4 An LSA model of retrieval structures

The notion of retrieval structure has appeared in several places in the lit¬
erature on comprehension and problem solving in recent years. Richman,
Staszewski, and Simon (1995) propose a formal model of retrieval struc¬
tures in recall in the context of the EPAM theory of memory. Clark and
Marshall (1981) describe how what they call “reference diaries” function
as retrieval structures in a conversation, somewhat like macrostructures
do in reading a text (Figure 7.4). Myers, O’Brien, Albrecht, & Mason
have reported a series of experiments in which they explore how the con¬
tent that is in the focus of attention during reading accesses related prior
information in the text or in the reader’s knowledge base (e.g., Albrecht
& Myers, 1995; O’Brien, 1995; O’Brien & Albrecht, 1991). Myers uses a
well-chosen metaphor for this process, namely, that long-term memory
resonates with the contents of short-term memory; the items that res¬
onate most strongly become accessible in working memory (Myers et al.,
1994). Myers’s resonance theory is a specific, mathematical model of
retrieval structures in comprehension. It builds on memory theoiy, in
particular the SAM formulation of Gillund and Shiffrin (1984), to
describe how resonance works in comprehension and is in principle quite
compatible with the Cl model. Indeed, Varma and Goldman (1996) have
implemented a version of the Cl model that includes a resonance-based
reinstatement mechanism.
Retrieval structure, therefore, is not just a useful but vague concept;
rather, it is something that can be modeled formally. I he possibility of
formally specifying retrieval structures brings up another problem, how¬
ever. How can we objectively and with some degree of accuracy hope to
model the reader’s long-term memory? As suggested in section 3.3, latent
semantic analysis may provide at least a partial solution to this problem.
Latent semantic analysis (LSA) can be used to model retrieval struc¬
tures. Retrieval structures are fixed, well-established links in long-term
236 Models of comprehension

memory. Such links should be related to semantic distance as measured

by LSA. Therefore, one may hypothesize that in text comprehension the
probability that a particular piece of knowledge is activated in long-term
memory is proportional to the cosine between the vector of the eliciting
item and the knowledge item.
With this assumption, we can readily model knowledge elicitation, or
reminding, in discourse comprehension. Consider example (12) dis¬
cussed in section 6.2.3, which is reprinted here.

(6) SI John was driving his new car on a lonely country road.
52 The air was warm and full of the smell of spring flowers.
53 He hit a hole in the road and a spring broke.
54 John lost control and the car hit a tree.

What are likely words that LSA might be reminded of when reading SI,
John was driving his new car on a lonely country road?Table 7.2 shows two
of the 10 most closely related terms in the encyclopedia LSA space for
each of the content words in that sentence.
For John, LSA thinks of two historical Johns; for new it knows no bet¬
ter than New York and New Jersey. But suppose we add these remindings
to the network created for that sentence in (6), using the cosines from the
LSA analysis for link strengths. Integrating, we obtain the results shown
in Figure 7.6, where the areas of the squares used for each node are
approximately proportional to the final activation value of that node.
Note that lonely has completely disappeared from the network, as have
most of the intuitively irrelevant remindings from the list in Table 7.2.
Only rail and York remain as obviously irrelevant associations but with
activation values barely above zero.
If we similarly select two associates from the top 10 neighbors of each

Table 7.2. Most closely related terms in the LSA

space for each content word in example (6)

John Butterfield, Milton

drive driver, gears
car driver, automobile
new York,Jersey
lonely tomboy, girl
country nation, landlocked
road rail, highway
 The role of working memory in comprehension 237

Figure 7.6 The activation values for the words from the sentence John mas
driving his nem car on a lonely country road (gray squares) as well as for several
remindings (white squares) and the sentence node S (black square). The area
of each square indicates the activation value of the node.

content word in the other three sentences of example (6), the following
knowledge intrusions can be obtained.

52 The air was warm and full of the smell of spring flowers:
smell drops out, humidity, dew, showy, and leaves are added.
53 The car hit a hole and a spring broke:
driver, automobile, smash, and album are added.
54 John lost control and the car hit a tree:
automobile, trunk, and bark are added.

Obviously we have a method to predict knowledge intrusions in read¬

ing comprehension, which may eventually lead to some empirical tests.
Equally obviously, the method is not perfect yet, probably because of the
nature of the LSA space used. As discussed in section 3.3, the LSA space
on which all calculations are based is derived from the analysis of about 4
million words from an encyclopedia. Thus, what LSA knows is heavily
weighted toward technical information such as one finds in an encyclo¬
pedia, but its knowledge of everyday things is woefully inadequate. For
instance, all top 10 neighbors of hit from example (1) have to do with pop¬
ular music, and it knows nothing about cars hitting a hole in the road or a
238 Models of comprehension

tree along the road. Therefore, LSA misunderstands things because of its
inadequate knowledge base. For instance, the nearest neighbor in the
LSA space for the sentence SI, John driving his car around, is the docu¬
ment “New Frontier.” This is not as puzzling as it appears: Put together
John, country, new, road - quickly you get into Camelot country. One
might say LSA is doing a good job here. If it knew more about the every¬
day worlds - knowledge that LSA might get from reading newspapers,
novels, movie scripts - it might also be reminded of more prosaic drives
in the country. The example also illustrates another limitation of the pre¬
sent version of LSA: its disregard of syntax. Sentence SI is not about
John (FK) leading his country along a new road but about driving on a
country road.
In spite of these caveats about LSA, its use for modeling the operation
of retrieval structures in comprehension opens exciting new possibilities.
Only further research will show how much of this promise can be real¬
ized. It should be possible, however, to employ LSA to derive specific, a
priori predictions for experiments such as those of Myers and his col¬
leagues - with a better, more informed LSA space.

7.5 Capacity differences in working memory?

In section 9.1 there is a discussion of the factors that make for a good or
a poor reader. Decoding skills, language skills, and domain knowledge are
emphasized. Another factor that has often been suggested as a determi¬
nant of reading comprehension is the capacity of working memory. The
theory of long-term working memory described earlier implies that indi¬
vidual differences in comprehension may be the result of skill and knowl¬
edge differences. Capacity differences in working memory may or may
not play an additional role (see also Ericsson & Kintsch, 1995).
At first glance, the data seem to contradict any relation between read¬
ing comprehension and STM capacity. Reading comprehension improves
enormously between childhood and adulthood, but measures of STM
capacity such as the digit span show only a slight improvement (e.g., Chi,
1976; Huttenlocher & Burke, 1976; and others). Good and poor readers
cannot be distinguished on the basis of their memory span (e.g., Farnham-
Diggory & Gregg, 1975; Rizzo, 1939). However, such data are not really
conclusive, because the memory span merely measures a subject’s abil¬
ity to store unfamiliar digit sequences and does not provide a direct indi¬
cation of the STM capacity available during reading. Daneman and
 The role of working memory in comprehension 239

Carpenter (1980), therefore, developed an alternative measure of STM

capacity, called the reading span, which directly indicates how much capac¬
ity a reader has available during reading. They gave subjects a list of unre¬
lated sentences to read and later questioned the readers about the content
of these sentence to make sure the readers actually comprehended them.
After reading the sentences, subjects were asked to recall as many of the
last words of the sentences as they could. The reading span was defined in
terms of the number of words subjects could recall in this task. Thus, it
provides a measure of memory during reading. Unlike the traditional
memory span, the reading span correlates highly with reading compre¬
hension on a variety of tasks. In the studies reported by Daneman and Car¬
penter (1980) and others, correlations between reading span and text com¬
prehension have ranged from .5 to .9.
Thus, a clear and strong relation between STM capacity and reading
comprehension appeared to have been established. Cantor, Engle, and
Hamilton (1991) and Just and Carpenter (1992) further developed the
notion that STM memory capacity, as measured by the reading span, is
the basis for individual differences in reading comprehension.
However, questions arose early on whether the reading span really was
a measure of STM at all. Masson and Miller (1983) reported evidence
that the whole sentences were stored in LTM (rather than the final word
being maintained in STM). Similarly, Baddeley (1986) showed that mea¬
sures of reading span could be obtained in situations in which the sub¬
jects were not told beforehand which word to recall and hence had to
retain the whole string in memory. Baddeley’s measures of reading span
correlated with text comprehension just as well as the original one, but
they must have been based on LTM storage, for such a large amount of
information could hardly be retained in STM. Thus, the reading span is
certainly a good predictor of reading comprehension but apparently not
a measure of STM capacity.
The theory of long-term working memory proposed by Ericsson and
Kintsch (1995) provides an alternative interpretation for these findings.
What the reading span measures is the efficiency with which readers can
comprehend sentences and hence store them in long-term memoiy.
More skilled readers construct better representations and hence have
available more effective retrieval structures; if they recall anything at all
from a sentence, they are more likely to reintegrate the whole sentence
and retrieve the last word than are low-skill readers. To use a metaphor, it
is not that good readers have a larger box to put things in for temporary
240 Models of comprehension

storage, but that they are more skilled in putting things into long-term
storage and retrieving them again.
This reinterpretation of the findings of Daneman and Carpenter
(1980) fits many of the results that have been reported in support of the
capacity theory. Differences in language skills, not STM capacity, explain
these findings. Cantor and Engle (1993), for instance, had subjects with
high and low reading spans memorize either a list of unrelated sentences
or a list of thematically related sentences. The number of statements in
these sentences that related to a particular concept was varied. A fan
effect is typically observed in such situations when subjects are asked to
verify these statements: The more statements there are about a particular
concept, the slower the verification time. Indeed, this was what Cantor
and Engle observed in both low- and high-span readers when the sen¬
tences were unrelated. When the sentences were related, however, low-
span readers still showed the typical fan effect, whereas this effect was
reversed for high-span readers: The more statements about a particular
concept, the faster were the latters’ verification times. The fan effect
occurs because the more unrelated statements subjects have to remember,
the larger the search set becomes on a verification test. For thematically
related sentences, however, high-span readers could use their superior
organizational skills to construct a single, coherent representation. The
more information there was about a given concept, the more strongly
integrated this representation became and the better it was able to serve
as a retrieval structure. Hence, the negative fan effect. The low-span
readers, on the other hand, did not fully succeed in constructing a coher¬
ent representation of the thematically related sentences because of their
inferior organizational skills. The mental representation they created was
therefore only partially coherent, which resulted in a typical fan effect.
We can similarly explain the results of Just and Carpenter (1992) in
terms of LT-WM processes. In one of their experiments, Just and Car¬
penter showed that only high-span subjects took advantage of pragmatic
cues when comprehending reduced relative clauses. Consider the follow¬
ing two sentences with unreduced relative clauses:

(7) The evidence that was examined by the lawyer shocked the jury.
(8) The defendant who was examined by the lawyer shocked the
jury.

From these sentences, a baseline measure can be obtained for the reading
 The role of working memory in comprehension 241

time for the phrase by the lawyer. Reduced relative clauses can be con¬
structed by omitting the that was or who was from the original sentences:

(9) The evidence examined by the lawyer shocked the jury.

(10) The defendant examined by the lawyer shocked the jury.

Omitting this important syntactic cue makes the sentence harder to

understand, which should increase the reading time for the by the lawyer
phrase. However, there is a difference between these sentences: The
reduction in the defendant sentence created an ambiguity that is not pre¬
sent in the evidence sentence. A natural way to continue The defendant
examined . . . would be with an object of examine, such as the documents.
The continuation with by the lawyer is unexpected and should create a
garden-path effect - the reader needs to take extra time to correct wrong
expectations. This is not the case for the evidence sentence, because we
know that an inanimate noun could not be the subject of examine. High-
span readers in fact react like this; they read sentences of the evidence type
about 50 ms faster than ambiguous sentences with animate subject nouns
{defendant). Low-span readers read both sentences equally fast, that is,
they do not take advantage of the pragmatic cue provided by the inani¬
macy of the sentence subject. Just and Carpenter (1992) claim that this is
so because they do not have the STM capacity to take pragmatic infor¬
mation into account. Ericsson and Kintsch (1995) claim that this effect is
a consequence of the fact that low-span readers are low-skill readers.
They know how to use obvious syntactic cues like that, but they cannot
take advantage of more subtle cues such as the animacy of the noun and
hence do not form expectations about the continuation of the sentence
based on it. High-span readers, in contrast, know and use such informa¬
tion in the comprehension process.
In another, very informative experiment, Just and Carpenter (1992)
show that knowing a lot is not always helpful but may even work against a
good reader. They used sentences like:

(11) The experienced soldiers warned about the danger before the
midnight raid.

This sentence has a simple structure, and low-span readers read through
it without difficulties. High-span readers, on the other hand, pay more
attention to the subtle cues the language provides and notice a potential
242 Models of comp rehension

syntactic ambiguity; this might be another reduced relative clause, in

which case it might continue something like

(12) The experienced soldiers (who were) warned about the danger
were ready.

Readers who form an expectation like this will stumble over the last word
of the actual sentence and experience a garden-path effect. Indeed, that is
what happened to high-span readers, who took about 130 ms longer to
read raid than did mid- and low-span readers.
Just and Carpenter (1992) attribute this result to STM capacity. Low-
span and high-span readers alike form both expectations (that the sen¬
tence will continue as a normal sentence, as it actually does, or that there
is a reduced relative clause). However, low-span readers have insufficient
memory capacity to retain both alternatives in STM and hence drop the
less likely parse before they come to the end of the sentence. High-capac¬
ity readers retain both parses to the end, which results in unnecessary
processing difficulties. An alternative possibility is that the less skilled
low-span readers never consider the possibility of a reduced relative
clause and hence have no difficulty. High-span readers habitually use
more of the information available to them in the parsing process, which
in this case gets them into trouble. Pearlmutter and MacDonald (1995)
have actually investigated whether both low- and high-span readers are
sensitive to the cues in question and found results in accord with our
claims. They found that only high-span readers were sensitive to proba¬
bilistic constraints in the sentences used in their experiment. Low-span
readers ignored these constraints. It was not that low-span readers did
not know about these constraints. They had the knowledge when appro¬
priately questioned, but that knowledge was not sufficiently automated
to be used for the encoding of sentences in an actual comprehension
task. This illustrates a very important feature of retrieval structures.
Simply knowing something - that is, being able to retrieve it under opti¬
mal circumstances - is not sufficient for a retrieval structure; the knowl¬
edge must also be strong, stable, well practiced, and automated, so that
it can be employed for encoding without additional resource demands.
Retrieval structures are a characteristic of expert memory. Unskilled,
low-span readers may have the knowledge, but it is not in readily usable
form.
Thus, much of what has been taken as studies of working memory
capacity turn out to be studies of comprehension skills. This may also be
 The role of working memory in comprehension 243

the case in an experiment on lexical ambiguity resolution by Miyake, Just,

and Carpenter (1994), who wrote sentences in such a way that a lexically
ambiguous term does not become resolved until several words later, as in
(13):

(13) Since Ken liked the boxer, he took a bus to the nearest pet store
to buy the animal.

There were large differences between high- and low-span readers with
such sentences. Low-span readers became confused when they had to
activate the nonpreferred sense of the ambiguous noun at the end of the
sentence, whereas high-span readers had no such difficulties. However,
when the disambiguating phrase followed right after the homograph,
low-span readers performed just like high-span readers. Clearly, low-
span readers knew both word senses.
The capacity interpretation of this finding is that both low- and high-
span readers activated both word senses, but only the high-span readers
had enough working memory capacity to maintain both meanings until
the end of the sentence, when it was needed for understanding. Low-span
readers were forced to drop the less likely meaning because of the limita¬
tions of their working memory capacity, and hence they had difficulties
understanding the end of the sentence. The trouble with this interpreta¬
tion is that we know (e.g., section 5.1.2) that word meanings in context are
fixed in about 350 msec; hence, by the end of (7) all readers must have
deactivated the nonpreferred word sense. However, high-span readers
had built a good retrieval structure that enabled them to reactivate with¬
out difficulties the lost word sense when it was needed, whereas low-span
readers had to engage in more cumbersome recovery processes because
the tentative sentence structure they had built did not serve them as an
efficient retrieval structure.
This interpretation is strengthened by another finding Miyake et al.
(1994) reported. Differences between high- and low-span readers were
observed only with sentences like (13), where the homograph has a pre¬
ferred and a nonpreferred meaning. With equibiased homographs, there
were no differences between high- and low-span readers. I his observa¬
tion is difficult for a capacity theory but is readily explained if knowledge
and skill differences are involved. Low-span readers know all about word
senses, but only in the case of the more frequent, equibiased homographs
is that knowledge sufficiently automated and usable. Again, knowledge of
word meanings is not sufficient to support the formation of a retrieval
244 Models of comprehension

structure — the knowledge must be well integrated and well connected to

the knowledge base.
There is also considerable evidence that high- and low-span readers
employ different strategies. In a study of discourse ambiguity, resolution
of Whitney, Ritchie, and Clark (1991) used texts like the well-known
washing-machine story of Bransford and Johnson (1972) without titles
and obtained think-aloud protocols from subjects reading these texts.
High-span readers maintained several possible scenarios while they were
reading, which they tested against each new sentence. Low-span readers,
in contrast, either made numerous wild guesses, or they gave up trying to
make sense out of these stories. Thus, the discourse processing of high-
span readers was characterized by orderly, systematic, and effective
strategies, whereas low-span readers simply did not know what to do.
Low- and high-span readers do different things when it comes to dis¬
course processing, with the result that high-span readers are able to make
more efficient use their ST-WM and LT-WM capacity.
Singer and Ritchot (1996) argue quite plausibly that individual differ¬
ences abound everywhere we look in cognition, and they report a study
that they interpret as evidence that both knowledge/skill and capacity
differences play a role in comprehension. However, their results can be
reinterpreted as showing the effects of two types of retrieval structure:
general knowledge and skills involved in reading, as measured by the
reading span test and specific fact and world knowledge, which was
needed in their experimental task to make certain required inferences.
Perhaps it is true that eventually there will be a place for both capacity
and skill explanations of individual differences in comprehension. Fur¬
ther research that more clearly differentiates these two aspects will be
needed. It is quite clear, however, even at this point, that the availability
of reliable retrieval structures is a major factor in reading comprehension,
as Ericsson and Kintsch (1995) have argued.

This chapter started out with a well-known dilemma for memory

theory; it suggested a solution for it and then went on to explore
the rich and manifold implication this solution has for the under¬
standing of the role of memory in discourse comprehension. The
dilemma for memory theory lies in the contradiction between
the laboratory findings of severe capacity limitations of short¬
term memory and the immense demand most higher cognitive
The role of working memory in comprehension 245

processes make on short-term memory capacity and the appar¬

ent ease with which real people meet these demands. Memory
theorists usually prefer to forget this dilemma, and ecologically
minded psychologists use it as an argument against laboratory
psychology. Ericsson and Kintsch (1995) have shown that the
solution to this dilemma lies well within the laboratory and that,
indeed, the knowledge that we have gained through laboratory-
research provides a ready explanation for it. Retrieval structures
help people to overcome the capacity limits of short-term mem¬
ory and thus make possible all higher cognitive functions. All
information that is linked by a retrieval structure to a cue in
short-term memory can be retrieved in a single retrieval opera¬
tion. Thus, as long as we can rely on retrieval structures, all
potentially relevant and necessary information in long-term
memory is just one retrieval step away during problem solving
and comprehension. However, retrieval structures are not a nat¬
urally given part of memory but must be acquired through long
and hard practice. Retrieval structures are not just knowledge
links, but links that are strong and allow fully automatic perfor¬
mance. Such structures are typically acquired in becoming an
expert in a cognitive domain - chess, medical diagnosis, and so
on. They are relevant to text comprehension because most edu¬
cated adults are experts in text comprehension, at least in famil¬
iar domains.
It has been argued in earlier chapters of this book that the
term inference is too loosely applied in text comprehension.
When one reads in a familiar domain in which stable retrieval
structures have been acquired through lifelong practice, any item
in working memory that is linked to long-term memory is able to
tap into a system of rich retrieval structures that makes available
a large body of relevant information on demand. The compre-
hender does not have to “infer” this body of information, or even
less, install it in short-term memory/consciousness. It is just
there if it is needed for one reason or another. It is there only if
the comprehender stays within his or her area of expertise, how¬
ever. Take the scholar away from his specialty and the scientist
from her subject, let the chess player play bridge instead, and
give a college student a philosophical essay to read instead of
246 Models of comprehension

the sports page of the newspaper or a trite little story, and their
memory for what they read and did and their ability to activate
relevant knowledge will not be characterized by the theory of
retrieval structures but by the classical theory of short-term
working memory.
 8

Memory for text

I am intrigued as I continually rediscover that what looks effort¬

less in nature can be so laborious to compute.
Brian Hayes in an article on the collision of billiard, balls
in American Scientist (July 1996)

8.1 Recognition memory

8.1.1 List-learning data and theory

The typical recognition memory experiment consists of a study trial fol¬

lowed by a test trial. On the study trial, subjects are shown a list of items
- words, for instance. On the test trial, subjects are shown a test list con¬
sisting of the previously shown (“old”) items and a set of new, distrac-
tor” items. Their task is to say “yes” or “no” whether each test item has
been presented before. The principal features of the results of item recog
nition experiments are as follows.

1. Recognition memory is very good and long-lasting in contrast to

other memory tests. In a classical experiment, Shepard (1967) employed
a pool of 300 frequent and 300 infrequent words, of which 540 weie
shown on a study trial. Recognition memory was then tested with a
forced-choice test: 60 of the study words were randomly selected, and
each word was paired with one of the 60 words not shown. Recognition
was 88% correct, which means that subjects were able to remember about
475 of the 540 words they had seen. Similar results were obtained with a
pool of 612 sentences; performance was at 89% correct. Even more
impressive were Shepard’s results on picture memory. After seeing 612
248 Models of comprehension

pictures a single time, subjects were tested for forced-choice recognition

immediately and after two hours, three days, one week, or four months.
The percentage of correct recognitions were 97%, 100%, 92%, 87%,
and 58%, respectively. Thus, there was very little memory loss even after
seven days. After four months, however, performance was no longer sig¬
nificantly different from the 50% chance level.
2. As described by signal detection theory for sensory detection, rec¬
ognition performance must also be decomposed into two separate com¬
ponents: (a) the discriminability of old items and distractor items in
memory, and (b) the decision criterion adopted by the subject. Dis¬
criminability depends on the way study items are encoded, the length
of the retention interval, and, most important, the similarity between
study and test items (e.g., Underwood & Freund, 1968). By varying the
similarity between study and test items, recognition can be made arbi¬
trarily easy or difficult. The decision criterion depends on instruc¬
tions, the retention interval, and the benefits or costs associated with the
decisions.
3. In many ways, recognition memory is like recall memory. For
instance, context effects are as important in recognition (Thomson, 1972)
as in recall; the short-term memory function for recognition (Shepard &
Tegthsoonian, 1961) looks much like the short-term memory functions
for other memory tests. The main difference between recognition and
recall is that recognition is much less sensitive to organizational factors.
The structure and organization of a memory trace are central for retrieval
in recall tasks but play a much lesser role in recognition (e.g., Kintsch
1977a).
4. The primary recognition procedure can be described as a familiarity
judgment. On a recognition test, the subject assesses how “familiar” a test
item is and responds “yes” if the familiarity exceeds some criterion value.
Models of recognition memory describe how this familiarity assessment
is achieved.

Early models of recognition memory, such as Anderson and Bower

(1972) and Kintsch (1966), were flawed because their central assump¬
tion about the recognition process was wrong. In these models, a famil¬
iarity value was computed by comparing a test item to its correspond¬
ing memory image and computing a match score. Empirical evidence
 Memory for text 249

failed to support this assumption, and it is rejected by all current mod¬

els of recognition memory. Instead, it is assumed that a test item is com¬
pared to the whole set of study items in memory. Several recognition
models have been proposed that share this common assumption. These
models differ radically in the form of the memory representation they
use. In Murdock’s TODAM, for instance, all items in a list are repre¬
sented by a single vector (Murdock, 1993); in contrast, in Hintzman’s
MINERVA, not only each item but each presentation of each item
results in a separate memory trace (Hintzman, 1988). In Gillund and
Shiffrin’s (1984) model, the representation of memory is a network, or
matrix. In spite of these differences about what memory really is, all
these models make rather similar predictions that account quite well for
the existing data.
Perhaps behavioral data alone cannot settle the basic issues about the
nature of memory representations. But they certainly have provided good
models that describe and predict recognition data, and they have allowed
us to identify the essential features of recognition processes: that strength
and bias must be distinguished according to signal detection theory, and
that a test item must be matched against the whole memory set.
I propose here to model sentence recognition data by using the Cl
model to describe the memory representations generated from sentences
in a discourse and a recognition model to describe the recognition pro¬
cess. Because none of the three recognition models is obviously superior
to the others, we are free to choose our favorite on the basis of other con¬
siderations. The Gillund and Shiffrin (1984) model is most suitable for
our purposes because it employs a matrix representation that is closely
related to the one used in the Cl model. Hence, structures generated in
the Cl model can be directly used as input to the Gillund and Shiffrin
model, whereas they would have to be transformed in nontrivial ways to
make them suitable for the other two models.
Gillund and Shiffrin (1984) analyze item recognition experiments. A
list of items is studied, resulting in an interconnected network of mem¬
ory images. On a recognition test, old items (I;) as well as distractor items
(D) serve as cues. In addition, the general list context (CX) is a cue that
plays an important role. 1 he strength values S(Ij,Ij) of the links between
the various cues and memory images are given by a cues-by-memory
images matrix:
250 Models of comprehension

MEMORY IMAGES

h
CX S(CX,Ij)
II
h
Ii SCI,,!,)
Cues

D S(D, I,) S(D, Ij) S(D, In)

If a cue is used to probe memory, its effectiveness depends not only on

how strongly it is linked to a memory image but also how strongly that
image is linked to the current list context. In fact, for cue i and memory
image j, the probe strength is given by the multiplicative rule:

S(CX,Ij) * SCI,,!,.) (1)

Even though items i and j may be strongly associated in memory, if item

j is unrelated to the list context, it will not be linked to the probe. Associ¬
ations unrelated to the context play no role, as demanded by Tulving’s
encoding specificity principle (Tulving & Thomson, 1973).
The familiarity of a test item i is the sum of the strength values of the
links between all probe sets of which it is a part and the images in memory:

(2)
all j

In many situations, we are interested in the familiarity value of a com¬

pound cue consisting of more than two items in which case the products
in equations 1 and 2 will contain more than two terms (see equation 3).
A yes-no recognition decision is made by employ ing a signal-detection
decision rule with the familiarity values calculated in this manner: If the
familiarity of a test item is greater than some critical value, respond
“yes”; if it is less, respond “no.”
This brief excursion into the results and theories of recognition mem¬
ory experiments prepares the ground for a discussion of sentence recog¬
nition.
 Memory for text 251

8.1.2 Sentence recognition

How well does the Cl model account for the data from sentence recogni¬
tion experiments? An earlier attempt was made by Kintsch, Welsch,
Schmalhofer, and Zimny (1990) to answer this question. Some simula¬
tions consistent with the present version of the model are presented
below (see also Singer & Kintsch, in preparation).

Sentence recognition data. Sentence recognition data have much in com¬

mon with list item recognition data. Recognition performance with sen¬
tences is generally very good, decreases relatively little with delay, and the
probability of a false alarm is a function of the similarity of old and dis-
tractor items. Beyond that, sentence recognition data are much more
complex and richer than item recognition data. Whereas a word list is
either unstructured or minimally structured (e.g., by categories), texts
are highly structured, which has important implications for the results of
sentence recognition experiments.
In particular, the probability of recognizing a sentence from a text can
be predicted from the position of the sentence in the text hierarchy, with
structurally important propositions, especially macropropositions, being
recognized better than unimportant, detail propositions (e.g., Walker &
Yekovich, 1984).
The probability of a false alarm to a distractor sentence depends on the
nature of the relationship between the text and the distractor sentence.
False alarm probabilities are ordered from highest to lowest for para¬
phrases, inferences, topically related distractor sentences, and unrelated
distractor sentences (Kintsch et al., 1990). Delay has only minor effects
on the hit rate for old items but raises the false alarm rate for paraphrases,
inferences, and related distractors. Figure 8.1 shows the data of Zimny
(from Kintsch et al., 1990) on which these conclusions are based.
Kintsch et al. (1990) have further analyzed this pattern of delay effects
in terms of the levels of representation of a text in episodic memory by
computing separate decay rates for the situation model, the propositional
textbase representation, and surface memory. Surface memory was esti¬
mated by the difference between verbatim sentences and paraphrases;
textbase memory was estimated by the difference between paraphiases
(in the text but not verbatim) and inferences (true, but not in the text);
and situation model memory was estimated by the difference between
inferences and distractors (true vs. false; neither appeared in the text).
252 Models of comprehension

—□— New-Inap.

.O. New-Ap.

—•O'— Inference

—a— Paraphrase

---83--- Old

Figure 8.1 Recognition probabilities for old sentences, paraphrases, infer¬

ences, contextually appropriate new sentences, and contextually inappropriate
new sentences as a function of delay. After Kintsch et al. (1990).

Their results are shown in Figure 8.2. Under the conditions of their
experiment, no memory loss was observed for the situation model in a
four-day period; the propositional textbase lost about half of its strength
over that period, and surface memory was lost completely.
Differences in instructions are another factor that contributes signifi¬
cantly to the complexity of sentence recognition data. Results are similar
whether subjects recognize a sentence as having been part of a text or
whether the sentence is true with respect to that text (Schmalhofer,
1986). However, differences are obtained when subjects are asked to rec¬
ognize a sentence or to judge its plausibility (Reder, 1982, 1987). Her
recognition data are about the same as those shown in Figure 8.1, but the
plausibility data show a different pattern of results. There are almost the
same number of “yes” responses to old sentences and inferences (roughly
an 8% difference), on both the immediate and delayed test, and there is
essentially no forgetting.
Sentence recognition data, therefore, are fairly complex. The challenge
to the Cl model is to account for this pattern of results. Which of the
empirical phenomena are simply a consequence of the basic architecture
of the Cl model? What additional assumptions are needed to account for
these data, and how justifiable are these assumptions?
1 he Cl model is not a model of recognition memory and hence it must
 Memory for text 253

Surface

Q. Textbase

—O— Model

Figure 8.2 Estimates of trace strengths for surface, textbase, and situation
model traces as a function of delay. After Kintsch et al. (1990).

be combined with one. As stated earlier, the easiest choice for that pur¬
pose is the recognition memory model of Gillund and Shiffrin (1984).
This model represents memory in the same form as the Cl model does,
as a network of interconnected nodes. The Cl computation generates a
network that can be used as an input to the Gillund and Shiffrin memory
model. Other memory models, such as TODAM (Murdock, 1993) and
MINERVA (Hintzman, 1988), use different representations that would
require a transformation from the network representation of the Cl
model to the format appropriate for TODAM or MINERVA.
To model delay effects, we need to know what happens to the mem¬
ory trace as a function of delay. Figure 8.2 provides the necessary con¬
straints. Thus, the simulations of sentence recognition are constructed
from three components: the simulation of comprehension by the Cl
model, the simulation of recognition by the Gillund and Shiffrin model,
and the empirically determined decay functions for different levels of
sentence memory.

The effects of importance in sentence recognition. Lists of words in a recog¬

nition experiment are usually constructed in such a way that all items are
as much alike as possible. That is, relevant variables such as word fre¬
quency, length, imagery value, and the like are controlled. Any remaining
254 Models of comprehension

differences between items are random to the extent that other factors
have been successfully controlled.
Sentence recognition experiments are fundamentally different in this
respect, because the sentences in a natural text are never equal. Indeed, it
has been shown that important, central sentences in a text are recognized
better than unimportant, less central sentences (Walker & Yekovich,
1984). The Cl model allows one to calculate the strength of each sentence
in a textbase. Hence, if we use these calculations as the input to the
Gillund and Shiffrin recognition model, the model should predict cor¬
rectly that sentences high in the textbase hierarchy should be recognized
better than sentences low in the hierarchy.
A simple test of this prediction can be obtained by constructing an
arbitrary textbase and then computing how well a sentence correspond¬
ing to a proposition high in the hierarchy is recognized versus a sentence
corresponding to a lower-order proposition. Figure 8.3 shows the net¬
work I have used for that purpose. It contains seven nodes forming a
textbase, the propositions PI to P7, the corresponding surface structure
nodes SI to S7, and a situation model consisting of a general context
node, CX, and two macropropositions, Ml and M2. This is an abstract
example, in that no particular text is associated with it. Its structure is
arbitrary but designed to allow us to investigate recognition for important
and unimportant sentences, that is, sentences high or low in the textbase
hierarchy. Thus, the propositional nodes are arranged so that they occupy
three levels of a hierarchy: P3 and P6 are at the top of the hierarchy
(directly subordinated to the macropropositions), P2, P4, P5, and P7 are
at the intermediate level (subordinated either to P3 or P6), and PI is at the
lowest level (subordinated to P2). This allows us to test recognition for
sentences based on PI and P3, that is, for a low-importance and for a
high-importance sentence.
As a first step, comprehension of the structure shown in Figure 8.3 was
simulated. (At this point, the stars in the upper left corner of Figure 8.3
play no role.) The simulation was done in two cycles: The first contained
CX, Ml, and all nodes under Ml; the second cycle included CX and M2
and all the nodes dominated by M2, plus P3, which was retained in the
S PM buffer as the most strongly activated proposition from the first
cycle. Thus, a long-term memory matrix Ltext, which became the basis
for the recognition tests, was calculated. The program of Mross and Rob¬
erts (1992) was used for these calculations.
The recognition probe is represented in Figure 8.3 by the stars labeled
Memory for text 255

si

pi

cD— —GD
s/

P3 S3 |
M

Ml
-
—CD

CD"'"—-(D
/
/
/

.(D
\

CD-- --CD
Figure 8.3 A network representation of a text and a test sentence high in the
textbase hierarchy. Nodes that do not participate in the test are linked by
dashed lines.
 256 Models of comprehension

M, P, and S for its three components at the situation model (MOD),

propositional (PRO), and surface levels (SUR). When the high-impor¬
tance sentence (P3-S3) was tested for recognition, the PRO element was
connected to its counterpart, P3, and the SUR element was connected to
its counterpart, S3, as shown in Figure 8.3. The MOD element was con¬
nected to Ml, because the test sentence fits into that part of the situation
model. All other elements in Figure 8.3 were deactivated, that is, they did
not take part in the processing of the test sentence. The activated struc¬
ture was then integrated and the results added to the existing LTM
matrix, Ltext. The resulting matrix Lhigh was then used as the input to the
recognition module. -'
For the test of the low-importance Pl-Sl sentence, SUR was con¬
nected to SI, PRO to PI, and MOD as before to Ml, because that sen¬
tence, too, is part of the same portion of the situation model. All other ele¬
ments (including P3, etc.) were deactivated. Once again, the remaining
network was integrated and the results added to the original LTM values.
Thus, another LTM matrix Llow was obtained that could be used as input
to the recognition module.
The matrices Lhigh and Llow are both sparse - that is, most of the
entries are zero, because only the direct links between nodes have
strength values in the LTM matrices generated by the Cl model. The
Gillund and Shiffrin model (1984), however, requires information not
only about the direct links between items but also about indirect links.
Hence, full matrices have to be calculated from Lhigh and Llow. To illus¬
trate this problem, note that in Figure 8.3 there is no direct link between
SI and S2. However, there is a third-order connection between these two
nodes by PI and P2. The strength of this connection can be calculated by
multiplying the strengths of the path segments: r(Sl,S2) = r(Sl,Pl) *
r(Pl,P2) * r(P2,S2). In this manner, full matrices Rhigh and R|ow can be
obtained that give the strength values of all interconnections between all
nodes in the network for the high- and low-importance test sentences.1
These matrices correspond to the retrieval structures of Gillund and
Shiffrin (1984).
Because the Gillund and Shiffrin model does not care whether an ele¬
ment in memory is a proposition or a surface feature, the Ms, Ps, and Ss
in Figure 8.3 will be relabeled as memory images, Ib I2, . . . , I]6. The
retrieval cues CX, MOD, PRO, and SUR will be relabeled as 0.1,-CU-

1 I thank hrnie Mross for writing the “Matrix” program that performs these operations.
 Memory for text 257

Then the familiarity of the probe (Q_1? . . . Q_4) is given by equation 1 of

Gillund and Shiffrin as

(3)

For the high-importance sentence, the familiarity value calculated in this

way is .536; for the low-importance sentence, the familiarity value is .038.
These values are sensible, given our expectations. Transforming them
into observable statistics such as the probability of a “yes” response,
requires some further calculations, however.
Gillund and Shiffrin (1984) suggest a signal-detection model for cal¬
culating the probability of “yes” or “no” decisions based on familiarity
values. Assume that the familiarity values we have calculated are the mean
values of normally distributed random variables. A plausible source of
this variability may be subject differences. Thus, we need to obtain from
the familiarity values we have calculated d' and [3 values. For this purpose,
we arbitrarily fix the d' value of the high-importance sentence to be d'high
= 2.00 and (3hlgh = -.84, so that P(yes)high becomes .80, which is a reason¬
able value according to Figure 8.1. Assuming that the standard deviations
of the familiarity values for the high- and low-importance sentences
are the same, we can calculate d\ow = .14. Therefore, [3low — 1.02, and

p(yes)iow = -17-2
Thus, in our example, a high-importance sentence is recognized much
more often than a low-importance sentence. Of course, these values
depend on how the example has been constructed; a different and con¬
siderably higher value would be obtained had we chosen P2 as the low-
importance proposition. Yet different values would be obtained with
different networks. The point here is merely that when we compare
recognition of a high-importance sentence and a low-importance sen¬
tence, the former is correctly predicted to be higher than the latter. 1 his
prediction is obtained without ad hoc assumptions, merely by using the
Cl model together with an off-the-shelf recognition model and standard
signal detection analyses.

2 mhi h = .536 and i'high = 2 implies shigh = .536/2 = .268. For d' - 2, a b value of -.84
corresponds to a b value of 1.16 for d' = 0. For the distribution of the low-importance
sentence with a mean of .038, d' = .038/.268 — . 142. The criterion value of 1.16 coi-
responds to 1.16 - .142 = 1.02 on the low-importance distribution. Hence P(yes)low
= .172.
258 Models of comprehension

The effects of level of representation in sentence recognition. Now let us turn

to a somewhat more complex task, namely, modeling the pattern of results
seen in Figure 8.1 for recognition of old sentences, paraphrases, infer¬
ences, and distractors. Two phenomena in Figure 8.1 deserve attention.
First, old sentences, paraphrases, inferences, and unrelated distractor
sentences yield systematically different recognition probabilities. Sec¬
ond, recognition probabilities remain essentially constant over a four-day
period for old sentences but increase systematically for paraphrases,
inferences, and distractors.
To keep this illustrative example as simple as possible, a smaller net was
used than the one in Figure 8.3. This is sufficient, because there is no need
here to distinguish between high- and low-importance propositions. Fig¬
ure 8.4 shows the net chosen for this simulation. The network is similar to
the previous one, but the number of propositions and surface nodes has
been reduced to four. The test item is always related to PI, but there are
three different cases. (1) If the old sentence is tested, all three test nodes
are linked to corresponding nodes in the LTM structure. (2) If a para¬
phrase of that sentence is tested, the link between SUR and SI is deleted
because the surface structure is no longer the same. (One could make more
subtle distinctions, of course, such as decreasing the strength of the link
between SUR and SI proportionally to the amount of change in the para¬
phrase). (3) If an inference is tested, neither the SUR node nor the PRO
node is directly linked to the LTM structure because both the surface
form and the propositional content are new. Only the link between the
model component of the test sentence (MOD) and the situation model
component of the text (Ml) remains because the inference pertains to the
situation model that was constructed for this particular recognition test.
Comprehension of the network in Figure 8.4 was simulated in two
cycles, the first comprising the upper branch, dominated by Ml; the sec¬
ond the lower branch, dominated by M2. The situation model element
Ml was retained from the first to the second cycle in the STM buffer
because it was the most strongly activated node in the first processing
cycle. The CX node also participated in both processing cycles because it
is connected to both parts of the network. The resulting LTM strengths
are shown in the first column of Table 8.1.
Recognition of an old sentence (PI-SI) was simulated by adding three
test nodes to the network, MOD, PRO, and SUR, corresponding to the
situation model, proposition, and surface components of the test sentence,
as shown by the stars in Figure 8.4. These three nodes were linked to their
 Memory for text 259

off-—off-Qf)
Figure 8.4 A network representation of a text and an old test sentence.
Nodes that do not participate in the test are linked by dashed lines.

corresponding nodes in the LTM structure, as has been described. All

other nodes in the LTM structure were deactivated, and the test nodes
were integrated together with the active portion of the text. The resulting
LTM values were added to the LTM values calculated from the compie-
hension simulation, as shown in column 2 of Table 8.1. Column 3 of Table
8.1 shows the results of testing a paraphrase of Pl-Sl. For these calcula¬
tions, the node SI was deactivated. Otherwise the calculations were the
same as for the old test sentence. Finally, testing an inference was simu¬
lated, for which purpose the PI node was also deactivated. 1 hese results
are shown in the last column of Table 8.1. Note the changes in the mem¬
ory strength values for the test nodes MOD, PRO, and SUR in lable 8.1
(last three rows) for old, paraphrase, and inference test items: All three
260 Models of comprehension

Table 8.1. The LTM strength values for the network shown in Figure 8.4 and
the strength values calculated for the recognition of an old sentence, a paraphrase
sentence, and an inference sentence

LTM OLD TEST PARA TEST INF TEST

cx .712 .787 .810 .779

CX,M1 .740 .949 .987 .886
CX,M2 .669 .667 .667 .667
Ml 1.112 1.691 1.739 1.436
Ml,PI .708 1.000 1.000 .708
M1,P2 .708 .708 .708 .708
Ml,MOD .740 .792 .568
PI .501 1.253 .984 .501
PI,SI .260 .824 .260 .260
PI,PRO .867 .676
SI .135 .558 .135 .135
S1,SUR .613
P2 .501 .501 .501 .501
P2,S2 .260 .260 .260 .260
S2 .135 .135 .135 .135
M2 1.000 1.000 1.000 1.000
M2,P3 .667 .667 .667 .667
M2,P4 .667 .667 .667 .667
P3 .445 .445 .445 .445
P3,S3 .223 .223 .223 .223
P4 .445 .445 .445 .445
P4,S4 .223 .223 .223 .223
S3 .112 .112 .112 .112
S4 .112 .112 .112 .112
MOD .945 1.000 1.000
PRO 1.000 .946 .677
SUR .888 .602 .677

nodes are strong for old items, whereas for paraphrases only the MOD and
PRO nodes are high, and for inferences only the MOD node is high.
The Cl model having done its work, we turn to the recognition mod¬
ule. We first need to transform the sparse matrices obtained from the Cl
model into full matrices, showing indirect as well as direct connections
between all nodes in the network. Table 8.2 shows the sparse matrix
obtained from testing the old sentence; this is equivalent to column 2 of
Table 8.1, except presented in a different form.
 Memory for text 261

Table 8.2. The sparse matrix showing the direct links between nodes for an old test sentence

Memory Images

Cues Ml PI SI P2 S2 M2 P3 P4 S3 S4 MO PRO SUR

CX 0.95 0 0 0 0 0.67 0 0 0 0 0 0 0
Ml 1.69 1 0 0.71 0 0 0 0 0 0 0.74 0 0
PI 1 1.25 0.82 0 0 0 0 0 0 0 0 0.87 0
SI 0 0.82 0.56 0 0 0 0 0 0 0 0 0 0.61
P2 0.71 0 0 0.5 0.26 0 0 0 0 0 0 0 0
S2 0 0 0 0.26 0.14 0 0 0 0 0 0 0 0
M2 ‘0 0 0 0 0 1 0.67 0.67 0 0 0 0 0
P3 0 0 0 0 0 0.67 0.45 0 0.22 0 0 0 0
P4 0 0 0 0 0 0.67 0 0.45 0 0.22 0 0 0
S3 0 0 0 0 0 0 0.22 0 0.11 0 0 0 0
S4 0 0 0 0 0 0 0 0.22 0 0.11 0 0 0
MOD 0.74 0 0 0 0 0 0 0 0 0 0.95 0.97 0.92
PRO 0 0.87 0 0 0 0 0 0 0 0 0.97 1 0.94
SUR 0 0 0.61 0 0 0 0 0 0 0 0.92 0.94 0.89

Table 8.3 shows the corresponding full matrix, in which the strengths
of all indirect paths have been computed by multiplying together the
strengths values of all path segments connecting to nodes. Because the
network in Figure 8.4 is fully connected, all entries in this matrix are
nonzero. These entries are the strengths values with which nodes i and j
are connected in memory, S(I;,Ij).
In the Gillund and Shiffrin model (1984), the Ms, Ps, and Ss of Fig¬
ure 8.4 are the memory images, which we now label as Ik. The recognition
probe consists of the context element CX and the test sentence, which are
represented by the three elements MOD, PRO, and SUR in Figure 8.4.
The familiarity of the probe (CX, MOD, PRO, SUR) according to equa¬
tion 1 of Gillund and Shiffrin is given by
9
F(CX,MOD,PRO,SUR) = £ S(CX,Ik) * S(MOD,Ik) *
k=i
S(PRO,Ik) * S(SUR,Ik). (4)

The familiarity values calculated in this way for OLD sentences Irom
Table 8.3 is 1.378. These calculations were repeated for paraphrases and
inferences, starting with the values for these test sentences shown in Table
8.1. The familiarity value of the paraphrase test sentence turns out to be
262 Models of comprehension

Table 8.3. The full matrix showing both direct and indirect links between nodes
for an old test sentence

Memory Images

Cues Ml PI SI P2 S2 M2 P3 P4 S3 S4 MODi PRO SUR

CX 0.95 0.95 0.78 0.67 0.17 0.67 0.44 0.44 0.1 0.1 0.7 0.82 0.64
Ml 1.69 1 0.82 0.71 0.18 0.63 0.42 0.42 0.09 0.09 0.74 0.87 0.68
PI 1 1.25 0.82 0.71 0.18 0.63 0.42 0.42 0.09 0.09 0.84 0.87 0.82
SI 0.82 0.82 0.56 0.58 0.15 0.52 0.35 0.35 0.08 0.08 0.56 0.71 0.61
P2 0.71 0.71 0.58 0.5 0.26 0.45 0.3 0.3 0.07 0.07 0.52 0.61 0.48
S2 0.18 0.18 0.15 0.26 0.14 0.12 0.08 0.08 0.02 0.02 0.14 0.16 0.12
M2 0.63 0.63 0.52 0.45 0.12 1 0.67 0.67 0.15 0.15 0.47 0.55 0.43
P3 0.42 0.42 0.35 0.3 0.08 0.67 0.45 0.44 0.22 0.1 0.31 0.37 0.29
P4 0.42 0.42 0.35 0.3 0.08 0.67 0.44 0.45 0.1 0.22 0.31 0.37 0.29
S3 0.09 0.09 0.08 0.07 0.02 0.15 0.22 0.1 0.11 0.02 0.07 0.08 0.06
S4 0.09 0.09 0.08 0.07 0.02 0.15 0.1 0.22 0.02 0.11 0.07 0.08 0.06
MOD• 0.74 0.84 0.56 0.52 0.14 0.47 0.31 0.31 0.07 0.07 0.95 0.97 0.92
PRO 0.87 0.87 0.71 0.61 0.16 0.55 0.37 0.37 0.08 0.08 0.97 1 0.94
SUR 0.68 0.82 0.61 0.48 0.12 0.43 0.29 0.29 0.06 0.06 0.92 0.94 0.89

.834, and of the inference test sentence .221. These values are shown in
Table 8.4.
Familiarity values for test sentences after a delay can also be calculated
by making suitable assumptions about the decay of probe strengths as a
function of delay. These assumptions were motivated by Figure 8.2.
Specifically, it was assumed that there was no loss of strength as a func¬
tion of delay for situation model nodes in memory but that propositional
nodes lost half their strength and surface elements all their strength.
Thus, the terms in equation 2 were multiplied by .5 if the imagery ele¬
ment was a proposition and by 0 if it was a surface element. The result¬
ing familiarity values for old sentences, paraphrases, and inference test
sentences are also shown in Table 8.4.
Finally, a signal-detection model was used to calculate recognition
probabilities from the familiarity values of the test sentences. As in the
previous example, d'(M was arbitrarily set to 2, and Pold was chosen so that
P(yes)0id = -80, as suggested by Figure 8.1. The d' values for paraphrases
and inferences were obtained by scaling their familiarity values in the
 Memory for text 263

Table 8.4. Familiarity values, d'- and Rvalues and P(yes) for old
sentences, paraphrases, inferences, and unrelated distractor sentences for
immediate and delayed recognition

Immediate

Familiarity d' P P(yes)

Old 1.38 2.00 -.84 .80

Paraphrase .83 1.20 .52
Inference .22 .32 .20
Distractor 0 0 .12

Delay

Familiarity d' P P(yes)

Old .84 1.21 -.84 .80

Paraphrase .64 .93 .71
Inference .16 .23 .44
Distractor 0 0 .36

same way as for old sentences (i.e., dividing by .69). The (3 and P(yes) val¬
ues could then be calculated for paraphrases, inferences, and distractors,
as shown in the table. The calculations for the delayed recognition test
were identical, again choosing a criterion value that yields a recognition
probability of .80 for old test sentences. Figure 8.5 summarizes the pre¬
dicted recognition probabilities for both the immediate and delayed test.
These predictions capture the essential features of the empirical data of
Figure 8.1. That the “yes” probability for old sentences remains at .80 for
both immediate and delayed tests was built into Figure 8.5 by the way the
parameters of the signal detection model were estimated. What the model
predicts correctly is the ordering of the test items and the increase with
delay in false alarms for items that were not actually part of the original
text (paraphrases, inferences, and distractors).
The two examples of how well-known phenomena of sentence recog¬
nition can be derived from the Cl model and current memory theory
make an important point. We don’t need a special theory of sentence
memory: If we understand sentence comprehension (the Cl theory) and
recognition memory (the list-learning literature), we have all the parts we
need for a sentence recognition model. Conceptually it is very simple.
264 Models of comprehension

0.8

<ZJ
<0 —□— Old
C 0.6
c
G. .o-. Paraphrase
<D
Oi
—-O"— Inference
c/l
u 0.4
--A- New

0.2

Figure 8.5 The predicted probabilities of recognizing old sentences, para¬

phrases, inferences, and unrelated distractor items on an immediate and
delayed recognition test.

The actual calculations are complex, and the pattern of data that must be
accounted for is complex, too. But underlying all this complexity is a sim¬
ple model of comprehension and a simple memory model.
The examples discussed have been intentionally simplified in order to
be able to work through the actual calculations step by step. This is
important if one wants to show that a complex new phenomenon can be
accounted for by simple old theories. It is, of course, also necessary to
work with real, nonarbitrary texts, fitting actual data, and to extend the
domain of phenomena modeled - for example, by including the plausi¬
bility results mentioned in the introduction to this section. Singer and
Kintsch (in preparation) attempt to do so, thus demonstrating that the
kind of account proposed here of sentence recognition in terms of famil¬
iar theories about comprehension and memory is indeed feasible, and not
just a theoretical possibility.

8.2 Recall and summarization

To provide an account of how readers recall a text they have read has been
one of the oldest and most important applications of the Cl theory and its
predecessors (Kintsch, 1974, 1976, 1977b; Kintsch & van Dijk, 1978;
Miller & Kintsch, 1980; van Dijk & Kintsch, 1983). In general, one can
 Memory for text 265

say that the theory has done a reasonably good job of predicting recall,
especially for short texts and stories. The way these predictions were
derived was to simulate comprehension of a text and then assume that
recall probabilities for propositions are a monotonic function of the acti¬
vation values, or better, the long-term memory strength values calculated
in the simulation. Correlations between predicted recall and observed
recall frequencies typically range between .4 and .8. This is as good as or
better than intuitive judgments of what will or will not be recalled from a
text.
Exceptions have been encountered, however. In my experience, there
are two cases in which the theory regularly yields poor predictions. First,
recall for long texts cannot be predicted without an explicit account of
macrostructure formation. Brief, paragraph-long texts are recalled repro-
ductively, but longer texts are summarized. Their macrostructure may be
reproduced, but the microstructure is either deleted or reconstructed, as
Bartlett (1932) showed long ago. The solution to this problem is to include
the macrostructure explicitly in the simulation. When this is done, as in
Kintsch and van Dijk (1978), the predictions of the model can be sub¬
stantially improved.
Another reason for prediction failures of the model occurs when read¬
ers do not properly understand the text that they are recalling. The model
assumes an ideal reader, that is, it assumes that the reader has processed
the whole text and has formed a mental representation of the text as the
author intended - essentially complete and essentially correct. If the
reader misunderstands the text and makes various kinds of errors in
recall, the model simply does not apply. Such misunderstandings occur
quite frequently, especially with scientific texts. For instance, recall data
recently collected in my laboratory on a text discussing several hypothe¬
ses concerning the extinction of the Irish elk - a huge animal with giant
antlers - were beyond the scope of the comprehension theory because
most subjects egregiously misunderstood the discussion of the compet¬
ing hypotheses. They formed idiosyncratic mental text representations,
depending on the nature of their misconceptions, which is unpredictable.
Thus, some conspicuous failures of the model to predict recall data that
have been reported in the literature may very well be attributable to the
subjects’ inadequate comprehension of the text they had read.
The model does, overall, a fairly good and reliable job of predicting
story recall. Some issues, however, need to be considered and might help
266 Models of comprehension

us to improve the model predictions further. I hese concern parameter

estimations, the propositional unit of analysis, and the possibility of com¬
bining an explicit model of free recall with the comprehension theory, as
was done for recognition.
The Cl model has a large number of parameters, at least potentially.
Might not the success of the model be due the fact that the model judi¬
ciously exploits this parameter space? One often hears the complaint that
with that many degrees of freedom, any structure could predict any data
set. It is an ill-founded complaint, however. It is the structure of the
model that yields the good predictions, not the free parameters. Although
the model has many" potential parameters (conceivably, every link
strength might be estimated separately), actual simulations use this free¬
dom with restraint. Some research has begun to explore the parameter
space of the model with respect to free recall predictions (Miller &
Kintsch, 1980; Tapiero & Denhiere, 1995). Their conclusions are rather
reassuring, in that the model does not appear to be overly sensitive to the
selection of particular parameter values within some reasonable range.
Direct links may be assigned a value of 1 and all other links a value of 0;
short-term memory buffer values between 1 and 4 have been used, with
1 or 2 yielding the best results; the appropriate cycle size (i.e., the num¬
ber of atomic propositions included in a processing cycle) appears to be
around 7. Tapiero and Denhiere (1995) obtained better results when links
were defined by “minimal predication” than by strict argument overlap:
If A is an argument and P and Q^are predicates, P[ A] is linked to Q_[P[A]]
by argument overlap, but not by Tapiero and Denhiere’s criterion of
predication. Thus, not only does the Cl model use only relatively few
parameters in practice, but these remain mostly the same from one appli¬
cation to the next. Parameter estimates don’t have much to do with the
good fit of the model, at least at the qualitative level. More careful para¬
meter estimation would be required if quantitative fits were desired.
Most published research involving the Cl model and its predecessors
use atomic propositions as the unit of analysis. The simulations presented
in the following section employ complex propositions, instead. There are
several reasons to prefer complex propositions. First, the data are often
difficult to score in terms of atomic propositions. Higher scoring reliabil¬
ity can usually be obtained for complex propositions. Second, atomic
propositions may or may not be expressed in a recall protocol for a vari-
 Memory for text 267

ety of reasons, even if they are remembered. For instance, a modifier may
be omitted in a recall protocol, not because it was not encoded as part of
the text representation, but because it was considered redundant in the
production phase. Such editing processes introduce further and spurious
variability into data scored in terms of atomic propositions. Thus, com¬
plex propositions appear to be better units for analyzing recall data and
for comparing predictions with data.
In the previous section I showed how memory strength estimates
derived from the Cl model of comprehension can predict sentence recog¬
nition data when plugged into a standard model of recognition memory.
In contrast, no recall theory was used in deriving recall predictions in the
studies of story recall; it was simply assumed that recall was proportional
to the activation values calculated in the comprehension simulation.
Although this simplifying assumption yielded reasonable predictions, it
is certainly as questionable as in the case of sentence recognition. Why
should activation values predict frequency of recall? In theories of list
recall, we do not assume that memory strengths directly predict recall;
instead, it has been necessary to introduce complex processing assump¬
tions in order to describe the experimental phenomena typically observed
in list recall experiments. Thus, the SAM theory of Gillund and Shiffrin
(1984) assumes recall to be a complex process of cue-based retrieval and
recovery. Whatever cues are present in short-term memory (for example,
a context cue plus an already recalled list item) serve as retrieval cues. The
retrieval process then may or may not produce implicitly another as yet
not recalled item from the list, depending on the strength of the relation
between the retrieval cue and the items of the list. This implicitly
retrieved item may then be recovered with a probability depending on its
memory strength. This process is formulated mathematically, and pre¬
dictions derived in this way indeed mimic all the major features of list
recall data that have been observed in the laboratory. Why should text
recall be different? Should not the same model of recall that has proven
its worth elsewhere also be used to derive predictions for text recall? The
current practice in predicting story recall from comprehension simula¬
tions assumes without any justification that recall is directly determined
by memory strength. Can we enhance our predictions based on the Cl
model by using a proven memory theory to model the recall process, as
we did for the recognition process in the previous section?
268 Models of comprehension

8.2.1 Simulations of story recall

To address this question, I simulate the comprehension of two stories and

then derive recall predictions from these simulations in the traditional
oversimplified way (that is, by assuming that the long-term memory
strength values in the Cl simulations are proportional to recall fre¬
quency), as well as in ways suggested by memory theory. Specifically, I
adapt the Gillund and Shiffrin (1984) theory of list recall to story recall.
The stories are two Aesop fables previously used by R. M. Golden (per¬
sonal communication). The first text is shown in Table 8.5.
The propositionalized text is shown in Table 8.6. There are thirteen
complex propositions, each corresponding to a sentence. Links among
complex propositions are shown in Figure 8.6. Repeated arguments are
indicated by first letters in bold face type. Note that verbs of psychologi¬
cal state and verbs of saying and thinking are treated as modifiers. Thus,
for The frogs decided to find a new home the core proposition is FIND
[FROGS,HOME], which has the modifiers DECIDE[FROGS, FIND
[. . .]] and NEW[HOME], as shown in Table 8.6 in a abbreviated form.
Similarly, The secondfrog replied: Suppose the well’s bottom is not a good home
has the core proposition NOT-GOOD[BOTTOM,HOME] with the
modifiers REPLY[SECOND-FROG,NOT-GOOD[. . .]], SUPPOSE
[NOT-GOOD[. . .]], and OF[BOTTOM,WELL].
For the simulations, a buffer size of 1 atomic proposition was used.
Cycle size was varied systematically between 4 and 8 atomic propositions.
Thus, for a cycle size of 4 the first processing cycle would contain only

Table 8.5. The Frog story: Two Frogs and the Well

Two frogs dwelt in the same pool of water.

The pool dried up in the summer heat.
The frogs decided to find a new home.
The two frogs left the pool of water.
The two frogs began their search for another pool of water.
As the two frogs traveled along, they reached a well.
The well was very deep.
The first frog said the well would provide them with shelter and food.
The first frog wanted to enter the well.
The second frog wanted to think about the well as a new home.
The second frog replied: “Suppose the well’s bottom is not a good home.”
The second frog said, “How will we get out of the deep well in that case?”
Table 8.6. Propositional analysis of the Frog story

PI Two frogs and the well.

AND[FROGS,WELL]
■—- TWO
P2 Two frogs dwelt in the same pool of water.
DWELL[F, POOL]
|— SAME
OF-WATER
P3 The pool dried up in the summer heat.
DRY [P]

t SUMMER
IN-HEAT
P4 The frogs decided to find a new home.
FIND [F, HOME]]
L NEW
DECIDE
P5 The tw o frogs left the pool of water.
LEAVE [F, P]
P6 The tw'o frogs began their search for another pool of water.
SEARCH[F, POOL]]
L OF-WATER
BEGIN
P7 As the two frogs traveled along, they reached a well.
REACH [F, W]
I- AS-TRAVEL
P8 The well was very deep.
DEEP [W]
I_ VERY
P9 The first frog said the well would provide them with shelter and food.
PROVIDE[W,SHELTER,FOOD]
I_ SAID [FIRST-FROG]
P10 The first frog wanted to enter the well.
ENTER[F, W]
I_WANT[FF]
PI 1 The second frog wanted to think about the well as a new home.
THINK [F, W, H]
WANT[SEC-FROG]
P12 The second frog replied:
“Suppose the well’s bottom is not a
good home.”
NOT-GOOD[ BOTTOM,H]
I- SUPPOSE 1—OF-W
P13 The second frog said.
“How will we get out of the deep well in
that case?”
HOW-GET-OUT[F, W]
I_ IF-N-G
- SAID[SF]
270 Models of comprehension

the first two complex propositions (which actually contain 5 atomic

propositions, because the second complex proposition consists of 3
atomic propositions that cannot be split apart). For a cycle size of 7, the
first three complex propositions would be included in the first cycle, by
the same argument.
Repeated arguments were introduced as separate nodes in the network,
Thus, a node FROGS was formed that was connected to all propositions
with that argument, (FROGS or F in Table 8.6). In this way, repeated
arguments were treated as separate discourse entities.
All links were assigned a value of 1. Links were constructed between a
proposition and its modifiers and between a discourse entity node and the
propositions containing that discourse entity. These links are shown in
Figure 8.6.
Trabasso and van den Broek (1985) and others have shown that causal
links in a story are more important than other types of links. Hence, a sec-

POOL

HOME

FIRST-FROG

P12-NOT-GOOD/ SECOND-FROG

PI 3-GET-OUT

Figure 8.6 The argument-repetition network constructed for the proposi¬

tion list in Table 8.6. Only predicates and repeated arguments are shown;
modifiers and the links among them have been omitted for clarity.
 Memory for text 271

ond coherence network was constructed in addition to Figure 8.6 that

included causal links in addition to the links established by mere argu¬
ment overlap. Whenever a causal relation was found between two com¬
plex propositions (e.g., between P3 and P4 or between P5 and P6),
another link was added to the argument overlap network. Comprehen¬
sion was simulated for this causal net with an input size of 7 atomic
propositions only.
Tables 8.7 and 8.8 show the text and proposition list for a second story,
the Miser story. There are seventeen complex propositions, each corre¬
sponding to a sentence. Links among complex propositions are not
shown. Repeated arguments are indicated by first letters in boldface type.
Thirty-two college students were asked to read and recall both the Frog
and the Miser stories. Figure 8.7 shows the correlations between the acti¬
vation values of atomic propositions and the observed recall frequencies
as a function of cycle size. It is reassuring that these correlations are not
strongly affected by cycle size. The highest correlations for the two sto¬
ries were obtained for cycle sizes of 6 and 7. Adding causal links did not
help; indeed, it significantly decreased the correlation with the data in the
case of the Miser story. Note the relative robustness of these results with
respect to input size. Input size is the most important parameter in this
model (more than, say, link strength or buffer size), but it is reassuring to

Table 8.7. The Miser story. The Miser, the Neighbor, and the Gold

A miser bought a lump of gold using all his money.

The miser buried the gold in the ground.
The miser looked at the buried gold each day.
One of the miser’s servants discovered the buried gold.
The servant wanted the miser’s gold.
The servant stole the gold.
The miser, on his next visit, found the hole empty
The miser was very upset.
The miser pulled his hair.
The neighbor wanted to make the miser feel better.
The neighbor wanted the miser to realize a stone was as useless as the gold.
The neighbor told the miser not to be upset.
The neighbor said, “Go and take a stone and bury it in the hole.”
The neighbor said, “And imagine that the gold is still lying there.”
The neighbor said, “The stone will be as useful to you as the gold.”
The neighbor said, “When you had the gold, you never used it.”
Table 8.8. Propositional analysis of the Miser story

PI The miser, the neighbor, and the gold.

AND [MISER,NEIGHBOR,GOLD]
P2 A miser bought a lump of gold using all his money.
BUY [M, G,USE[MONEY]]
Lall
P3 The miser buried the gold in the ground.
BURY [M, G,GROUND]
P4 The miser looked at the buried gold each day.
LOOK[M,G]
L EACH-DAY
P5 One of the miser’s servants discovered the buried gold.
DISCOVER [SERVANT, G]
Lof-m
P6 The servant wanted the miser’s gold.
WANT [S, G]
P7 The servant stole the gold.
STEAL [S, G]
P8 The miser, on his next visit, found the hole empty.
FIND [M,HOLE]
L_ EMPTY
1-NEXT-VISIT
P9 The miser was very upset.
UPSET [M]
P10 The miser pulled his hair.
PULL [M,HAIR]
PH The neighbor wanted to make the miser feel better
MAKE[N, FEEL-BETTER [M]]
1— WANT[N]
P12 The neighbor wanted the miser to realize a stone was as useless as the gold.
REALIZE [M, AS [USELESS [STONE],USELESS [G]]]
L_ WANT[N]
P13 The neighbor told the miser not to be upset.
NOT-UPSET [M]
L_ SAY[N]
P14 4 he neighbor said, “Go and take a stone and bury it in the hole.”
BURY [M, ST, IN-H]
I— SAY[N]
P15 The neighbor said, “And imagine that the gold is still lying there.”
IMAGINE [M, G, IN-H]
L SAY [N]
P16 The neighbor said: “The stone will be as useful to you as the gold.”
AS [USEFUL[ST,TO-M],USEFUL[G,TO-M]]
I— SAY [N]
PI^ The neighboi said When you had the gold, you never used it ”
NOT-USE [M, G]
L WHEN [HAVE [M, G]
SAY [N]
 Memory for text 273

Figure 8.7 The correlations between observed recall frequencies and model
predictions for two stories. The connected lines show the correlations with
models based on argument repetition only (Arg) or argument repetition plus
causal links (Caus) for input cycle sizes between 4 and 8 atomic propositions.
The correlations between predictions and recall with complex propositions as
units, with links based on argument repetition and an input size of 7, are also
shown.

see that not even that matters very much. It is the structure of the model
that yields the predictions, not the specific parameter values chosen.
However, an input size of 7 appears to be best for these two stories.
Figure 8.7 also shows the correlations for complex propositions, for
cycle size 7 only. These were obtained by adding together the activation
values of all atomic propositions as well as the corresponding recall fre¬
quencies. The activation values for complex propositions predict recall
much better than do the activation values for atomic propositions and
indeed, account for over half of the variance in the recall data.
Although correlating activation values for complex propositions with
recall data gives good predictions, it remains to be investigated how much
better can we do if we use an established memory model for free recall,
such as the Gillund and Shiffrin (1984) theory. In the present case there
is probably not much to be gained because, for these very simple stories,
activation values predict recall quite well - the reliability of the data is
274 Models of comprehension

probably not much better than the predictions! Nevertheless, it is worth

investigating how better-motivated theoretical recall predictions could be
derived from the Gillund and Shiffrin model.
Unlike in the case of sentence recognition, however, we cannot simply
take over Gillund and Shiffrin’s recall model wholesale and apply it to
story recall. Much of that model is designed to deal with features of list
recall that are absent or different in story recall. When subjects recall a
list of words, they exhaust their current retrieval cue and then return to
the context cue to construct a new retrieval cue, using that as long as it
is productive. Thus, a recall episode typically involves many new starts,
resulting in bursts of, recall separated by pauses during which the
retrieval cue is being reconstructed. Story recall is quite different: It is
ordered, backtracking is rare, and the bursts characteristic of list recall
are absent because each recalled story proposition serves as a retrieval
cue for a later one. Thus, lists are recalled like semantic categories,
whereas stories are recalled like scripts (Walker & Kintsch, 1985; see also
the discussion in section 3.2.2). The control strategies for list recall and
story recall are therefore quite different, and the portions of the Gillund
and Shiffrin model that deal with these aspects do not apply to story
recall.
It is informative, however, to calculate the predictions of the Gillund
and Shiffrin model for story recall in spite of these reservations. To do so,
we must first compute the full matrix of links between propositions. The
Cl model generates only direct links (as in Table 8.2), but what we need
are the strengths between all interitem links, direct links as well as indi¬
rect links (as in Table 8.3). For instance, although proposition 3 DRY
[POND], and proposition 4, FIND[FROGS,HOME], in the Frog story
are not directly connected (Figure 8.6), there is an indirect connection via
POOL, DWELL, and FROG. The strength of that indirect connection
can be computed as the product of the strength values of the links that
form the path between proposition 3 and proposition 4.3
Recall that predictions from the Gillund and Shiffrin model were
obtained through a Monte Carlo simulation, using the full link matrices
and the same parameter values that Gillund and Shiffrin employed in
their study of list recall.
These computations are performed on the matrix of atomic proposi¬
tions (37 for Frog and 42 for Miser). Link strengths for complex proposi-

3 However, path strengths are never allowed to be greater than 1.

 Memory for text 275

tions (13 for Frog, 17 for Miser) can be calculated by summing the
strength values of all atomic propositions that belong to the same com¬
plex proposition. Note that the self strengths thus calculated are not just
the sum of the self-activation values but include also the links between
the atomic propositions that make up the complex proposition.
Table 8.9 shows the correlations between predictions and recall data
for complex propositions when the activation values obtained from the Cl
simulation of comprehension were used directly as predictors and when
these values were used in conjunction with the Gillund and Shiffrin
model, as it has been described. Both sets of predictions are about equally
good irt predicting the overall frequency with which complex text propo¬
sitions were recalled. Neither predicts the order of recall: The Gillund
and Shiffrin model makes order predictions, but, as has been discussed,
the list-learning control structure is not suited for story recall. Thus,
there are numerous jumps from one part of the story to another, forward
and backward, in the Monte Carlo simulations, that have no counterpart
in the data.
Wolfe and Kintsch (in preparation) have explored modifications of the
control structure of the Gillund and Shiffrin (1984) model that would
allow the model to predict the order in which stories are recalled. The
Gillund and Shiffrin model of recall is a two-stage model. In the first stage,
memory nodes (text propositions in our case) are implicitly reti leved with
probabilities proportional to their relative memory strength; in the second
stage, an attempt is made to recover these implicitly retrieved memory
items, the success of which depends on the absolute memory strength of
the item. In our simulations, this recovery process played no role, because
absolute memory strength was so high that recovery was almost assured.
The recovery process may be likened to a recognition judgment of implic¬
itly retrieved items: Does the item that was retrieved come from the right

Table 8.9. Correlations between observed and predicted recall

Frog Miser

Activation-recall .77 .72

G&S-recall .63 .73

G&S(order)-recall .66 .68

276 Models of comprehension

list? With stories, such a decision seems straightforward, so the recovery

process plays no important role. But suppose we assign the recovery
process a more useful control function in story recall: Propositions out of
order, or propositions far away in the story may have a reduced recovery
probability. It would be unreasonable to assume that readers sequentially
number the propositions of a story during comprehension and then use
this information in the recovery process. However, we know that readers
categorize story propositions in terms of story schema categories such as
setting, problem, and resolution. In Wolfe and Kintsch (in preparation),
we manipulated recovery probabilities in such a way that (1) transitions to
a distant category (e.g.^ from setting to resolution) were prohibited, (2)
transitions to the next category (e.g., from setting to problem) had a much
lower recovery probability than within-category retrievals, and (3) even
within a category, forward transitions had a higher recovery probability
than backward transitions (such decisions can be made on the basis of local
cues - e.g., if a subject has already recalled that the servant stole the gold,
he or she is unlikely to say next that the servant discovered the gold). This
control structure was devised simply to reflect the empirical data on recall
order without making psychologically implausible assumptions. It did
what it was supposed to. With this modification, the Gillund and Shiffrin
model predicts recall order quite well (less than one within-category
reversal per protocol and none between categories in both data and simu¬
lations), and the overall recall frequencies are still predicted as well as
before (the last line in Table 8.9).
Figures 8.8 and 8.9 show the detailed correspondence between
observed and predicted values (total activation) for atomic propositions
for the Frog and Miser texts, respectively. The predictions for Frog
roughly correspond to the trends in the data. For Miser, there is one
exception, when a proposition that was recalled very well by the subjects
did not get a particularly high activation value in the simulation (The
miser found the hole empty.).
Overall, therefore, the recall predictions of the model are quite good.
Furthermore, it mattered little whether a formal theory of recall -
Gillund and Shiffrin (1984) — was used to derive the predictions, or
whether the activation values of propositions, as calculated by the Cl
model directly, were used to predict recall. This may not always be the
case, however, because the stories used here were extremely simple. With
longer, structurally complex texts, the power of the memory module may
be required for adequate recall predictions.
 Memory for text 111

Figure 8.8 Observed and predicted recall frequencies for the Frog story.

Figure 8.9 Observed and predicted recall frequencies for the Miser story.

8.2.2 Evaluation of summaries

The most straightforward way to derive predictions from the model for
summarization would be to assume that subjects employ the same mem¬
ory representation as for recall but that their output criterion is higher.
Thus, the implicit retrieval process would remain the same as has been
described for recall, except that when a story proposition is retrieved,
278 Models of comprehension

recovery is attempted only if its self-strength is above a threshold value.

This threshold value may vary, depending on whether the subject is try¬
ing to construct a high-level summary or a more detailed one. The two
stories used here are not suitable for the empirical investigation of sum¬
marization because of their extreme simplicity, but an early version of
such a model has been used successfully for a longer story by Kintsch
(1976).
Summaries generated in this way are strictly reproductive, however.
That is, they are based on only one of van Dijk’s macrorules, the selec¬
tion of important propositions (van Dijk, 1980). The other two macro¬
rules - generalization and construction - are just as important, however,
and summaries are usually generated by applying all three rules. As was
discussed in section 6.2, the theory therefore can explain summaries after
the fact, but it is not easy to predict what a summary will look like,
because we cannot anticipate when the generalization and construction
operations will be used.
What we can do is to analyze the summaries people have produced to
determine whether they can in fact be accounted for by the operations of
selection, generalization, and construction. However, until these opera¬
tions are precisely formulated in a simulation model, such judgments are
difficult to validate. A better way to evaluate summaries, as was suggested
in section 6.2.3, is to employ LSA to measure how close a particular sum¬
mary is to the text it was derived from.
1 his was done in a study by Nathan Chatfield at the University of Col¬
orado. 7 his study had two parts. In one condition, the participants, 33
undergraduate students, read 15 unrelated paragraphs selected from a
wide variety of texts, essays as well as stories, and underlined what they
judged to be the most important sentence in each paragraph. They then
wrote summaries for three of the paragraphs. Their instructions were
simply to write a one-sentence summary for each paragraph. In the sec¬
ond part of the experiment, the participants read 10 three-paragraph
texts and rank-ordered the three paragraphs of each text in the order of
their importance “to the overall meaning of the passage.” Again, a variety
of texts was used, varying in length between a page and half of a page.
Finally, the subjects summarized three of these texts in no more than two
sentences. All summaries, for the passages as well as for the single para¬
graphs, were graded by two judges on a scale from A to F. The judges
were instructed to use all five grades approximately equally often.
Consider the rating results for the single paragraphs first. The empir-
 Memory for text 279

ical data consist of the frequencies with which each sentence of a para¬
graph was chosen as the most important one. To see whether this pattern
of choices agrees with LSA, all paragraphs and their sentences were
expressed as vectors in the LSA space obtained from the encyclopedia
scaling (see section 3.3). The cosines were computed between a para¬
graph and each sentence in the paragraph. These cosines indicate the
strength of the semantic relationship between a sentence and the para¬
graph. More important sentences should be closer to the overall meaning
of the paragraph than less important sentences. Hence, these cosines
should predict the frequency with which readers choose each sentence as
most important in a paragraph. 1 he overall correlation between the fre¬
quency judgments and the cosines was r = .51, p < .01. The split-hall
reliability of the frequency data is r = .75. Thus, the LSA predictions are
not quite as good as the agreement among subjects but are reasonably
close to human performance.
The importance ranking data for the three-paragraph passages were
analyzed in the same way: The mean empirical rank for each paragraph
was compared with the cosine between the paragraph and the whole pas¬
sage. The overall correlation was r = .58, p < .01. To obtain an idea of how
high a correlation might be expected, a split-half reliability coefficient
was computed for the empirical rankings. This correlation was r - .80. As
was the case for sentence judgments in paragraphs, LSA’s ability to pre¬
dict the importance ranking of paragraphs in a text is not quite as good as
human judgments. This may indicate a weakness of LSA (almost cer¬
tainly the fact that LSA was trained on an encyclopedia limits its perfor¬
mance in the present case, where commonsense knowledge is more rele¬
vant than the kind of formal knowledge that is usually found in an
encyclopedia), or it may indicate that semantic relatedness is not the only
factor in importance judgments. Further research is needed to clarify this
question.
The summaries for the single paragraphs as well as the whole texts
were evaluated by computing the cosine for each summary with the text
it was supposed to summarize. The better a summary, the closer its vec¬
tor in the LSA space should be to the text it summarizes. Not surpris¬
ingly, there were substantial differences between paragraphs. The first
paragraph was summarized quite well (the average cosine between all
summaries and the paragraph was .57), the second paragraph was sum¬
marized quite poorly (average cosine .28), and third paragraph was inter¬
mediate (average cosine .42). The overall correlation between the cosines
280 Models of comprehension

and the average grade assigned to a summary was r = .57, p < .01. The
correlation between the two human graders was only r = .48, however.
Thus, LSA did even a little better than the two human graders in this
case.
Similar results were obtained for the passage summaries. The differ¬
ences between texts were less pronounced (the average cosines for the
three texts were .36, .43, and .38, respectively). The overall correlation
between the average grade and cosines was r = .48, p < .01, about the
same as the correlation between the two human graders, r = .45.
Although these results need to be extended with more texts, systematic
variation between text types, summaries of different length, and human
graders who achieve a higher level of agreement, they suggest that LSA
is quite successful as an automatic method for evaluating the quality of
summaries, thus opening up a new research area as a well as a potentially
fruitful field of application.

The details of the simulations described in this chapter should

not obfuscate its message. A model of text comprehension does
not by itself provide an account of all behaviors that have some¬
thing to do with text understanding. The model takes care of
only the input side - the transformation of the text into a mental
representation. We need a separate, additional model to account
for any behavior based on this mental representation, whether
this is a lexical decision response, a comprehensibility judgment,
a recognition response, the free recall of a whole text, solving a
mathematical problem, or an action, as discussed in later chap¬
ters. In some cases these behavioral output models are trivial and
the theorist gets away with assumptions like “behavior is propor¬
tional to activation strength,” or the like. Even then one must not
forget that this is merely a shortcut, a simplifying assumption
that happens to be adequate in some cases. In general, an explicit
account of the behavioral module is required.
In the two applications discussed here, this explicit account of
the behavioral module is simple conceptually but complex to
implement. At the rational level of theorizing, it is enough to say
that the compiehension module feeds into a standard recogni¬
tion or free-recall module. At the implementational level, how¬
ever, fairly extensive calculations are required. Perhaps one
should not be surprised at the complexity of these computations.
Memory for text 281

The phenomena we are dealing with here are indeed complex,

and to get from simple principles to complex phenomena by
extensive but mechanical straightforward computations may not
be too serious a price to pay.
Using a standard theory for recognition memory and recall,
which was developed for quite different purposes and on the
basis of totally different empirical data, in combination with the
Cl theory of text comprehension is also a way of working toward
a unified theory of cognition. The unified theory does not have
to come from a single overarching theoretical scheme, but may
also be constructed from the bottom up by building on prior
established theories.
 9

Learning from text

9.1 What makes for a good reader?

The answer to that simple question is surprisingly complicated.

Reading is a complex activity, with several factors that can com¬
pensate for one another to a considerable extent. Thus, many
highly intelligent persons and fluent readers may be poor
decoders, and a person who knows a great deal about a particular
domain but has low reading skills can outperform a highly skilled
reader under the right set of circumstances.

This chapter discusses three factors that determine whether a person

is a good reader and what kind of a reader the person is: decoding skills,
language skills, and domain knowledge. All three are important for read¬
ing but in different ways.

9.1.1 Decoding skills

For a large sector of the reading research and instruction community, it

often seems as if decoding and reading were one and the same. Decoding
is certainly a very important component of the reading process, but there
is obviously more to reading than that.
Good readers are good decoders (Perfetti, 1985). They recognize
words almost twice as fast as poor readers do. Graesser, Hoffman, and
Clark (1980), for instance, have estimated the reading time per word for
poor adult readers to range from 198 to 227 ms, depending on the text,
compared to 114 to 135 ms for good readers. Bisanz, Das, Varnhagen, and
 Learning from text 283

Henderson (1992) obtained estimates ranging from 355 ms per word for
poor fifth-grade readers to 170 ms for good seventh-grade readers. Rapid
decoding is important because better word recognition frees up resources
for higher-level processing (Perfetti’s verbal efficiency hypothesis). Thus,
better decoders build more accurate and complete representations of text
content. Their decoding capabilities feed the growth of vocabulary, they
know more word meanings, and they tend to have richer representations
of individual word meanings.
However, slow decoders can use higher-order processes to compensate
for their lack of decoding skills. Stanovich (1980) has shown that poor
readers can use the sentence context to speed up their word recognition.
Good readers, on the other hand, can recognize words rapidly and accu¬
rately either within a sentence context or without context; although con¬
text may speed their processing, they do not depend on it. Many studies
investigating this phenomenon are summarized and discussed in Perfetti
(1985, pp. 143ff): vocalization latency to words in context and in isola¬
tion, word identification latencies, predicting next word, and degrading
of visual stimuli. A number of other studies are in agreement with these
findings. Good readers, for example, fixate every word in a sentence,
whereas poor readers show more irregular fixation patterns (Just & Car¬
penter, 1980). Therefore, good readers detect misspellings better than
poor readers (McConkie & Zola, 1981). Note that poor readers, accord¬
ing to Perfetti, also are less able to exploit orthographic patterns in decod¬
ing when processing demands are high, even though they apparently do
not lack knowledge of such structures.
Good decoding skills make good readers less dependent on discourse
context in order to recognize a word, but certainly they are able to use
contextual information when they need to, and furthermore they do so
more effectively than poor readers. Frederiksen (1981) provided com¬
pelling evidence of this ability in comparing good and poor readers on a
word recognition task. The target word was shown either without context
or as the last word of a sentence. When the sentence made the target word
highly predictable, both good and poor readers saved 125 ms in compar¬
ison to the context-free presentation. However, when the sentence con¬
text was only mildly constraining, poor readers could not take advantage
of such a context, whereas good readers still showed a saving. Thus, good
readers are not only better decoders but also better top-down processors.
284 Models of comprehension

9.1.2 Language skills

Most of the variance in reading times is taken up by the number of words,

which is unfortunate because this factor overshadows everything else. It
is not surprising that this is so because word recognition is executed more
frequently in reading a sentence than anything else, and hence decoding
must play a dominant role. However, other skills are involved in text com¬
prehension in additioa to word decoding, and they have been studied
experimentally. These skills are those involved in constructing the kind of
mental representations discussed in this book: figuring out the proposi¬
tional elements of a text and organizing them into a coherent structure,
including a macrostructure.
Lexical factors are obviously relevant. Just and Carpenter (1980) have
shown that word frequency is related to reading speed in adults. Indeed,
the very process of meaning construction for a word may be different in
good and poor readers. Long, Oppy, and Seely (1994) observed that for
both good and poor readers 300 ms after the presentation of a homonym
lexical decision responses to context-appropriate associates of the homo¬
nym were faster than lexical decision responses to context-inappropriate
associates. Thus, both good and poor readers were sensitive to the context
of a word at this point. However, using a somewhat different procedure
Gernsbacher and Faust (1991) and Gernsbacher, Varner, and Faust (1990)
suggested that there were nevertheless important differences between
good and poor readers in the way word meanings are constructed. Gerns¬
bacher and her co-workers gave subjects sentences to read such as He dug
with the spade versus He dug with the shovel, followed by a test word, such
as ace. Subjects were asked to judge whether the test word matched the
meaning of the sentence. Significant interference was observed for the
test word right after reading the sentence by both good and bad readers.
That is, both good and bad readers took longer to reject ace after reading
the spade sentence than after reading the shovel sentence. After a 850-ms
delay, poor readers still were slower to reject the test word ace, whereas
there was no interference for good readers. Gernsbacher et al. (1990)
argue that once the contextual meaning of a homonym is established, it
suppresses the inappropriate meaning in good readers but not in poor
readers. In terms of the GI model, this might mean that the integration
process reduces the activation of contextually inappropriate associates of
a homonym essentially to zero for good readers, whereas context only
 Learning from text 285

weakens (the Long et al. results) but does not fully suppress (the Gerns-
bacher et al. results) such associations for less skilled readers. Further
research is clearly needed to clarify these issues.
Efforts to determine the role of syntactic complexity in reading have
not yielded uniform results. Graesser et al. (1980; also Haberlandt &
Graesser, 1989) have used as a measure of syntactic complexity an aug¬
mented transition network analysis but without notable success. Simpler
measures, such as the number of main and subordinate clauses per sen¬
tence, tend to be redundant with propositional analyses.
The number of propositions per sentence is indeed related to reading
speed and recall. In an early study, Kintsch and Keenan (1973) observed
an increase in reading times of 1 sec per proposition recalled. However,
the effects of other potentially important variables were not partialled out
in this study. When this is done, estimates of the additional reading time
per proposition range from 191 to 238 ms for poor readers and 75 to 122
ms for good readers in the Graesser et al. (1980) experiment. Estimates
obtained by Bisanz et al. (1992) for younger readers are consistent with
these values: 166 ms per proposition for fifth-grade poor readers, down
to 94 ms for seventh-grade good readers.
Of the skills that are important for the organization of the textbase,
especially its macrostructure, the ones that have been studied most thor¬
oughly are the role of new propositional arguments in the coherence
structure and the reader’s use of causal relations in narratives. Kintsch et
al. (1975) found longer reading times and poorer recall for paragraphs
containing many new arguments than for paragraphs containing few new
arguments. For short history and science paragraphs, subjects required
on the average 1.6 s of reading time per proposition recalled when the
paragraphs contained few new arguments, versus 2.5 s reading per
proposition recalled when the paragraphs contained many new argu¬
ments. The overall level of recall was 75% for the paragraphs with few
new arguments and 58% for those with many new arguments. In other
words, subjects read at a fairly constant rate but recalled more from the
paragraphs with few new arguments than from the paragraphs with many
new arguments. Haberlandt et al. (1986) and others obtained similar
results for reading times, and Bisanz et al. (1992) for recall.
A number of studies have pointed out the important role that causal
links play in a text. For instance, Trabasso and van den Broek (1985) have
shown that for both adults and children recall of a text unit is a function
286 Models of comprehension

of the number of causal links connecting it with other text units. If sto¬
ries are analyzed into units that fall on the primary causal chain of states,
events, and actions that make up the story and dead-end units that are not
located on this causal chain, units on the causal chain are recalled better
than units off the chain (e.g., Omanson, 1982; Trabasso & van den Broek,
1985).
Bisanz et al. (1992) have shown that poor readers (fifth- and seventh-
grade pupils) employ their causal knowledge in a compensatory manner.
The more causal relations there were in a text, the faster poor readers
could read it. In contrast, the number of causal relations had no effect on
reading times for good readers. Thus, we have here another instance
when readers are using a top-down process (causal inferences) to com¬
pensate for a weakness in other components of the reading process (they
are slow at decoding, require more time to form propositions).
Bisanz et al. (1992) also make the interesting observation that different
skills are related to reading times and recall. Specifically, reading time is
primarily a function of the number of words and the number of proposi¬
tions in a sentence, whereas recall is a function of the causal links in a text
and the number of new arguments, both of which are important for the
overall organization of the textbase.
Topic inferences by skilled and less skilled readers have also been inves¬
tigated in the study by Long et al. (1994) already cited. Their experiment
used a design like that of Till et al. (1988): At SOA intervals between 200
and 1,000 ms after reading a sentence, lexical decision responses were
made to appropriate and inappropriate topic words. Response times to
appropriate topic words were significantly faster than to inappropriate
topic words at the 500 ms SOA for good readers, and an even larger dif¬
ference was observed at the 750 ms SOA, but no differences in the lexical
decision times for appropriate and inappropriate topic words were
observed for poor readers, even after 1 s. It was not the case that poor read¬
ers could not infer the topic of these sentences, however, because when
asked to provide topics for these sentences they performed just as well as
skdled readers. Fhus, they had the knowledge, but what was missing was
the automatic retrieval structure necessary to make available this knowl¬
edge during normal reading. Referring back to Table 6.1, for good read¬
ers, the topic inference belonged to the inference category in cell A of that
table (automatic retrieval of bridging information), whereas it fell into cell
B (controlled search for bridging information) for poor readers. Or,
 Learning from text 287

expressed as a tautology, we might say that skilled readers were experts in

this task and less skilled readers were not.

9.1.3 Domain knowledge

In addition to decoding and language skills, domain knowledge is the

third factor that determines whether someone is a good reader. Readers
with high domain knowledge tend to understand texts better and remem¬
ber them better than readers with low domain knowledge. Domain
knowledge may even compensate for other factors, such as low IQ^, low
verbal ability, or low reading ability. Recht and Leslie (1988) and Walker
(1987) performed factorial experiments in which reading ability and
domain knowledge — knowledge about baseball — were either high or low.
Their subjects read a description of a baseball game and then answered
memory and comprehension questions. Similarly, Schneider, Korkel,
and Weinert (1989) designed a study comparing high and low IQ_against
high or low knowledge of soccer. In every case, readers with high domain
knowledge outperformed readers with low domain knowledge, irrespec¬
tive of their general ability. Thus, low IQ_or low verbal ability could be
fully compensated for by domain knowledge.
Three factors were varied in an experiment by Moravcsik and Kintsch
(1993): (1) domain knowledge (high or low), (2) the way the text was writ¬
ten (good writing, well organized, vs. poor writing, disorganized), and (3)
reading ability (high or low scores on the comprehension subtest of the
Nelson-Denny Reading Test). Domain knowledge was manipulated in a
different way in this study. Subjects read texts, such as the washing-
machine paragraph from Bransford and Johnson (1972), that could easily
be understood when given a title. The title allowed subjects to use their
knowledge and to disambiguate the otherwise obscure text. Without the
title, the texts were difficult to understand, because subjects could not
interpret the situation as a familiar one. Hence, subjects could use their
knowledge in the title condition but not in the no-title condition. 1 he
texts were either well written or poorly written. In the well-written ver¬
sions, syntactic cues as well as the overall organization of the text allowed
the reader to identify the significant elements of a text and their relation¬
ship to each other, even without knowing what was really going on. I he
poorly written texts provided no such cues to facilitate the construction
of a textbase.
 288 Models of comprehension

All three factors - knowledge, writing quality, and reading ability -

significantly influenced the amount of reproductive recall (Figure 9.1).
These effects were additive. There was no indication of an interaction
and hence of compensation (see also Voss & Sillies, 1996). However, there
was an interesting difference in the kind of mental representation sub¬
jects constructed in the different experimental conditions. Even though
high skill and good writing enabled low-knowledge readers to form ade¬
quate textbases that were capable of supporting reproductive recall, these
readers could not form correct situation models to support their elabora-
tive recall.1 Their elaborations tended to be wrong and fanciful. Only
high-knowledge readers were capable of good elaborations. Figure 9.2
shows that inadequate situation models did not keep the subjects from
elaborating but that their elaborations and inferences were erroneous,
whereas the elaborations of readers who could use their knowledge to
construct an adequate situation model were appropriate.
It is interesting to note the effects of writing quality in this study. All
passages were written in two versions, preserving their content but vary-

Knowledge Skill Writing

Figure 9.1 Reproductive recall: The mean number of propositions recalled

as a function of knowledge, reading skill, and writing style. After Moravcsik
and Kintsch (1993).

1 Elaborativ e recall is that portion of a recall protocol that is left over when verbatim or
paraphrased reproductions of the text are deleted.
 Learning from text 289

Correct

Erroneous

Figure 9.2 Reconstructive recall. Correct and erroneous elaborations and

inferences as a function of knowledge and writing style. After Moravcsik and
Kintsch (1993).

ing their style. In one version, the language was as helpful as we could
make it in signaling discourse importance to the reader. The other version
was as unhelpful as we could make it and still write an English text. Writ¬
ing quality had a major effect on reproductive recall, facilitating repro¬
duction about as much as domain knowledge did, but it did not help
understanding. Whether or not a text was well written did not have a sta¬
tistically significant effect on the proportion of erroneous elaborations.
Thus, although good writing can help the reader to construct a better
textbase that is sufficient for recall, it does not by itself guarantee the
deeper understanding that is a prerequisite for learning.
Just how low-ability students go about using their domain knowledge
to achieve good comprehension results was investigated by Yekovich,
Walker, Ogle, and Thompson (1990). One group of their subjects had
high knowledge of football, and one had low knowledge. With texts that
had nothing to do with football, the two groups performed equivalently.
With a football text, the high-knowledge subjects outperformed the low-
knowledge subjects, as in the other studies reported here. The questions
on which the high-knowledge subjects showed the greatest advantage
290 Models of comprehension

over low-knowledge subjects were inference questions and integrative

summary statements. There was less of a difference on memory and detail
questions. This is exactly what the theory of comprehension would lead
one to expect. Even readers with little domain knowledge can understand
information that is explicitly given in the text (though they might not
remember it because their retrieval structures might not be rich enough).
However, inferences and thematic integration that build retrieval struc¬
tures require knowledge.
The study that most thoroughly investigates the relationship between
reading skills and domain knowledge is one by Voss and Silfies (1996),
who observed what students learned from history texts that were either
causally explicit or not. They clearly identified two sets of relevant vari¬
ables, one having to do with reading skills in general and one with knowl¬
edge of history. But Voss and Silfies are no longer concerned only with
whether students are good readers; they also ask whether the students are
good learners. The Cl model provides the background for this shift in
focus from text memory to learning, and that is where we turn next.

9.2 Learning and memory

There are important psychological differences between learning from a

text and remembering the text. Text memory - that is, the ability to
reproduce the text verbatim, in paraphrase, or by summarizing it - may
be achieved on the basis of only superficial understanding. In the extreme
case, one can learn to recite a text by rote without understanding it at all.
Learning from text, on the other hand, requires deeper understanding. I
define learning from text as the ability to use the information acquired
from the text productively in novel environments. This requires that the
text information be integrated with the reader’s prior knowledge and
become a part of it, so that it can support comprehension and problem
solving in new situations. Mere text memory, on the other hand, may
remain inert knowledge - reproducible given the right retrieval cues but
not an active component of the reader’s knowledge base.
This distinction is not always made in the literature. In the memorv lit-
erature, for instance, one talks about a subject learning a list of paired
associates or, synonymously, remembering it. On the other hand, there is
an analogous distinction in the problem-solving literature, made espe¬
cially by Wertheimer (1945), in differentiating superficial problem solv-
 Learning from text 291

ing, which is only mechanical, from real problem solving, which involves
deep understanding. The same difference exists in text comprehension.
In most cases, deep understanding and mere memory for the words of a
text are intermingled to various degrees, but there are experiments in the
literature that represent fairly pure cases of one or the other extreme.
One is the well-known study of Bransford and Franks (1971), which can
be characterized as all understanding and no memory. In this study, the
texts consisted of four simple sentences, such as The ants mere in the
kitchen. The ants ate the jelly, and so on. These ideas could also be expressed
in more complex sentences, such as The ants in the kitchen ate the jelly. Sub¬
jects were given several such texts, either in the form of four one-idea sen¬
tences, two two-idea sentences, one three-idea sentence, and one one-idea
sentence, and other such combinations. Later, they were given a recogni¬
tion test consisting of some sentences they had actually read and others
they had not seen before. The results of the study were clear: Subjects
remembered very well the stories they had read (e.g., they remembered
the ants and the jelly and whatever else there was to that text) but did not
know which particular sentences they had read. They remembered the
meaning of each minitext, a scene, an image, perhaps - but not the way it
had been presented verbally. The memory for the actual text they had read
was wiped out by heavy interference (the subjects read many sentences, all
very similar), but they had no trouble keeping in mind the few simple and
distinct situation models they had formed for each of the several texts they
had read.
Close to the other extreme - all memory and little understanding - is
the study by Moravcsik and Kintsch (1993). Their readers had formed a
good textbase but only a very sketchy or erroneous situation model. 1 he
subjects in the Bransford and Franks experiment had formed a good sit¬
uation model but no textbase at all. Given the extreme simplicity and arti¬
ficiality of the texts used by Bransford and Franks, it seems likely that
their subjects merely combined what was in the text into an integrated
scene to form a situation model.

9.2.1 Textbase and situation model

Because the distinction between textbase and situation model is impor¬

tant for the work discussed in this chapter, it is worth reviewing the main
points here. First, it must be remembered that the textbase and the situ-
292 Models of comprehension

ation model are aspects of the same episodic memory trace of a text. We
distinguish these aspects for the purposes of scientific analysis because
this distinction is often useful in research as well as instruction.

1. Those elements or links in the mental representation that have a

direct correspondence to elements in the text make up the textbase; to
construct the textbase one needs syntactic, semantic, and (some, depend¬
ing on how this is defined) pragmatic knowledge.
2. Texts usually describe real or imaginary situations in the world. The
situation description that a comprehender constructs on the basis of a
text as well as prior knowledge and experience is called the situation
model.
3. In a totally complete, explicit text, a complete and adequate situation
model is described. Thus, the textbase that tells it all is also a good situa¬
tion model. Usually, however, texts are incomplete and rely on the com¬
prehender to fill in gaps and make links to prior knowledge. This must be
done on the basis of the comprehender’s general knowledge — domain
knowledge, knowledge of the world, and knowledge of the specific com¬
municative situation. Therefore, in the general case, the situation model
that a reader generates from a text is a mixture of text-derived (the
textbase) and knowledge-derived elements. If the reader has no relevant
background knowledge or does not employ it in understanding a text, the
text representation will be dominated by the textbase. At the other
extreme, if rich, relevant background knowledge is available and the text
itself is poorly written and disorganized, the reader’s knowledge elabora¬
tions may come to dominate the mental representation of the text, and a
good situation model may be obtained at the expense of the textbase.
Normally the two components, text-derived and knowledge-derived, are
more balanced, resulting in a more or less complete textbase and a more
or less adequate situation model.
4. The distinction between the micro- and macrostructure of a text is
orthogonal to the textbase-situation model distinction. Microstructure
refers to local text properties, macrostructure to the global organization
of text. Either may be directly cued by the text or constructed on the
basis of the comprehender’s knowledge (e.g., by using a schema to orga¬
nize a text whose macrostructure is not made explicit in the text, or by
adding some detail to the text representation that would be expected on
 Learning from text 293

the basis of the comprehender’s knowledge but was not actually stated in
the text).
5. A third distinction that must be kept in mind is how complete and
good each one of these structures is. One may have a poor textbase (either
at the micro- or macrolevel), perhaps because the text is poorly written or
perhaps because the comprehender did not encode properly what was
there. One may have a poor situation model, perhaps because the com¬
prehender lacked the required knowledge, perhaps because he or she
failed to apply it. A poor textbase may be combined with a poor situation
model but may also be associated with a good situation model (as was
shown in Figure 4.1 earlier).
6. As long as one tests for information that is directly given in the text,
one is measuring textbases. For instance, in many laboratory experiments
recall is almost purely reproductive, as are summaries. They may involve
some semantic knowledge, as in generalizing a concept, but that still
remains at the textbase level. Recall and summaries are not always repro¬
ductive, of course, and can be perfectly good indicators of well-developed
situation models, when and if they go beyond the text.
As soon as one tests for things not directly in the text, one is testing for
aspects of the situation model. Inference questions and sorting tasks are
obvious examples. In some cases, recall or summaries may also function
as situation model measures to the extent that they go beyond reproduc¬
tion.
7. Bridging inferences often involve merely semantic information; one
does not need a situation model construct for that case. They may, how¬
ever, involve more than that in some texts. Often a comprehender needs
a good model of the situation under discussion in order to make some of
the bridging inferences required by the text.

9.2.2 Learning

Situation models not only integrate the text but integrate the text with
the reader’s knowledge. An example from Kintsch (1994c) makes this
clear. In Figure 9.3 a brief text is shown, together with its propositional
textbase and a situation model in the form of a graph. The textbase con¬
sists of three complex propositions linked by sentence connectives. 1 he
situation model consists of an image (shown here as a graph), plus back-
294 Models of comprehension

Text
When a baby has a septal defect, the blood cannot get rid of enough carbon
dioxide through the lungs. Therefore, it looks purple.

Textbase and Situation Model:

WHEN

Figure 9.3 A text, its textbase, and situation model. The textbase is printed
in italics and is linked to prior knowledge, forming a situation model. After
Kintsch (1994c).

ground information about the role of some of the components of the

image. It is mostly based on the reader’s knowledge about the circulatory
system, rather than on the text directly. Only the fact that there is a gap
in the septal wall so that purple blood gets mixed with the red blood is
derived from the text itself.
I do not claim that every reader will construct exactly this textbase and
situation model. There may be some variation in the textbase (e.g., some
readers probably will not encode the modifier ENOUGH). And there
will certainly be even more variation in the situation model; not every
reader will employ imagery. Some readers may form an even more
sketchy model (or even none at all); others a much more elaborated one.
Some readers will undoubtedly get some of the model wrong.
Text memory may be reproductive (e.g., the subject retrieves the
 Learning from text 295

textbase that has been formed and reproduces it), or it may be recon¬
structive (e.g., the subject retrieves the situation model that has been
formed and reconstructs the text on this basis). Obviously, the two cases
are not mutually exclusive. Learning from text, according to the defini¬
tion given earlier, requires the formation of a situation model. Thus, the
subject has learned something from the text in Figure 9.3 if the septal gap
is added to the reader’s knowledge about how the blood circulates in the
heart.

9.3 The measurement of learning

To be useful in research, the theoretical distinction between learning from

a text and memory for a text requires empirical methods to assess learning
separately from memory. However, because learning and memory cannot
be separated cleanly even in the theory (textbase and situation model are
not two separate structures, but the text-derived and knowledge-derived
components of a single structure, as in Figure 9.3), measurement proce¬
dures are not precisely separable into textbase and situation model mea¬
sures, either. Instead, empirical measures reflect one or the other aspect
of the structure to a stronger degree. Thus, one can ask questions that
demand a specific detail from the text - or that require the integration of
textual information and prior knowledge - in order to solve a new prob¬
lem. Even recall reflects both aspects: the textbase to the extent that the
recall is reproductive, and the situation model to the extent that it is recon¬
structive. Usually, recall is a mixture of the two, but in some cases it is pri¬
marily one or the other.
Text memory is measured in the conventional way, as described in the
previous chapter: through free recall, cued recall, summarization, various
types of recognition tests, and text-based questions. Different methods
are needed, however, for the measurement of learning. Psychology shares
the need for such measures with AI, insofar as AI is interested in the con¬
struction of knowledge-rich expert systems (e.g., Olson & Biolsi, 1991),
and with education, where the assessment of learning is of obvious
importance. Education and AI have relied for the most part on direct
methods for knowledge assessment, that is, various forms of question ask¬
ing. That is still by far the most widely used method in psychology too,
although more indirect scaling methods have also been developed for
purposes of psychological research.
296 Models of comprehension

Asking questions as a method for the assessment of knowledge is

fraught with problems. Developers of expert systems rely on this method
almost exclusively, but it is difficult to determine the correctness or com¬
pleteness of the answers that are elicited. Educationally, the problem is
that asking questions is artificial and sometimes yields invalid results. It
is an unnatural act when a teacher asks a student for something the
teacher knows better than the student. Furthermore, the answers that the
students give may indicate much else besides real learning. Students may
acquire specific strategies that allow them to generate acceptable answers
without having deeper understanding. Or questions may be answered
correctly or wrongly for various accidental reasons that have nothing to
do with the students’ understanding. These problems are widely appre¬
ciated but not easily avoided.
Asking questions to assess knowledge for scientific purposes is limited
in its effectiveness because we do not have a detailed theory of question
answering. As long as it is not known just what psychological processes
and what knowledge are involved in answering a particular question, we
simply do not have a reliable way of constructing the right questions for
our purposes. Psychologists, like teachers, must and do rely on their intu¬
ition, which sometimes yields satisfactory results and sometimes does
not. It is not surprising, therefore, that psychologists have looked for
alternative and more objective ways of assessing knowledge. In particular,
various scaling methods have been developed for the indirect assessment
of knowledge.
Scaling methods require a set of key words or phrases that are charac¬
teristic of a certain knowledge domain. (One can ask experts for such
words, or use more objective methods, such as frequency counts of tech¬
nical terms in relevant scientific publications.) The knowledge of a sub¬
ject is inferred from the way the subject organizes these key words. If the
structure generated by the subject resembles the structure generated by
domain experts, we infer that the subject’s knowledge organization is
similar to that of the experts. To the extent that the subject structures the
set of key words in ways that differ from the experts, a lack of correct
domain knowledge is revealed.
The basic technique for finding out how a subject organizes a set of key
words is to ask the subject for relatedness judgments between all pairs of
key words in the set. A similarity matrix between all keywords is thus
 Learning from text 297

obtained, showing the rated closeness between all word pairs. Such a
matrix can then be used as the basis for multidimensional scaling (e.g.,
Bisanz, LaPorte,Vesonder, &Voss, 1978; Henley, 1969; Stanners, Price, &
Painton, 1982). A low-dimensional space is generated in which the key
words are embedded. One can then ask whether the space is the same for
the students as for the experts, and whether the location of the key words
in this space is similar for students and experts. This method has been
used successfully a number of times. For instance, the semantic field of
animal names has been scaled in this way, yielding a space with the two
dimensions of size and ferocity, which account for 59% of the variance of
the paired-comparison judgments (Henley, 1969).
However, these scaling methods are of limited usefulness. The pairwise
comparison method is laborious for the subject and rapidly becomes
impossible to use as the number of keywords increases. Furthermore,
multidimensional scaling methods work with group data, but we often
need to work with data from individual subjects. Most important, how¬
ever, it has become apparent that very few knowledge domains (other
than animal names) are regular and simple enough to be described by a
space of a few namable dimensions.
A more practical method for indirect knowledge assessment has been
developed by Ferstl and Kintsch (in press). It applies to the problem of
measuring the amount of learning that occurs from reading a text. If the
text has an effect on the reader’s memory and knowledge, it should
change the way the reader organizes a knowledge domain, and the change
should be in the direction of the text organization. Thus, we assess a
reader’s knowledge about a particular domain, have the reader study a
related text, and reassess the reader’s knowledge organization to see
whether it has changed in accordance with the text organization.
Ferstl and Kintsch chose a knowledge domain that is quite rich for
most subjects and fairly stereotyped: the birthday party. We then wrote a
story about a somewhat weird birthday party, designed to distort our
readers’ conventional birthday party schema. We assessed this particular
schema both before and after reading the story. The question we were
concerned with was whether the story had an effect on the way subjects
organized the birthday party domain, whether this effect could be attrib¬
uted to reading the story, and whether we could measure this effect with
our procedures.
298 Models of comprehension

We selected a set of 30 key words characteristic ot the birthday party

domain from free associations subjects gave to the stimulus word birthday
party. To these we added 30 words that were important in the story we
had written about a birthday party but that were not otherwise associated
with birthday party (such as clown, credit card). It is, of course, impossi¬
ble to do pairwise comparisons with such a large set of keywords. Instead,
we used two less demanding methods to obtain relatedness matrices from
our subjects. One was a sorting task for which we gave each subject a stack
of 60 index cards, each printed with one of the key words, and asked them
to sort these cards into piles according to relatedness. We made sure that
the subjects understood that there was no right or wrong way of sorting,
that we were interested only in their intuitions about which of these
words were related to each other. Thirty-two subjects performed this
sorting task, and from their results a relatedness matrix was constructed,
showing for each word pair how many times it had been sorted together
into the same group.
The other method we used was cued association. The subjects were
asked to read silently through the list of key words twice (to familiarize
themselves with this word set), and then they were asked to generate up
to three associations to each keyword. No restrictions were placed on the
associations subjects generated, but because most of our keywords were
associated with the birthday party theme and because subjects were
primed by having just read these words, most of the associations they
generated were actually from our set of key words. Thus, the number of
times a key word was given as a response to another key word could be
used as a measure of relatedness between the two words. Notice that the
relatedness matrix generated in this way is asymmetric, whereas the
matrix generated from the sorting data is necessarily symmetric.
Two analyses were performed on these relatedness matrices, a hierar¬
chical clustering analysis after Johnson (1967) and a Pathfinder analysis
after Schvaneveldt (1990). The former is particularly suited for symmet¬
ric relatedness matrices, the latter for asymmetric matrices. The results of
the clustering analysis of the sorting data prior to reading are shown in
Figure 9.4. The results of the Pathfinder analysis of the cued association
data prior to reading are shown in Figure 9.5.
Figure 9.4 shows that the sorting procedure and cluster analysis yields
readily interpretable results. The key words are neatly organized by cate-
 Learning from text 299

age clown games cake basket buying Charlie crying

balloon music hideseek candy sack creditcrd Tim screaming
decoratn excitemnt pintail icecream bedroom camera children swimrnng
confetti fun videogm hamburgr livingrm car friends dream
streamers laughing toys hotdogs kitchen keys dog nap
partyfvr playing soda table garage father
partyhat singing lunch ride mother
bows parents
ribbons
wrapping
candles
brithday
party
HappyB
presents

Figure 9.4 The two top layers in the clustering hierarchy before reading
based on sorting data. Only the two top layers are shown; long words and
compounds are abbreviated. After Ferstl and Kintsch (in press).

gory membership. Indeed, the results are almost too neat, suggesting that
subjects performed some sort of semantic analysis when sorting the key
words. We have previously argued (Walker & Kintsch, 1985) that the sort¬
ing task produces results that are a bit too orderly and logical and do not
necessarily reflect the memory structures that are operative in memory
retrieval. In comparison, the structure derived from the cued association
data (Figure 9.5) is much richer. It still shows much the same clusters as
Figure 9.4, but there is a complex, rich pattern of interconnections evi¬
dent that Figure 9.4 lacks. My guess is that the cued association data more
accurately reflect knowledge organization than the less spontaneous sort¬
ing data do, but whether or not this is so would have to be established
empirically.
If Figures 9.4 and 9.5 represent our subjects’ birthday party schema,
300 Models of comprehension

Figure 9.5 Part of the Pathfinder structure based on the cued association
data. Only the links that were present both before and after reading are
shown. Heavy lines indicate links that were stronger after reading, and thin
lines indicate links that were stronger before reading. After Ferstl and Kintsch
(in press).

how did this structure change after they read the birthday party story?
One obvious way to answer this question is to have subjects re-sort the
key words after reading the story, or ask them again for their cued associ¬
ations, and construct from the resulting data postreading structures to be
compared with Figures 9.4 and 9.5. Ferstl and Kintsch showed that it is
indeed possible to see how these structures changed as a result of reading
and argue that the changes that have taken place can be attributed to the
new relations among the key words that were established by the story the
subjects had read. Flowever, this is a bit like interpreting inkblots, and
there is cei tainly a danger of reading more into the data in accord with
one’s expectations than the facts warrant.
 Learning from text 301

Fortunately, objective measures of change can be devised that have the

advantage that they are not restricted to group data but can be computed
for individual subjects, hence making standard statistical hypothesis test¬
ing procedures available. The strength with which two concepts are
linked in a text can be determined from the propositional structure of the
text. If the concepts belong to the same propositions, they are strongly
linked; if they belong to different propositions, they are linked to the
degree that these propositions are related in the textbase. Thus, the num¬
ber of links that must be traversed from a proposition containing one
concept to reach a proposition containing the other concept provides a
convenient metric for the relatedness of concepts in the text structure. It
is therefore possible to determine the extent to which the associations a
subject gives coincide with the text structure. Figure 9.6 shows the pro¬
portion of associations subjects gave in the Ferstl and Kintsch experi¬
ment that coincide with text links before they read the text, immediately
after reading the text, and one week after reading. Because a percentage
value can be computed for each subject, these values can be analyzed sta¬
tistically. In this case, we can conclude that the proportion of associations
coinciding with text links is significantly higher after reading than before

Figure 9.6 The percentage of associations before reading, immediately after

reading, and one week after reading that coincide with text links. After Ferstl
and Kintsch (in press).
302 Models of comprehension

reading the text, both immediately after reading and one week later. How¬
ever, the proportion after one week is significantly lower than the pro¬
portion immediately after reading.
The story subjects read thus clearly influenced the pattern of associa¬
tions they produced. Was their knowledge changed thereby, or was what
we see in Figure 9.6 merely due to the influence of the episodic text rep¬
resentation they had formed? We don’t know this, but if episodic mem¬
ory and knowledge are all part of the same dynamic network, this is not a
question we necessarily have to answer. Our subjects had learned some¬
thing new that changed the way they thought about birthday parties,
more strongly right after they read the story than a week later, but even
then some change could still be documented.
The measurement of learning from text, as distinct from text memory,
is probably best approached through a combination of direct and indirect
procedures: questions that require inferencing and problem solving on
the one hand and scaling methods based on sorting or association tasks,
as in Ferstl and Kintsch (in press), on the other. Both approaches have
their limitations, so it is important to obtain results with different meth¬
ods that confirm each other. The experiments discussed in the next sec¬
tion demonstrate the practicality of this approach.

9.4 A simulation of learning with the Cl model

In what way is learning from a text different, in terms of the construc¬

tion-integration model, from text memory, which was investigated in
chapter 8? I have discussed this question in Kintsch (1994c) and have
provided a simulation to illustrate the crucial differences. It is assumed
that the goal of learning is to form a situation model that integrates the
textual information with prior knowledge, as in Figure 9.3, not merely
the formation of a textbase. Thus, any simulation of learning from text
must explicitly include a knowledge component.
In Kintsch (1994c) this was done by first assessing a subject’s knowl¬
edge about a particular domain (heart disease) by means of a large set of
questions. The questions that were answered correctly were used to con¬
struct a knowledge map for each subject. Figure 9.7a shows an example
of such a knowledge map for a high-knowledge subject. All questions that
were correctly answered are turned into statements and expressed as
Lea rning from text 303

CONTRACT! V ALVEI
I
OPEN [VALVE] - -CAUSE PUMP[BLOOD] -
I
ALLOW[VALVE.S, ' IN-HEART

. FLOWfBLOOD, FROM-HEART, TO-BODY, IN-BLDVESSEL]

1— NAME[$, ARTERY] — PULART[TO-LUNG]

^ AORTAfTO-BODY]

. FLOW[BLOOD, FROM-BODY, TO-HEART, IN-BLD VESSEL]

'''nAME[$,VEIN]-MAIN [VEIN]-VENACAVA

- FLOW[BLOOD, FROM-BODY, TO-HEART, TO-LUNG]

REASON

PROVIDE[LUNG, $, OXYGEN]

CAUSE "S>ION-OXY[BLOOD]

OXYtBLOOD]-CAUSE -RED[BLOOD]

BEAT[HEART] -CAUSE -PULSE[IN-WRIST, IN-NECK]

INDICATE

. SPEED[$]-HEARTRATE

HAVE[HEART, CHAMBER]-FOUR

UPPER[CHAMBER] ^ ~ ~~~ LOWER[CHAMBER]

NAME[$, ATRIUM] NAME[$, VENTR]

LEFT[ ATRIUM] LEFT[VENTR] RIGHT[ ATRIUM] RIGHT [VENTR]

DIVIDE[SEPTUM ,$,$] X. /
ART]_ J I_RIGirr [HEART]

CONSIST[HEART,MUSCLE]

(a)

BEAT[HEART] -CAUSE-PUMP[HEART, BLOOD, TO-BODY, IN-BLVESS]

DICATE

‘ PULSE[IN-WRIST, IN-NECK]

(b)

Figure 9.7 Knowledge maps for (a) a high-knowledge subject and (b) a low-
knowledge subject. After Kintsch (1994c).
304 Models of comprehension

propositions. Links are drawn between core propositions and their mod¬
ifiers, and whenever propositions are linked by a specific semantic rela¬
tion (like CAUSE, or CONSEQUENCE), or when one proposition is
embedded as an argument of another proposition. An inspection of Fig¬
ure 9.7a shows that this subject knew quite a bit about the heart and that
his understanding about the functioning of the heart was basically cor¬
rect. Compare this with the skimpy knowledge map also shown in Figure
9.7b for a low-knowledge subject.
Two separate processing simulations of a text about heart disease (not
shown here) were performed, one for the high- and one for the low-
knowledge reader. The assumption was made that whenever a concept
occurred in the text, it retrieved all the knowledge the reader had about
this concept. Thus, the phrase the heart supplies blood to the body retrieved
PUMP[FIEART,BLOOD] for the low-knowledge subject, and FLOW
[BLOOD, FROM-HEART,TO-BODY] for the high-knowledge subject.
Valve retrieved ALLOW[VALVE,PUMP[BLOOD], FLOW [BLOOD,
FROM-HEART,TO-BODY]] for the high-knowledge subject and noth¬
ing for the low-knowledge subject, and so on. In this manner 15 nodes
from the knowledge map for the high-knowledge subject and one node
from the knowledge map of the low-knowledge subject (the nodes printed
in boldface in Figure 9.7) were added to the simulation at the appropri¬
ate points. It is not necessarily the case that every bit of knowledge a
reader knows that is relevant to a certain text is actually retrieved when
processing this text, but for present purposes this is a reasonable simpli¬
fying assumption.
For the simulation, working memory was set to five atomic proposi¬
tions. Hence, the model reads the proposition(s) from the first sentence,
goes on to the next sentence, and continues until it has read at least five
elements. The model then retrieves from its knowledge base (the maps
shown in Figure 9.7) any related information and adds these elements to
the network. The network in working memory is then integrated. The
most strongly activated proposition in each processing cycle is carried
over to the next processing cycle in the short-term memory buffer.
The results of the simulation (not shown here) were not very different
for high- and low-knowledge subjects so far as text propositions are
concerned. The pattern of activation is similar, and the average activa¬
tion strengths are about the same. However, there are two important
differences. First, because 15 knowledge nodes were combined with
 Learning from text 305

the text structure for the high-knowledge subject but only one for the
low-knowledge subject, the text representation generated by the high-
knowledge subject is solidly anchored to the subjects’ prior knowledge.
This is not the case for the low-knowledge subject. Second, the prior
knowledge that has become incorporated into the text representation for
the high-knowledge subject bridges some coherence gaps in the textbase
of the low-knowledge subject. Where the text is coherent, the textbase for
the low-knowledge subject is also coherent, but where there is a gap in the
text, requiring special knowledge for a bridging inference, the textbase
lacks coherence. This is not the case for the high-knowledge reader,
where gaps in the text can be bridged by prior knowledge.
What empirical predictions do these simulations imply? Behaviors that
depend mostly on the textbase, such as questions about facts explicitly
stated in the text, should not differ much for high- and low-knowledge
subjects. High-knowledge subjects, however, should be better on ques¬
tions requiring inferences or problem-solving tasks. The predictions for
free recall are somewhat ambiguous. The reproductive component should
be about equal for high- and low-knowledge readers, especially for a text
that does not require many bridging inferences, but the former readers
could successfully reconstruct portions of the text on the basis of their
deeper understanding of the situation and hence end up with higher recall
scores overall.
Predictions for the sorting task are illustrated in Figure 9.8. Consider
how subjects with different knowledge would sort the four key words oxy¬
gen, purple, septal defect, and brain defect before reading the text on heart
disease. A subject who knows nothing about septal defects would put the
two defects together and sort the remaining two words into separate cat¬
egories. A more knowledgeable reader, on the other hand, should sort the
first three words together and use a separate category for brain defect. Fig¬
ure 9.8 shows that the simulation makes just these predictions. 1 he figure
shows a segment of the mental representation constructed by the high-
knowledge reader, together with the four key words. An inspection of the
figure shows that there exist pathways with strong links between oxygen,
purple, and septal defect, which therefore would tend to be sorted together,
whereas brain defect is only weakly linked to septal defect. The predictions
are quite different for the low-knowledge reader, however. The low-
knowledge reader’s text representation is the same as that of the high-
knowledge reader, except that it lacks the knowledge component (all the
306 Models of comprehension

Figure 9.8 Network formed by four key words from the sorting task, related
text propositions from reading, and related knowledge items. The numbers
refer to the memory strengths of the links. After Kintsch (1994c).

nodes below the lower dashed line in Figure 9.8). It is easy to see that in
this case there are no pathways linking purple with any of the other words.
Hence, after reading the text a low-knowledge subject should keep purple
separate in sorting these four key words but group oxygen and septal
defect, which are linked together on the basis of the text alone. A low-
knowledge subject who had not read the text would group the two defects
only on the basis of their preexperimental association.
The most important predictions that can be derived from the simula¬
tion described concern learning from text, that is, the usability of the tex¬
tual information in novel situations, which should be fairly high for the
high-knowledge subject. According to the simulation, there are 12 entry
points that allow access to the textual information from the reader’s
knowledge base — the items in boldface type in Figure 9.7 that form links
between the text representation and background knowledge. Thus, even
at a time after reading when the contextual and temporal cues no longer
permit retrieval of the episodic text structure, it remains still accessible
via the reader’s general knowledge. Indeed, it has become a part of it and
ceased to exist effectively as a separate episodic memory structure.
Learning has occurred. With only one reading, the links may be too weak
 Learning from text 307

to permit reliable retrieval at a later date — typically one does not learn
much from a single reading - but they could be strengthened through
further study. The high-knowledge reader of our simulation is on the way
to learning something he or she did not know before about the heart.
This is not the case for the low-knowledge reader, who in the simula¬
tion forms much the same text structure as does the high-knowledge
reader, except that it is only very weakly linked to his knowledge base.
This reader has stored in his memory the same information as the high-
knowledge fellow student, but the only way to retrieve it is through the
contextual and temporal retrieval cues associated with the episodic text
memory. It is thus inert knowledge - the student knows something but
cannot spontaneously make use of this knowledge in novel tasks because
what is known is not tied in with the rest of his knowledge. The informa¬
tion may be retrievable given a retrieval cue such as “Think of the chap¬
ter on heart disease that you read here the other day,” but once such
episodic retrieval is no longer effective, the text information is lost.

9.5 Using coherent text to improve learning

According to the simulation described in the previous section, learning

from text requires that the learner construct a coherent mental represen¬
tation of the text, and that this representation be anchored in the learner’s
background knowledge. Thus, one reason students might fail to learn
something from reading a text could be that they are unable to form a
coherent textbase linked to their pre-knowledge.
It is easy to see why this might be the case for low-knowledge readers.
Not all links either within a text or between the text and the readers’
knowledge are always spelled out in a text but are often left for the reader
to fill in, for example, as bridging inferences. This is fine and, as we shall see
in the next section of this chapter, can be quite advantageous, but it often
creates problems for low-knowledge readers. If readers simply do not
have the necessary background knowledge to fill in the gaps in the text
that an author has left, they will be unable to form a coherent representa¬
tion of the text or to link it with whatever little they do know. Consider
some trivial examples.

(1) The heart is connected to the arteries. The blood in the aorta is
red.
308 Models of comprehension

For a reader who does not know what the relationship between arteries
and aorta is, there will be a coherence problem. Or consider

(2) To stop the North Vietnamese aggressors, the Pentagon decided

to bomb Hanoi.

which may present all kinds of problems to a low-knowledge reader.

Namely, what is the Pentagon and how does Hanoi get into this sen¬
tence? A little rewriting of these problem texts to make the relations
between items in the text or between general knowledge and the text fully
explicit can avoid these problems. Thus, we might define aorta for a low-
knowledge reader by rewriting (1) as

(la) The heart is connected to the arteries. The blood in the aorta, the
artery that carries blood from the heart to the body, is red.

And we can help with (2) by inserting explicit links between the unknown
terms and what a reader might be expected to know.

(2a) To stop the North Vietnamese aggressors, the U.S. Defense

Department in the Pentagon decided to bomb Hanoi, the capital
of North Vietnam.

So far we have considered only local coherence problems, but the

global coherence of a text can often also be made more explicit. The
macrostructure of a text is not always explicitly signaled in the text but is
left for the reader to deduce. This is fine for knowledgeable readers but
can be a major source of confusion when the requisite background knowl¬
edge is lacking. Thus, most of us need no help to understand the struc¬
ture of a four-paragraph text, each describing the anatomical details of
one of the chambers of the heart. However, for a reader who does not
know that the heart has four chambers, a title like The four chambers of the
heart, plus appropriate subtitles or clearly marked topic sentences for
each paragraph, can be of great help.
Such relatively minor revisions of texts that ensure coherence at both
the local and global levels facilitates text memory as well as learning for
readers who lack background knowledge. Figure 9.9 summarizes the
results of three studies in which the effectiveness of such revisions were
explored.
Beyer (1990) used a computer manual as his learning material. He
 Learning from text 309

Figure 9.9 The effectiveness of text revisions in the experiments of Beyer

(1990), Britton and Gulgoz (1991), and McNamara, Kintsch, Singer, and
Kintsch (1996, Exp. 1). To make the data comparable, the improvement in
performance after reading the revised text is expressed as a percentage of the
performance with the original text.

revised the original manual by making its macrostructure explicit by

means of titles and subheadings and by improving the comprehensibility
of the instructions contained in the manual with illustrative examples.
The revised text proved to be significantly better than the original ver¬
sion, but the improvement was restricted to problem-solving tasks. When
questions were asked about facts that were explicitly stated in the text, the
original text and the revised text yielded about equal performance. How¬
ever, the students learned better with the revised text, as measured by
their performance on inference questions (Figure 9.9).
A more systematic approach to filling in the local coherence gaps in a
text was employed by Britton and Gulgoz (1991), who used a history text
as their learning material that described the U.S. air war in Vietnam.
The text was written at the time of the war and presumed considerable
prior knowledge on the part of the reader - knowledge that at the time
was probably readily available among the population to which the text was
310 Models of comprehension

addressed. Many years after the war, the people who participated in Brit¬
ton and Gulgoz’s study had very little specific information about the war
and hence found the text hard going. Although they were able to recall
quite a bit from the text and answered more than half of the fact questions
correctly, their performance was poor on the inference questions. In fact,
they really did not understand the text at all. This conclusion follows
from an analysis of the conceptual understanding of the situation as it was
assessed by a key word comparison task.
Britton and Gulgoz (1991) selected 12 key words that were of crucial
importance for the understanding of the text and had students give pair¬
wise relatedness judgments for these key words. They also had the origi¬
nal author of the text plus several other experts on the Vietnam war pro¬
vide relatedness judgments for the same set of key words. The author and
experts agreed quite well among themselves (average intercorrelation r =
.80). But the students who had obtained their information from reading
the text did not agree with the author or the experts at all (average r =
.08). Their understanding of the air war in Vietnam on the basis of read¬
ing this text was quite different from the one the author had intended
(and which the experts achieved from reading the same text). Thus, even
though the students recalled a good part of the text and answered ques¬
tions about it reasonably well, they really did not understand what they
were saying! Britton and Gulgoz also report a Pathfinder analysis of the
key word judgment data that makes clear some of the fundamental mis¬
conceptions of these readers. For instance, whereas the text emphasized
the failure of the operation Rolling Thunder, in the students’ judgments
Rolling Thunder was linked to success, instead. Britton likens this result to
reading the Bible and concluding that the devil was the good guy.
This dismal performance could be significantly improved by some
rather simple revisions of the original text to make it understandable to
readers without adequate background knowledge. Britton and Gulgoz
(1991) used the Miller and Kintsch (1980) simulation program to locate
all coherence gaps in their original text. Whenever that program encoun¬
ters a coherence gap (no argument overlap between propositions), it stops
and asks the operator to supply a bridging inference. For instance, if in
one sentence North Vietnam is mentioned, and the next sentence begins
with In response to the American threats, Hanoi decided, Britton and Gul¬
goz might have made this sentence pair coherent by adding to the second
 Learning from text 311

sentence the North Vietnamese government in Hanoi decided. These revi¬

sions were highly effective. Recall increased significantly, as did the per¬
formance on inference questions. The improvement on fact questions
was not significant, as in Beyer’s experiment (Figure 9.9). Most impor¬
tant, however, readers now understood the text more or less correctly.
Their relatedness judgments after reading the revised text correlate rea¬
sonably well with those of the experts (r = .52), and if one looks at the
structure of their judgments as revealed by a Pathfinder analysis (not
shown here), no glaring misunderstandings are apparent, as was the case
for the readers of the original version of the text. Hence, making this text
locally coherent by filling in the gaps that required bridging inferences
yielded a text that readers could understand, even though their back¬
ground knowledge was lacking.
A third experiment comparing the effectiveness of revising a text for
greater coherence has been reported by McNamara, E. Kintsch, Songer,
& W. Kintsch (exp. 1, 1996). In this study, the subjects were students in
grades seven through nine. As learning material, we used a junior high
biology text describing characteristics of mammals. The text was 1.5
pages long and was well written from the standpoint of local coherence,
so that there were no local coherence gaps that had to be filled in the revi¬
sion. However, the text had a listlike macrostructure that was not clearly
signaled, and the paragraph structure did not always correspond to the
subtopic structure of the text. Hence, revising the text amounted to
adding material that explicitly identified the major subtopics as traits of
mammals and reorganizing the paragraphs so that they corresponded to
the subtopic structure. Subjects were given a prior knowledge question¬
naire, after which they read either the revised or the original text at their
own rate. They then were given a recall test, a postreading questionnaire,
and a sorting test. For the latter, subjects were given 16 key phrases to
sort, which could be sorted in multiple ways. Eleven of these key phrases
were characteristics of mammals (e.g., has hair or fur) and five were non-
mammalian traits (e.g., is cold blooded). We were interested in learning
whether reading the text on mammalian traits increased the subjects ten¬
dency to use “mammalian” as a sorting criterion.
The recall data replicate Britton and Gulgoz (1991), in that subjects
recalled the revised text significantly better than the original text (Figure
9.8). Further analysis showed that this improvement was mainly a conse-
312 Models of comprehension

quence of the better recall of macropropositions in the revised text, which

would be expected, because the revision involved primarily the macro¬
structure of the text.
An indication that the way our subjects thought about this knowledge
domain was indeed influenced by reading the text - and evidence that this
influence was more pronounced for the revised text - could be obtained
from an analysis of the sorting data. Each subject sorted mammalian
traits together with some likelihood before reading the text on mammals.
From these data, change scores could be computed. These change scores
reveal an increase of 23% in mammal groupings after subjects read the
revised text, compared with only a 10% increase for the original text. In
addition, there was a corresponding decrease in the tendency to combine
mammal characteristics with nonmammal characteristics (-1% for the
original text, -7% for the revised text). Thus, subjects were more likely
to employ “mammal” as a sorting criterion after reading, and especially
after reading the revised text.
Several other studies confirm and extend these results. Beck, McKe-
own, Sinatra, and Loxterman (1991), working with fourth- and fifth-
grade students who studied a text about the American Revolutionary
War, significantly improved performance on open-ended questions and
recall by adding explanatory coherence to the text, especially emphasiz¬
ing causal relations. Similarly, McKeown, Beck, Sinatra, and Loxterman
(1992) manipulated prior knowledge by means of a 35-minute instruc¬
tional unit and text coherence (original and revised text) with fifth-grade
students. Recall and open-ended questions both improved by coherent
texts more than by knowledge alone, but the effects of the two factors
were cumulative. The authors suggest that background knowledge alone
was not sufficient to compensate for the poorly written text. Both prior
knowledge and a coherent text were necessary to construct a good text
representation.
Revising a text for coherence is thus clearly an effective technique to
further understanding and learning. Students without adequate back¬
ground knowledge cannot fill in gaps in the text on their own that readers
with greater familiarity with the domain bridge effortlessly and in fact
unconsciously. For such readers without background knowledge, provid¬
ing explicit bridging material in the text, at both the local and global level,
is a prerequisite for understanding and learning.
If this is so, why do authors ever leave gaps in their texts? Why do we
 Learning from text 313

not write fully coherent, explicit texts all the time? The answer is that to
write such a text is an elusive goal; a writer must always rely on the
reader’s knowledge to some degree. There is no text comprehension that
does not require the reader to apply knowledge: lexical knowledge, syn¬
tactic and semantic knowledge, domain knowledge, personal experience,
and so on. The printed words on the page or the sound waves in the air
are but one source of constraints that must be satisfied. The reader’s
knowledge provides the other. Ideally, a text should contain the new
information a reader needs to know, plus just enough old information to
allow the reader to link the new information with what is already known.
Texts that contain too much that the reader already knows are boring to
read and, indeed, confusing (e.g., legal and insurance documents that
leave nothing to be taken for granted). Hence, too much coherence and
explication may not necessarily be a good thing.

9.6 Improving learning by stimulating active processing

McNamara et al. (1996) conducted a second experiment that explores the

contention that if readers possess adequate knowledge, a fully explicit
text is not optimal for them. Such readers remember more and learn bet¬
ter from texts that require them to assume a more active role in compre¬
hension. Specifically, we hypothesized that the results obtained by Beyer,
Britton, and Gulgoz, and in our experiment 1, which were described in
the previous section, pertain to low-knowledge readers. Readers with
good domain knowledge might react in very different ways.
As the learning material for our experiment, we chose an encyclopedia
article at the junior-high level on heart disease. We assumed that there
would be a great deal of variability in background knowledge among
junior-high students in this area, so that we could meaningfully distin¬
guish between high- and low-knowledge students. We rewrote this text in
several ways, only two of which concern us here. In one case we attempted
to produce a text that was maximally coherent and explicit at both the local
and global level. For this purpose, we replaced potentially ambiguous pro¬
nouns with full noun phrases, added elaborations that linked unfamiliar
concepts to familiar ones, added sentence connectives whenever possible,
and made sure that the same concept was always referred to in the same
way (rather than by a synonym or paraphrase). In addition, we added titles
and subtitles to indicate the macrostructure of the text, as well as explicit
314 Models of comprehension

macropropositions to mark the role of each paragraph in the text. The

resulting version was our high-coherence text. The low-coherence version
was constructed by deleting all these signals but not otherwise changing
the content of the text. Of course, the high-coherence text was quite a bit
longer (1,053 words) than the low-coherence text (683 words).
The prior knowledge of our subjects was assessed by a knowledge
assessment test consisting of a series of questions on the basic anatomy
and functioning of the heart and a recognition test in which subjects were
asked to match parts ofjthe heart to a diagram showing a cross-section of
the heart and its major blood vessels.
After taking this test, subjects read either the high- or the low-coher¬
ence text twice, and then responded to a series of posttests. First they
were asked to recall the text in their own words. Then they were given 41
questions of four types: (1) text-based questions, (2) elaboration ques¬
tions that required relating text information to the reader’s background
knowledge, (3) bridging inference questions that required connecting
two or more separate text segments, and (4) problem-solving questions
that required applying text information in a novel situation. Finally, sub¬
jects performed a sorting task consisting of 18 keywords, some from the
text and others not, that could be grouped in various ways, but for which
there existed only a single text-driven sorting principle.
The results of this study reveal a strong interaction between the level
of prior knowledge of the students, the coherence of the text, and the
method of testing. Fligh-knowledge students always perform better than
low-knowledge students do. But when tests are used that assess primarily
text memory, the high-coherence text is better for all types of students.
This is shown in Figure 9.10 for free recall and in Figure 9.11 for text-
based questions. However, when tests are used that depend on the con¬
struction of a good situation model, the low-coherence text actually yields
better results for high-knowledge subjects than the high-coherence text
does. For low-knowledge subjects, the usual superiority of the high-
coherence version over the low-coherence version is observed. Figure
9.12 shows this interaction for the problem-solving questions and Figure
9.13 for the sorting task. As mentioned before, the sorting task admitted
to several rational groupings of the key words. Figure 9.13 shows the
extent to which subjects changed from whatever grouping they employed
before reading the text to a sorting principle that corresponded to the
text. That score was close to zero for the low-knowledge subjects reading
Learning from text 315

[?]] High-Coherence

Q Low-Coherence

High K Low K

Figure 9.10 Proportion of text propositions recalled as a function of prior

knowledge for high- and low-coherence texts. After McNamara et al. (1996).

o
5
B
o High-Coherence
U
c
o □ Low-Coherence
'£
o
D-
O

High K Low K

Figure 9.11 Proportion correct for text-based questions as a function of

prior knowledge for high- and low-coherence texts. After McNamara et al.
(1996).
316 Models of comprehension

o
<o
t:
o [2] High-Coherence
u
G
O
Low-Coherence
'fi
o
a,
o

High K Low K

Figure 9.12 Proportion correct for problem-solving questions as a function

of prior knowledge for high- and low-coherence texts. After McNamara et al.
(1996).

02
. -
Sorting Score

0.15-
[FJ High-Coherence

□ Low-Coherence
0.1 -

0.05-

High K Low K

Figure 9.13 Sorting scores as a function of prior knowledge for high- and
low-coherence texts. After McNamara et al. (1996).
 Learning from text 317

the low-coherence text. That text was apparently just too difficult for
these subjects, and they could get nothing out of it. However, the con¬
ceptual change induced by the low-coherence text in readers with a good
knowledge background, as indexed by the sorting scores in Figure 9.13,
was almost twice as great as for the high-coherence text. It appears that
the high-coherence text that spelled out everything these readers already
knew quite well induced an illusory feeling of knowing that prevented
them from processing the text deeply. They were satisfied with a superfi¬
cial understanding, which was good enough for recall and answering text-
based questions, but they failed to construct an adequate situation model
combining their prior knowledge with the information from the text.
Hence, their relatively poor performance with the text that should have
been the easiest and most effective text.
The interaction shown in Figure 9.13 has been replicated by McNa¬
mara and Kintsch (1996) with a different text and a different knowledge
manipulation. In this experiment, we used the original and revised ver¬
sions of the Vietnam text of Britton and Gulgoz (1991). The goal of the
experiment was to replicate Britton and Gulgoz but at the same time to
investigate whether their results depended on a low level of background
knowledge, as we have argued. To achieve more variation in background
knowledge among our subjects, we pretrained half of our subjects. These
subjects received a brief, 20-minute history lesson designed to teach them
the knowledge that was needed to understand the original version of the
Britton and Gulgoz text - a few facts about the geography and history of
Vietnam and about the role of the United States and its South Vietnamese
allies in the war. Perhaps not surprisingly, this minilesson was not very
effective. There was a considerable overlap in scores on the preknowledge
test for subjects with and without pretraining. Hence, not all of our pre¬
trained subjects had become high-knowledge subjects, and the division
between high- and low-knowledge subjects had to be made strictly on the
basis of the knowledge pretest, without regard to the subject’s pretrain¬
ing condition.
The results of this study were similar for both recall and questions
(text-based questions and bridging inferences). Subjects who read the
revised, coherent text performed better than subjects who read the orig¬
inal text, and high-knowledge subjects did better than low-knowledge
subjects. There were no significant interactions. Thus, these results repli¬
cate both Britton and Gulgoz (1991) and Figures 9.10 and 9.11. The cru-
318 Models of comprehension

cial results, however, are the sorting data (no problem-solving questions
were given in this study), which are shown in Figure 9.14. Obviously, the
same kind of interaction between prior knowledge and the coherence of
the text was obtained as in McNamara et al. (1996) (Figure 9.13). For
low-knowledge subjects, the original low-coherence text is quite ineffec¬
tive, as is the revised, high-coherence text for high-knowledge subjects.
Flowever, if readers who know very little about Vietnam read the coher¬
ent, well-written text, their sorting is influenced by it. Similarly, if the
better informed reader^ are given the more challenging text, the text
clearly influences the way they sort the key words.
These results are paradoxical. Should we start writing incoherent texts
and give disorganized lectures so that our better students will benefit
from them? The answer to this question seems to be a qualified “yes.”
Making things too easy for a student may be a significant impediment to
learning. However, just messing up a lecture is not a solution. Instead, we
need to challenge the student to engage in active, deep processing of a
text. This can be done, as we have shown here, by placing impediments in
the path of comprehension, but impediments of the right kind and in the
right amount. They must be impediments we have reason to think the
student can overcome with enough effort, and the activity of overcoming

0.15

0.1 -

<D
S-H
O
o [n] High-Coherence
co
bD 0.05-
a
□ Low-Coherence
'6o
00

-0.05
High K Low K

Figure 9.14 Sorting scores as a function of prior knowledge for high- and
low-coherence texts. After McNamara & Kintsch (1996).
 Learning from text 319

these impediments must be relevant to learning. We have shown that giv¬

ing students a study text with coherence gaps, for which they do not have
adequate background knowledge, is self-defeating. Such students need all
the help they can get, and we need to organize and explicate the text for
them as well as we can. But, as the literature on generation effects in
memory also demonstrates (e.g., McNamara & Healy, 1995), students
who are able to perform a task unaided should be encouraged to do so.
They remember better and learn better to the extent that the activity they
are engaging in is task relevant. This was certainly the case in the experi¬
ments we have discussed in which the incoherent texts forced the high-
knowledge students to establish local coherence relations on the basis of
their own knowledge, to figure out the macrostructure of the text on their
own, and to elaborate the textual material with what they already knew. As
shown by the simulation described in section 9.4, that is exactly what is
required for learning from text.
The fact that not just any self-generated activity is helpful for text
memory has been shown in a series of studies discussed in McDaniel,
Blischak, and Einstein (1995). Story recall can be improved by omitting
occasional letters from words so that the reader must fill in the missing
information from the context. Filling in missing letters forces readers to
focus on the details of the stories that readers often disregard in favor of
the story line. This aids their recall of the story. On the other hand, a dif¬
ferent orienting task, such as reordering sentences, has no effect on story
recall because readers of a story pay sufficient attention to the relational
information between sentences, even without the reordering task. These
relations are reversed for descriptive texts. The reordering task helps
recall by focusing the reader’s attention on otherwise neglected order
information, but the missing letters task has no effect because readers of
essays are quite careful about the details anyway. T hus, simply placing
obstacles in the path of readers to force them to expend extra effort does
not benefit their learning. Instead, positive effects can be expected only if
the extra processing they engage in is appropriate to the task.
If the effort is appropriate to the task, however, engaging the reader in
active processing can be quite helpful. We have seen this for learning
from text in the experiments previously discussed. I his is also the case
for sentence memory. There is a curvilinear relation between the strength
of a causal connection between two statements and the memory strength
of their connection measured by cued recall, indicating that neither too
320 Models of comprehension

weak nor too strong links are elaborated as successfully as intermediate

links (e.g. Myers, Shinjo, & Duffy, 1987; van den Broek, 1990). Analo¬
gous results have been reported by Battig (1979) for paired-associate
learning, who reported better retention when the learning was made
more difficult through intratask interference. There is also a literature on
skill acquisition (Schmidt & Bjork, 1992) which shows that making the
learning process too smooth is counterproductive. Learners acquiring a
new skill must have the opportunity to face difficulties and learn to repair
mistakes. Novice skiers, certainly do not learn anything positive when
their friends take them down an expert slope without preparation, but
neither do they make much progress when they spend all their time on
the bunny hill.
Using texts with coherence gaps to stimulate deeper processing in
readers is but one technique that can be employed to stimulate reader
activity. It is not an easy technique to use, as we have seen, for its effec¬
tiveness depends on the match between the nature of the gaps in the text
and the reader’s background knowledge, which is not easily established.
Furthermore, other instructional or experimental manipulations may
override it, as was shown by E. Kintsch and W. Kintsch (1995). In this
study we used the same texts again as in McNamara et al. (1996) but asked
readers to comment on their understanding as they read each sentence.
Thus, we forced readers to be active processors and did not allow high-
knowledge readers to assume superficial processing strategies. Under
these conditions, the interaction between prior knowledge and text
coherence disappeared, and the high-coherence text was as good as or
better than the low-coherence text for all readers. However, when no
experimenter (or teacher) is available to ensure that the student processes
the text deeply, letting the student read a challenging text rather than one
too readily comprehended might still be a useful technique to ensure
reader activity.
Mannes and Kintsch (1987) have reported a study that is similar in
spirit to the experiments discussed here. In their study, readers were
given an advance organizer that either fit perfectly with the target text and
hence made it easy to read, or that structurally mismatched the target text
so that the readers had to engage in some cognitive effort to relate the
advance information and the target text. The target text was a rather long
article from a popular science publication about the industrial use of
microbes. Because our students knew very little about microbes before-
 Learning from text 321

hand, we prepared an advance organizer that told them everything they

needed to know about them. We prepared two forms of this advance orga¬
nizer that differed in the order in which the material was presented but,
as much as possible, not in the content of the material. In one case, the
general information about microbes occurred in exactly the same order as
it was presented in the target text. We called this the congruent advance
organizer. In the other version, the material was presented as in the ency¬
clopedia article that we had used as our source. This organization was
incongruent with the text.
Subjects studied either form of the advanced organizer and took a short
test on it. They then read the target text, which required a certain amount
of knowledge about the properties of microbes. After reading the text, they
were tested on what they remembered from it and what they had learned.
The data shown in Figure 9.15 are representative for the results of this
study. We observed a dissociation between measures of text memory (cor¬
rect verifications of old sentences in Figure 9.15) and measures of learn¬
ing (correct verifications of inference sentences in Figure 9.15). When the
advance organizer and the content of the text were structurally congruent,
subjects readily understood the text and remembered it better than when

0.6

0.5-

O 0.4-
<D
fc
O
U [7] Same Structure
§ 0.3.
□ Different Structure
c
o
P,
O 0.2

0.1 -

Old Inference

Figure 9.15 Proportion correct responses on a sentence verification task for

old sentences and inferences when the structure of an advance organizer and a
text were congruent (Same Structure) or incongruent (Different Structure).
After Mannes and Kintsch (1987).
322 Models of comprehension

the advance information and the text were organized differently. The for¬
mation of a textbase was presumably facilitated because the macrostruc¬
ture they had formed for the advance organizer provided a fitting schema
for the text itself. When this was not the case, the textbase was not as per¬
fect, and behavioral measures depending primarily on the textbase were
lower. However, their situation model was enhanced because in under¬
standing the target text, they had to retrieve and integrate information
from the advance organizer that had presented the information in a dif¬
ferent context. Therefore, a richer, more interrelated network was con¬
structed, which later helped the readers with inference questions and
problem solving (see also Mannes, 1994).
Making comprehension difficult for students is an effective technique
to foster learning but only up to a point. If we make comprehension so
difficult that students cannot succeed, nothing positive is achieved. E.
Kintsch (1990) showed this very nicely in a study comparing the ability
of sixth-grade students, tenth-grade students, and college students to
summarize a text. The text was a descriptive, comparing two countries on
a variety of dimensions. In one version, the text had a well-ordered and
clearly signaled macrostructure; in a second version, the content was the
same but was presented in a disorganized, though locally coherent way.
Figure 9.16 shows some of the results of this study. An ideal summary
should consist entirely or mostly of sentences corresponding to macro¬
propositions expressed in the text or inferred. By that standard, sixth-
grade students write pretty poor summaries, though the quality of the
summaries improves considerably with age. The important point is not
this general improvement, however, but the interaction between the age
of the students and the quality of the text they read. The college students
actually wrote better summaries when they received the poorly organized
text, paralleling the findings of some of the other studies reviewed here.
For the sixth-graders, however, the poorly organized text was too diffi¬
cult, such that they performed only about half as well with the poorly
organized text as with the well-organized text.
Thus, there is evidence now from a number of studies from our labo¬
ratory as well as elsewhere that increasing the difficulty of the learning
phase can have beneficial effects. Moreover, these studies are not limited
to learning from text but involve other kinds of learning situations as well.
Task difficulty can stimulate active processing, with the result that a more
elaborate, better integrated situation model will be constructed. As we
 Learning from text 323

Good Text

Poor Text

Age

Figure 9.16 The percentage of statements in the subjects’ summaries that

correspond to macropropositions as a function of age for the good and poor
text. After E. Kintsch (1990).

have seen, this statement must be carefully qualified. The student must
have the necessary skills and knowledge to successfully engage in the
required activity, and the activity must be task relevant.

9.7 Matching students with instructional texts

Studies like McNamara et al. (1996) and Voss and Silfies (1996) show how
important it is to provide a learner with the right text, not too easy and
not too hard. But how can that be done? To perform a theoretical analy¬
sis with the Cl model of both the student and the text as was done in sec¬
tion 9.4 is clearly impractical. Of course, good teachers and friendly
librarians who know their students and their books do so intuitively all
time, and often quite successfully. Our goal is to individualize learning,
yet teachers cannot always be there to help, nor can they know everything
that is available, and large electronic databases just don’t come with a
friendly librarian.
Kintsch (1994c) has hypothesized the existence of “zones of learnabil-
ity,” in analogy with Vygotsky’s “zones of proximal development” (Vy¬
gotsky, 1986). If a student’s knowledge overlaps too much with an
324 Models of comprehension

instructional text, there is simply not enough for the student to learn
from that text. If there is no overlap, or almost no overlap, there can be no
learning either: The necessary hooks in the students’ knowledge, onto
which the new information must be hung, are missing.
One way to determine whether a particular text falls within a student’s
zone of learnability is to represent both the student’s knowledge and the
text as vectors in the semantic space constructed by LSA and determine
how far apart they are (see section 3.3 on LSA). Wolfe, Schreiner, Rehder,
Laham, Foltz, Landauer, and Kintsch (in press) have investigated the
potential of this idea.
Wolfe and his co-workers used the heart as their knowledge domain.
They first assessed what average college students knew about the func¬
tioning of the heart and then developed a 40-point questionnaire to test
for this knowledge. This questionnaire was used both as a measure of the
students’ background knowledge and, after instruction, as a measure of
their learning. Specifically, if a student scored x points before instruction
and y points after, the proportion of improvement was used as a measure
of learning:

learriquest = (y ~ x)/(40 - x)

Another, independent measure of the students’ knowledge and learning

was obtained by having each student write a 250-word essay on the func¬
tioning of the heart, both before and after instruction. These essays were
graded by professional graders at the Educational Testing Service, and a
second learning measure, learnessay, was computed, based on the propor¬
tion of improvement between the grades on the before and after essays.
The length of the essays was constrained to avoid the difficulty of com¬
paring the content of essays of different lengths.
Students (the usual college sophomores) were randomly assigned to
one of four instructional conditions. Actually, the instruction the students
received was minimal; they were allowed to study a four-page text on the
functioning of the heart for 20 minutes. One cannot expect much knowl¬
edge change from such instruction, but we hoped that the different effec¬
tiveness of the four texts we used would become apparent nevertheless.
One of the texts (Text A) was from a junior-high textbook; we assumed it
would be easy for everyone and perhaps too easy for some of the students.
Texts B and C were from college level texts, with Text C being the harder
 Learning from text 325

one in our informal judgment. We assumed that these two texts would be
approximately appropriate for the majority of our subjects. The fourth
text, Text D, was from a medical journal and was definitely too hard for
most or all of our subjects. All four texts were concerned with the func¬
tioning of the heart, and although their contents were by no means iden¬
tical in detail, they were all intended to describe, at rather different levels
of sophistication, how the heart functions. We wanted to find out how well
prior knowledge predicts what students learned from these texts and
whether we could use LSA for these predictions.
Latent semantic analysis was used in two ways: to represent the four
instructional texts as vectors in the semantic space, and to represent the
essays each student wrote before as well as after instruction as vectors in
the same space. A new semantic space was constructed for this purpose
that dealt specifically with knowledge about the heart. Thirty-six articles
of an encyclopedia that dealt with the heart were used to construct this
space. A comparison of the cosines among the four text vectors confirmed
our intuitions that the texts were ordered from easiest to most difficult in
the order A < B < C < D. In each case, the cosine between nearer texts
were higher than the cosines between more distant texts.
Student scores on the two measures of knowledge, the questionnaire
and the essay, were well correlated, r - .74. Furthermore, if we assume
that Text B and C are standard texts, in the sense that their content is what
college sophomores should know about the heart, the cosine between a
subject’s pretest essay and one of these texts provides another knowledge
measure - how distant is that student’s essay from where it should be in
the LSA space? The fact that this third, LSA-derived measure of prior
knowledge also correlated quite highly with either the question-based or
the essay-based knowledge measures shows that we are dealing with a
rather orderly data set and that LSA may be the right measuring stick for
our purposes. Specifically, the correlation between the cosine for a stu¬
dent’s pretest essay and Text C and scores on the questionnaire and essay
grades were r = .64 and .69, respectively.
The main questions of interest were, however, whether prior knowl¬
edge predicted learning and whether there was any evidence for a non¬
monotonic relationship, as implied by the hypothesis of the zones of
learnability. According to that hypothesis, the relationship between prior
knowledge and learning should depend on the difficulty of the text.
Specifically, students should not learn much from the too-easy Text A
326 Models of comprehension

unless they really know nothing about the heart; nor should they learn
much from the too-hard Text D unless they already know a lot.
We have a restricted range problem in this study, however. All our col¬
lege sophomores knew a little bit about the heart, but there were very few
really high-knowledge people, so we really did not have a text that was too
easy. We therefore ran a fifth group of subjects, a group of medical school
students, who were given the easy Text A to study, d he results are shown
in Figure 9.17. Figure 9.17 shows the mean improvement in learnquest and
learnessay as a function of the average cosine between the texts that stu¬
dents read and their pretest essays. Including the medical students in
Figure 9.17 clearly brings out the hypothesized nonmonotonic relation¬
ship between prior knowledge and amount learned. It is particularly
noteworthy that two very different methods of measuring learning - by
questionnaire and by essay grades — yielded comparable results.
According to Figure 9.17, a great deal can be gained by assigning stu¬
dents the right text for learning. We could easily have doubled the learn¬
ing scores for the college students who read Text D, as well as improved

□ leam quest

O leamessay

Figure 9.17 Mean learning scores as assessed by questionnaire and essay

grades as a function of mean essay-to-text cosine scores for the four experi¬
mental groups (Texts D, C, B, A) and a group of medical school students
(Aned)- After Wolfe et al. (in press.)
 Learning from text 327

the scores of students who read Text A, simply by giving them either of
the other two texts! Wolfe et al. (in press) made some calculations that
suggest that by assigning individual students to texts on the basis of their
background knowledge, learning scores could have been improved by
more than 50%. Even if in practice these calculations turn out to be opti¬
mistic, such a potential can hardly be ignored.
The complex interaction between learning, text characteristics, and
background knowledge certainly deserves more research. However,
Wolfe et al. (in press; see also Rehder, Schreiner, Wolfe, Laham, Lan-
dauer, & Kintsch, in press, for a more detailed analysis of the LSA meas¬
ures that can be used to predict learning) have shown one thing quite
clearly: LSA predicts learning about as well as the more cumbersome
empirical measures (essays have to be graded, questionnaires require
careful design), thus giving us a powerful research tool and perhaps even
a practical tool that could be used to match students with the texts they
could most profit from.

9.8 Educational implications

Learning from text, like all learning, involves the construction or modifi¬
cation of situation models. Hence, the first educational implication of the
work discussed here is that if we are interested in learning, we must make
sure that our measures are sensitive to learning — that is, they must reflect
properties of the situation model rather than merely the textbase. Such
measures are inference questions and problem-solving tasks, and, at least
for laboratory studies, keyword sorting and cued association tasks. Recog¬
nition and fact questions, in contrast, may merely reflect understanding at
the level of the textbase. Recall and summarization are intermediate in this
respect because reproductive recall can be quite good even in the absence
of an appropriate situation model, whereas the reconstructive component
of recall can be more indicative of the nature of the learner’s situation
model. In the studies we have reviewed, textbase and situation model mea¬
sures frequently gave different results. Failure to distinguish between
measures of learning and measures of memory invites theoretical confu¬
sion and cannot produce educationally satisfactory results. Many educa¬
tors have stressed the dangers of superficial understanding and the need
for engaging students in deeper understanding. I hope that by tying the
commonsense terms deep and superficial understanding into the theory of
328 Models of comprehension

comprehension, and in particular to the concepts of textbase and situation

model, I have been able to clarify the nature of these concepts and to
sharpen their meaning.
The emphasis in this chapter has been on the formation of situation
models that are well integrated with the learner’s prior knowledge. If this
integration is not achieved and if the information acquired remains
accessible only via the episodic text memory, two problems arise. First,
such information represents inert knowledge, in the sense that although
it is possible to retrieve at as part of an episodic memory structure, it can¬
not be retrieved when it is needed for the performance of new tasks, for
by the principle of encoding specificity, memory traces can be retrieved
only via retrieval cues that were encoded when the traces were formed.
Second, memory structures not linked to a person’s base of knowledge
and experiences are rapidly forgotten, because the only effective retrieval
cues, which often change quickly, are of a temporal and contextual nature.
If situation models with rich and stable links to prior knowledge are the
goal, how is such a result best achieved? The answer depends on condi¬
tions. First, it is necessary that the reader/learner be able to form a coher¬
ent mental representation of the text. This is a necessary but not suffi¬
cient condition for learning. A well-written text, in which the local
coherence relations are marked syntactically and lexically, and in which
the macrostructure is well organized and clearly signaled, usually guar¬
antees this result. A poorly organized text with coherence gaps may still
be understandable but only if the learner can rely on background knowl¬
edge that enables him or her to overcome the deficiencies of the text.
Nevertheless, a coherent text representation is not sufficient for learn¬
ing; it must be linked to prior knowledge in as many ways as possible. The
only way to encode such links with prior knowledge during reading is for
the reader to use prior knowledge in building a situation model. Here the
limitations of a well-written text become apparent. If a reader has no
background at all to link with a new text, a hospitable text may result in
text memory in the sense we have discussed but not in learning. To learn
something new we must have some hooks in long-term memory to hang
it on.
The fact that a reader has available appropriate background knowledge
does not in itself guarantee learning, for the knowledge must be used in
the processes of understanding. Otherwise, no links will be generated
between it and the new text. As we have seen, this is not always the case.
 Learning from text 329

Readers with appropriate knowledge do not always bother to employ that

knowledge for learning. Readers also tend to take the path of least resis¬
tance, and if they have the feeling that they are easily understanding the
text they read, they may not bother to activate their knowledge and form
the links between it and the text that alone guarantee learning. Hence,
there exists an instructional need to stimulate reader activity.
In some of our experiments, we encouraged readers to be active by giv¬
ing them a less than perfect text. Because this text had both local and global
coherence gaps, readers could not understand it without using their
knowledge. High-knowledge readers, therefore, brought their knowledge
to bear on the interpretation of the text and became successful learners.
The same readers could also have done so when they read a well-written,
explicit, and coherent text, but they did not feel the need to do so.
There are many other ways to make readers active learners. Students
who have their own learning goals, who own their questions and are not
merely responding passively to what they perceive as arbitrary and irrel¬
evant teacher demands, tend to be active learners who are not satisfied
with superficial understanding. Dewey (1897), for this reason, demanded
long ago that all learning should be incidental, part of a meaningful activ¬
ity in which the student engages. Many current instructional programs
have adopted this strategy (e.g., Bereiter & Scardamalia, 1989). Others
have devised specific teaching strategies that ensure the active participa¬
tion of the learner, such as the reciprocal teaching method of Brown and
Palincsar (1989), or that are specifically suited for learning from text, the
questioning-the-author method used by Beck, McKeown, Worthy, San¬
dora, and Kucan (1996). There are many ways to engage the reader. The
research discussed here shows why we need them and what we want to
achieve by using them.

A main theme of this chapter has been the need to distinguish

between memory for a text and learning from a text. It is possible
to reproduce a text from memory without being able to use it for
any other purpose; the information provided by the text remains
inert knowledge. Only if the episodic text memory is intimately
intertwined with the reader’s long-term memory structure will it
be accessible in novel situations, become a part of the reader’s
active knowledge. For school learning, for instance, it does not
make much difference whether the student can reproduce the
330 Models of comprehension

textbook from memory. What is important is that the information

in the book become integrated with the student’s prior knowl¬
edge. For a long time, studies of text comprehension did not pro¬
vide information on the factors that determine the success or fail¬
ure of this integration process. We could tell anyone who wanted
to know how to make a text more memorable and easier to com¬
prehend, but that was not an answer that was of much use.
Authors of textbooks, training programs, instructional manuals,
and the like are not concerned with text memory, but they want to
get a certain body of information across to the student. Until a few
years ago, we really had very little to offer such people.
The research results reviewed in this chapter clearly show that
this situation has greatly improved; we have learned a lot about
learning from text in recent years. It is not a simple story, but the
primary factors in learning from text are reasonably clear at this
point. A major determinant of learning from text is background
knowledge. To learn effectively, we need hooks in prior knowl¬
edge, long-term memory, or personal experience on which to
hang the information to be learned. Learning is most successful
when such hooks are plentiful and when there is a clear relation
between the hooks and the learning material so that the student
hangs things on the right hooks.
Although the existence of such hooks in prior knowledge is a
necessary condition for learning, it is not a sufficient condition.
It is not enough that a hook in long-term memory is available — it
also must be used. Hence, the importance of making learners
active, intentional agents rather than passive vessels into which
information is poured at will. It is clear that if learners have an
adequate background, they will perform better if they take an
active role in their learning - making inferences, filling gaps,
generating macrostructures, elaborating, and the like. It has also
been shown that at least in some cases learners with an adequate
background knowledge will perform below their capabilities if
they do not assume such an active role. How does one create an
active learner? We have used here for the most part the somewhat
counterintuitive technique of providing the learner with a poorly
written text, a text that lacks coherence, thereby forcing the
learner to fill in the gaps in the text. Further research will be
Learning from text 331

needed to determine more closely the boundary conditions

for the effectiveness of this strategy or of the other strategies
designed to induce active learning that have been used success¬
fully in recent years in text research. Even now, however, this is
an area of research that seems ready for the step from the labora¬
tory to practice.
 10

Word problems

People spend a great deal of time reading, some much more than
others. The amount of time spent in leisure reading, according to
one study (Guthrie & Greaney, 1991), is 63 minutes per day, 39
minutes for newspapers and 24 minutes for books. These times
are medians of highly skewed distributions and they are probably
underestimates because a substantial part of leisure reading is sec¬
ondary to some other activity (cooking, planning a vacation, wait¬
ing in a doctor’s office) and hence is underreported. In addition,
the median estimated time people spend in occupational reading
amounts to another 61 minutes (much more for professional
workers). Once again, it is hard to get accurate reports, because
so much of occupational reading is embedded in another activ¬
ity. Scanning a production schedule tends not to be reported as
reading.
It is clear from these numbers that reading takes a consider¬
able portion of our time, even if we live outside a university and
even if much of the reading we do occurs in the service of some
other activity. Comprehension, memory, and learning are the key
concepts when we look at reading for its own sake for relaxation
or acquiring knowledge. But we need to broaden our approach
when looking at reading in the service of other goals, such as
solving a mathematical problem or performing some action.

Word problems are popular in arithmetic and algebra classes, as well as

in physics and engineering courses, because they lend an air of realism
and authenticity to the drill that is necessary for the acquisition of formal
 Word problems 333

skills. Most students profess to hate them and find them confusing. They
know the math but not whether to add or subtract or what to multiply
with what. Teachers are challenged to make their students understand
what they are doing, instead of applying rules mechanically.
To the psychologist, word problems provide some wonderful research
opportunities. Not only are they texts, so that one can recall and summa¬
rize them just like stories, but one can also observe the student’s problem
solving and the eventual answer. What the student remembers and what
the student does are related in informative ways and mutually constrain
each other. Thus, a richer set of experimental observations can be obtained
than in other text research.

10.1 Word arithmetic problems

Word problems play a significant role in the teaching of arith¬

metic in schools all over the world. They are used very differ¬
ently, however, in different countries. An illuminating compari¬
son of Soviet and U.S. elementary mathematics textbooks for
grades 1 to 3 was reported by Stigler, Fuson, Ham, and Kim
(1986). In the books that they analyzed, the Soviet texts con¬
tained somewhat more word problems than did the U.S. texts,
but this difference was minor compared with the striking differ¬
ences in the nature of the word problems and how they were pre¬
sented. Across the different types of word problems (e.g., Table
10.1 on page 335) the Soviet texts maintained an approximately
even distribution. Each type of word problem was represented
equally often, and problem types were mixed up and inter¬
spersed among numerical problems, so that most pages con¬
tained at least some word problems. In contrast, the American
texts focused on a few problem types. A single Compare problem
and two kinds of Change problems are used almost exclusively;
among the Combine problems, the missing-whole version is
three to ten times as frequent as the missing-part version. That
is, most of the word problems found in U.S. texts are the ones
that U.S. children find easiest. Furthermore, problems of the
same type tend to be blocked, and word problems are generally
segregated from the rest of the material into special problem¬
solving sections at the end of chapters.
334 Models of comprehension

Most efforts to study how children learn to solve word arithmetic prob¬
lems have focused on the formal aspects of the problem. Thus, the first
models of word arithmetic problem solving by Riley, Greeno, and Heller
(1983) and Briars and Larkin (1984) had no explicit language-processing
component and instead dealt with only the problem representations the
children constructed, the inferences they made, and the counting opera¬
tions they used. However, the language used to express a problem makes
a difference. A well-known example of how important the language can be
has been provided by Hudson (1983). Only 39% of Hudson s first-grade
students were able to solve the following problem:

(1) Joe has 8 sticks.

He has 5 more sticks than Tom has.
How many sticks does Tom have?

However, the arithmetically identical problem

(2) There were 8 birds and 5 worms.

How many birds did not get a worm?

was solved by 79% of the children. Nevertheless, focusing first on the

formal properties of word problems was a good research strategy. Once
the formalisms were understood, problem-solving models could be
merged with text comprehension models to yield complete theories of
word problem solving. Thus, the Riley et al. (1983) theory of arithmetic
problem solving was combined with the van Dijk and Kintsch (1983) text
comprehension model by Kintsch and Greeno (1985).

10.1.1 Schemas for solving word arithmetic problems

Kintsch and Greeno (1985) argued that word problem solving involves a
text understanding phase that yields the kind of mental representation
needed by the second problem-solving phase to compute an answer.
Many texts have a conventional structure that readers know and use to
organize their mental representation of the text. Thus, stories are orga¬
nized by the story schema and legal briefs by an argumentation schema.
Word problems have their own schemas that have to be learned in school,
recognized as relevant when the problem is being read, and used as an
 Word problems 335

organizational basis for the text. The schemas are the structures identi¬
fied and described by Riley et al. (1983): the Transfer schema, the Part-
Whole schema, and the More(Less)-Than schema in their various ver¬
sions. They are all based on the notion of a set. A set is a schema, too, a
frame with slots to be filled by particular quantities and objects:

SET
object:<count noun>
quantity:<number, some, many>
specification: <owner, location, time>
role:<start, transfer, result;
superset, subset;
large-set, small-set, difference-set>

Thus, the sentence Joe had 5 marbles yields a set with Joe as the owner, 5
as the quantity, and past as specification. Its role depends on the problem
context. For example, it might be the start-set of a Transfer problem (He
gave away 3 marbles. How many marbles does he have now?), or the subset
of a Combine problem (Jack had 3 marbles. How many do they have alto¬
gether?), or the large set of a Compare problem (He had 2 more marbles
than Jack. How many marbles did Jack have?)
Riley, Greeno, and Heller (1983) classified word arithmetic problems
into 14 types in three main classes, as shown in the following list:

Table 10.1. Types of word problems

Change Combine Compare

Result unknown Superset unknown Difference unknown

1. Joe had 3 marbles. 1. Joe has 3 marbles. 1. Joe has 8 marbles.
Then Tom gave him Tom has 5 marbles. Tom has 5 marbles.
5 more marbles. How many marbles How many marbles
How many marbles do they have does Joe have more
does Joe have now? altogether? than Tom?

2. Joe had 8 marbles. 2. Joe has 8 marbles.

Then he gave 5 Tom has 5 marbles.
marbles to Tom. How many marbles
How many marbles does Tom have less than

does Joe have now? Joe?

font.)
336 Models of comprehension

Table 10.1. Types of word problems [cont.)

Change Combine Compare

Change unknown Subset unknown Compared quantity unknown

3. Joe had 3 marbles. 2. Joe and Tom have 8 3. Joe has 3 marbles.
Then Tom gave him marbles altogether. Tom has 5 more marbles
some marbles. Joe has 3 marbles. than Joe.
Now Joe has 8 How many marbles How many marbles does
marbles. - does Tom have? Tom have?
How many marbles
did Tom give him?

4. Joe had 8 marbles. 4. Joe has 8 marbles.

Then he gave some Tom has 5 marbles less
marbles to Tom. than Joe.
Now he has 3 How many marbles does
marbles. Tom have?
How many marbles
did he give to Tom?

State unknown Referent unknown

5. Joe had some marbles. 5. Joe has 8 marbles.
Then Tom gave him 5 He has 5 more marbles than
more marbles. Tom.
Now Joe has 8 How many marbles does
marbles. Tom have?
How many marbles
did Joe have in the
beginning?

6. Joe had some marbles. 6. Joe has 3 marbles.

Then he gave 5 He has 5 marbles less than
marbles to Tom. Tom.
Now Joe has 3 How many marbles does
marbles. How many Tom have?
marbles did Joe have
in the beginning?

Source: After Riley, Greeno, and Heller (1983).

 Word problems 337

These are the kinds of problems children are given in kindergarten to

third grade. They differ greatly in difficulty, with most children being
able to solve the easiest types of problems, such as Change-1 or Combine-
1, whereas only a small minority of American school children can solve
correctly the more unusual and more difficult problems, such as Com¬
pare-3.
The Kintsch and Greeno (1985) model understands these problems by
creating various sets from the information supplied by the text and using
one of its higher-order schemas to interrelate these sets. Every time the
model encounters a noun phrase with a quantifier it creates a set and
attempts to fill as many slots of the set schema with information from the
text as possible. It then uses semantic information in the text to select one
of its higher-order schemas. A phrase containing more-than, for instance,
cues the Compare schema, and a transfer verb (give, take, lose, etc.) cues
the Transfer schema. If one child has so many cars and the other so many
dolls and the question asks for the number of toys, toys is the superset and
cues the appropriate schema.
Consider how the model understands and solves a Change-3 problem:

(3) John had 3 marbles.

Then Tom gave him some more marbles.
Now Joe has 8 marbles.
How many marbles did Tom give him?

• The 3 marbles in the first sentence provide the cue for making a set Sj
of 3 marbles, with owner Joe in the past, role unknown.
• Similarly, a second set is formed on the basis of the next sentence,
the objects being marbles, quantity some, and specification that Joe is
the owner, at a later time. Furthermore, give cues the Transfer schema,
with the result that S, can be assigned the role of start-set and S2
the role of transfer-set. An expectation is created for the missing result-
set.
• It is filled by S3, which is created on the basis of the third sentence: 8
marbles, owned by Joe now, the result of the transfer process.
• The final sentence merely asks to calculate the as yet unspecified quan¬
tity of S2. This is done by using a counting procedure called Add-On,
as is typical for first-grade students.
338 Models of comprehension

A More-Than schema is used to solve a Compare-5 problem:

(4) Joe has 8 marbles.

He has 5 more marbles than Tom.
How many marbles does Tom have?

• Sp objects: marbles
quantity: 8
specification: owner Joe
role: unknown
• S2: objects: marbles
quantity: 5
specification: owner Joe
role: difference-set

S, is assigned the role of large-set; a request is made for a small-set.

• S3: objects: marbles

quantity: ?
specification: owner Tom
role: small-set

The solution procedure in this case is complex; first-graders do not

appear to have a solution strategy for this type of problem. Therefore,
they convert the More-Than schema into a Part-Whole schema with Sj
as the superset, S2 as a subset, and S3 as the goal set for solution by a pro¬
cedure called Separate-From. (These procedures are modeled after
empirical work by Carpenter and Moser, 1983.)
The Kintsch and Greeno (1985) model can be thought of as a produc¬
tion system with smart rules. It differs in that respect from the Cl model,
because it uses cues in the text to select the right schema to organize a
problem. Because only three schemas are needed for children’s word
problems, this poses no insurmountable difficulties. I discuss below, how¬
ever, the Cl version of the model proposed in Kintsch (1988), which
explores contradictory hypotheses about the role of sets in parallel and
uses an integration process to select the one that fits a problem best.
Computer simulations based on the Kintsch and Greeno (1985) model
were constructed by Fletcher (1985), Dellarosa (1986), and Cummins et
al. (1988). 1 hese simulations serve several purposes. First, they show
 Word problems 339

that the model actually works as originally conceived. All 18 problem

types are solved successfully by the simulation, as well as a number of
richer story problems described in Cummins et al. (1988). Second, these
simulations were able to account reasonably well for the observed prob¬
lem difficulty in terms of the nature and number of the operations the
simulations had to perform. Difficult problems tend to make greater pro¬
cessing demands and require less familiar operations and schemas (e.g.,
the More-Than schema is less familiar to kindergarten to third-grade
children than the Transfer schema is). Most interesting, however, is the
ability of the simulation to account for the kinds of errors the children
make in solving these problems. The way it does so provides some real
insight into why word problems are so hard.

10.1.2 Error simulation

What is it that children learn when they solve word arithmetic problems?
According to the logicomathematical development view, which has its
roots in the work of Piaget, children learn to solve certain kinds of prob¬
lems by acquiring the conceptual knowledge necessary for these prob¬
lems. This is the view taken by, among others, Riley et al. (1983) and Bri¬
ars and Larkin (1984). Riley et al. (1983), for instance, hypothesize that
the problem solving of good students involves complete schemas such as
those we have described, whereas poor problem solvers work with impov¬
erished schemas that represent individual sets but not the relations
among the sets.
Alternatively, one might hypothesize that the children have the neces¬
sary formal knowledge but that certain difficult problems employ lin¬
guistic forms that are unfamiliar to the children. This linguistic develop¬
ment view contrasts with the logicomathematical hypothesis because it
places the source of difficulty not in the children’s deficient conceptual
structures but in their lack of linguistic knowledge. The child may have
an adequate understanding of the Part-Whole schema, for example, but
might not know how to map a phrase like How many more Xs than Ys?
onto that schema.
Cummins et al. (1988) have shown that one can decide among these
competing hypotheses by considering the way errors occur in the simu¬
lation and comparing these errors to the ones children make. 1 he trick is
340 Models of comprehension

to lesion the simulation, to take out certain processing components or

pieces of knowledge, and see how it behaves. If a lesion makes no differ¬
ence, the piece obviously was superfluous in the first place. If a lesion
leads to a catastrophic breakdown, we know that the piece in question is
vital because catastrophic breakdowns do not (or only very rarely) occur
in our data set. The children always do something and say something,
though it might not be what the problem demands. Indeed, the children
in one of Cummins’s experiments produced errors on 45% of all trials,
but most of these errors were systematic - only 8% of the trials yielded
uninterpretable errors. Thus, we should examine the kinds of errors the
children make and determine what we have to do to our simulation so that
it produces such errors.
It typically happens that when a child is unable to solve a problem, he
or she solves a simpler problem that is within the child’s abilities. The evi¬
dence for this claim lies in the systematic correspondences that we
observed between the children’s recall of a word problem and their solu¬
tion behavior. What they were doing did not resemble random errors -
there clearly was method in their errors!
Interestingly, the major types of systematic errors that we found in our
data can be reproduced by the simulation by means of certain kinds of
lesions. The simulation contains two separate sources of knowledge: lin¬
guistic and world knowledge that allows it to understand a word arith¬
metic problem, and mathematical knowledge that is necessary for the
solution of the problem. Any lesions in the mathematical knowledge -
that is, primarily the notion of what a set is, the higher-order schemas
relating sets, and the counting strategies children use - produce cata¬
strophic failures in our simulation. The system breaks down or produces
a weird result. That is not what the children do. However, some lesions of
the linguistic knowledge that is used in the simulation produce precisely
the kind of errors that were observed in the children. The difficulty seems
to be in certain key words and phrases that are either unfamiliar to the
children from their everyday experience or that are used in a special
restricted way in word arithmetic problems. The data are described in
detail in Cummins et al. (1988); I present here only a few typical exam¬
ples for illustration.
Combine-2 problems are surprisingly difficult for first-graders. Errors
occurred on 32% of all trials. The most frequent error was that the chil-
 Word problems 341

dren turned the (difficult) Combine-2 problem into an (easy) Combine-1

problem. As an example, consider the following:

(5) Mark and Sally have 7 trucks altogether.

Mark has 2 trucks.
How many trucks does Sally have?

The correct propositional parse for the first line is

HAVE[AND[MARK,SALLY], SEVEN[TRUCKS]].

The problem for the children is the interpretation of the and-altogether.

As DeCorte, Verschaffel, and DeWinn (1985) have shown, children tend
to substitute each for and-altogether. If we lesion the simulation so that it
no longer knows the correct meaning of and-altogether, it parses the first
line of the problem as

HAVE[MARK, SEVEN[TRUCKS]] & HAVE[SALLY,

SEVEN[TRUCKS]]

As a consequence, the system will create 2 sets of 7 trucks - one for Sally
and one for Mark. The quantity of trucks in Mark’s set gets updated to 2
by the second sentence, and the problem becomes a superset problem. In
answer to the question, the simulation uses its “Already-know-the-
answer” strategy and responds with 7.
Real children make this sort of error 45% of the time. The children
also recall the problem as a Combine-1 problem:

(6) Sally has 7 trucks.

Mark has 2 trucks.
How many trucks do Mark and Sally have?

This type of misrecall occurs on 74% of the trials on which the children
made a superset error, whereas it was rare when other types of errors
occurred, and it did not occur at all when the problem was solved cor¬
rectly. Figure 10.1 shows the recall data, including a control group of sub¬
jects w4io read the problem but did not solve it.
Compare-5 problems were also a difficult problem type for our sub¬
jects. Errors occurred in our data again on 32% of all trials; when an error
occurred, either the difference set was given as the answer (32%), or the
two numbers in the problem were added (44%). The simulation yields
342 Models of comprehension

80

60 -

qj
ofl
3 vXvXxXx:
§ 40 -

0. ::::X:X;.;.;.;:;.;

2° -

o -1..i. J
Read Only Correct Superset Error

Figure 10.1 Misrecall of subset problems as a superset problem.

both of these error types if we lesion a key piece of its linguistic knowl¬
edge. Have-more-than requires a complex linguistic construction, involv¬
ing three sets: a large-set, a small-set, and a difference-set. If it is simply
parsed as more-than, the simulation makes the same errors that the chil¬
dren do:

(7) Joe has 8 sticks.

He has 5 more sticks than Tom.
How many sticks does Tom have?

The correct parsing of the second line is

HAVE-MORE-THAN[JOE, TOM, FIVE[STICKS]].

With the lesion, the model parses this line as

MORE-THAN[[EIGHT[STICKS],FIVE[STICKS]]
1— OF-JOE I—OF-TOM

This incorrect parsing does not yield a well-structured problem, but nei¬
ther the children nor the simulation is at a loss. The children use one of
their default strategies to come up with a reasonable answer. Two of these
default strategies apply: the “Already-know-the-answer” strategy yields
 Word problems 343

the difference set error - Tom has 5 sticks. The “Addition” strategy based
on the keyword more yields the addition error - Tom has 13 sticks. These
two errors in fact accounted for 76% of all errors the children made. Fur¬
thermore, the way the children recalled the problem correlated highly
with the way they solved it. If they made an addition error, the most fre¬
quent recall pattern was

(8) Joe has 8 sticks.

Tom has 5 more sticks.
How many sticks do they have?

Figure 10.2 shows the frequency with which this pattern of recall was
observed for Compare-5 problems when subjects made an addition error,
when they solved the problem correctly, and when they only read the
problem without solving it in a control condition.
The subjects’ tendency to simplify problems and solve the simplified
problem rather than the original one does not always result in errors. For
instance, the second line of the Compare-5 problem exemplified in (7) is
rather clumsily worded. The real structure of the problem is expressed
much better by Tom has 5 sticks fewer than Joe. Here the fewer directly keys
subtraction, and the owner of the set is kept in subject position. The

Figure 10.2 Misrecall of a compare problem as an addition problem.

344 Models of comprehension

Read Only Correct Superset Error

Figure 10.3 Misrecall of a Compare-5 problem as a Compare-4 problem.

structure of the problem remains the same, however, by this transforma¬

tion; it is simply better formulated because in effect a Compare-5 prob¬
lem is turned into the easier Compare-4 type. As Figure 10.3 shows, this
kind of structure transformation occurs quite frequently when the prob¬
lem is solved correctly and even when it is only read without attempting
a solution, whereas it is rare when an error occurs.
Problem solving and problem recall are intimately intertwined in these
data. Recall is reconstructive; the subject recalls what he or she did - not
necessarily what was on the page. This does not look like a language¬
processing module impervious to outside influences a la Fodor and
Chomsky. Flow the problem is solved depends on how the problem lan¬
guage is understood and, correspondingly, how it is recalled depends on
how it was solved. The same structures are the bases for both language
understanding and problem solving.
If the Cummins et al. (1988) simulation was missing certain pieces of
linguistic knowledge, it succeeded in reproducing the most frequent
error patterns that we observed in our first- and second-grade corpus. If
it lacked mathematical knowledge, it did not produce the kinds of errors
the children made. Therefore, the considerable difference that exists in
the children’s’ ability to solve word and equivalent numerical problems
 Word problems 345

appears to be due mostly to linguistic factors. Certain words and phrases

that play an important role in word problems are not part of the children’s
everyday vocabulary, and their meaning and use must be learned, or
relearned, specifically in school.
Data that further support these conclusions have been reported by
Stern (1993a) and Cummins (1991). Cummins also showed that by suit¬
ably rewording word problems to avoid common linguistic ambiguities,
performance on word problems can be significantly improved.
This- is not to say that mathematical knowledge plays no role. On the
contrary, it plays a crucial role; neither the simulation nor the students get
anywhere without it. Rather, what makes word arithmetic problems espe¬
cially difficult is the lack of language skills that would allow the child to
represent correctly certain relationships among sets. Indeed, numerical
competence is much more strongly related to the ability to solve word
problems than is verbal intelligence in a large-scale longitudinal study
reported in Stern (1993b). For instance, in a causal path analysis, verbal
intelligence in grade 3 was related to different types of word problem
solving in grade 4 with coefficients varying between .16 and .24. On the
other hand, numerical competence in grade 2 showed a much stronger
relation to word problem solving in grade 4, with coefficients between .31
and .41. The sine qua non of word problem solving is certainly numeri¬
cal competence, but it is not the only factor and by itself is no guarantee
of success in word problem solving. The linguistic factors discussed here
matter, too. Furthermore, the rather low correlations betweeen verbal
ability and word problem solving observed by Stern (1993b) may be
slightly misleading because the linguistic skills that matter in this respect
are specific to mathematics and may not be adequately assessed by a
general verbal intelligence test.

10.1.3 Situation models

Word problems do not have to be as simple and impoverished as those

studied above (and those used in most American schools). Cummins et al.
(1988, experiment 2), in addition to simulating the standard problems
also worked with a set of sixteen complex story problems. Formally, these
problems could be reduced to one of the standard types, but they were
presented in the context of a little story with a familiar, everyday scenario,
346 Models of comprehension

with real goals and motivations for the actors. The simulation that was
developed was also given the new problem set to solve. Apart from telling
it the meanings of several new words, the simulation was not changed at
all. Nevertheless, it managed to solve thirteen of the complex problems
correctly. It also accounted quite well for the kinds of errors children
made in solving these problems when the simulation’s linguistic knowl¬
edge was degraded. Degrading its mathematical knowledge, in contrast,
yielded errors quite unlike those the children made. The simulation failed
to solve three of the problems. The reason for this failure was the same in
each case: It did not have enough linguistic or general world knowledge
to understand the situation described in the story. In one case, for
instance, it failed to connect a reference to some pencils a boy gave his sis¬
ter with a later reference the 5 pencils he gave her. In another case, it did
not know that trade implies a switch of ownership.
What we see here is the central importance of situational understand¬
ing in word problem solving. The reader must have a correct and com¬
plete situation model of the various actions, objects, and events described
in a story problem. In the textually simplest arithmetic problem, this
amounts to little more than learning how to use correctly a few key words
and phrases. As problem texts get richer, however, all kinds of linguistic
and world knowledge may be required for the construction of an adequate
situation model. Familiarity with the overall domain becomes very
important, as we have already seen from the example of Hudson (1983).
Stern and Lehrndorfer (1992) provide some nice experimental demon¬
strations that problems of a certain type can be solved by children at a cer¬
tain age when the problems are embedded in familiar contexts, but not
otherwise.
Reusser (1989) has stressed most emphatically the importance of situ¬
ation models in word problem solving. Reusser developed a simulation
model that extends the Cummins et al. (1988) model by building in an
explicit situation model. This was a significant innovation in the evolu¬
tion of such models. In the past, theorists of word problem solving
attempted to go directly from the text to the equation (and instructors
stressing key-word strategies followed suit). Kintsch and Greeno (1985)
and Cummins et al. (1988) presented a more complex model in which text
understanding is interwoven with the construction of a mathematical
problem representation (in terms of sets and their interrelations), from
 Word problems 347

which equations and calculation procedures are derived. Reusser (1989)

emphasizes the importance of another layer of representation intermedi¬
ate between the textbase and the mathematization of the problem: the sit¬
uation model that specifies actors, actions, states, and events in the prob¬
lem in terms of everyday concepts. It is necessary to constrain the formal
problem model; if the situation model is incorrect because of some lin¬
guistic ambiguity or because of an unfamiliar domain, it is unlikely that a
correct problem model will be constructed.
The inclusion of a situation model allows Reusser’s simulation to solve
complex story problems. It also has proved to be a rich source of peda¬
gogic implications, both for teaching word arithmetic (e.g., Staub &
Reusser, 1995; Reusser, 1993) and in our own work on word algebra prob¬
lem solving, which is described in the next section.

10.1.4 The Cl model for word problem solving

The Kintsch and Greeno (1985) theory and the simulations based upon
it (Cummins et al., 1988; Reusser, 1989) are top-down schema theories.
Problems are understood because cues in the text lead to the activation of
an appropriate arithmetic schema around which the text is organized and
on which solution procedures can be based. If the right schema is acti¬
vated, the problem will be solved correctly. This model has the same dif¬
ficulties as all top-down schema theories, in that the cues that tell a stu¬
dent which arithmetic schema is needed may be subtle, complex,
multiple, unreliable, and even contradictory. Instead of jumping to per¬
haps premature conclusions and committing oneself to a single hypothe¬
sis about the nature of a word problem, it might be better to consider all
the evidence for several alternatives in parallel in the context of the story
as a whole. That is what the Cl architecture is designed to do, and Kintsch
(1988) has shown that it is quite possible to embed the Kintsch and
Greeno theory within the Cl framework.
Which version of the theory is the better one, however? As we have just
seen, quite a bit of productive work could be done with the top-down ver¬
sion. Nevertheless, there are reasons to prefer the Cl architecture in this
case, too. If one just looks at the end result, both versions of the theory
achieve identical outcomes. They differ in the way a schema is selected.
In the earlier version, the organizing schema is selected on the basis of a
348 Models of comprehension

cue or cues in the text, which then controls the rest of the process in a
top-down fashion. In the Cl version, evidence for various alternative
schemas is collected in parallel, and the schema that best satisfies the mul¬
tiple constraints of a problem eventually wins out. I he fact that there are
multiple or even contradictory cues is easily incorporated in the Cl ver¬
sion of the theory. Thus, experimental evidence that during the course of
reading a word problem subjects consider more than one alternative, and
that they combine multiple cues, would support the Cl version of the
theory and contradict or at least require a revision and elaboration of the
single-schema version.
Such evidence has been obtained in an experiment by Lewis, reported
in Kintsch and Lewis (1993). Lewis asked subjects to make a choice
between two alternative hypotheses about a problem, addition or sub¬
traction, at several query points during reading the problem. Thus, the
time course of hypothesis formation could be recorded. The problems
she used contained two cues, the overall structure of the problem and a
keyword favoring either addition or subtraction.
Prototypical problems used by Lewis — four versions of Compare
problems from Table 10.1 - are shown in Table 10.2. The keywords in the
example in Table 10.2 are taller and shorter. Taller is associated with more
and hence with addition; shorter is associated with less and hence with
subtraction. Lewis used 16 problems of this kind with different keyword
pairs, such as fast-slow or hot-cold. The force of these keywords was
either consistent with the nature of the problem, or inconsistent. Lin¬
guistically, the two comparative terms are not equivalent. Taller is the
default term, which can be used without particular connotations, as in
How tall is Jeff ? Smaller is the marked member of the pair. Its use is not

Table 10.2. Problem types used by Lems

Consistent Inconsistent

Add Tom is 175 cm tall. Tom is 175 cm tall.

Jeff is 12 cm taller than Tom. He is 12 cm shorter than Jeff.
How tall is Jeff? How tall is Jeff?
Subtract Tom is 175 cm tall. Tom is 175 cm tall.
Jeff is 12 cm shorter than Tom. He is 12 cm taller than Jeff.
How tall is Jeff? How tall is Jeff?
 Word problems 349

neutral. How small is Jeff? implies that Jeff is small. One would therefore
expect that unmarked comparative terms would influence the course of
hypothesis formation little or not at all; marked terms, however, should
have a stronger effect, because normally speakers use such terms only if
they have something special in mind.
Lewis’s data for 144 college students are shown in Figures 10.4 and 10.5
for addition and subtraction problems, respectively. These figures show
the percentage with which subjects chose the correct hypothesis about the
appropriate mathematical operation at three query points located at the
end of each of the three sentences of the word problems. Logically, sub¬
jects should guess randomly after the first sentence and should respond
correctly at the end of the second sentence. The final question is logically
redundant. The keyword should be disregarded, and hence the consistent
and inconsistent problem version should be equivalent. This is not what
happened. Performance was imperfect after the second sentence and gen¬
erally improved after the third sentence. When the keyword was consis¬
tent with the problem structure, the correct hypothesis was chosen earlier
and with a higher final accuracy. Inconsistent keywords depressed both
the rate and asymptote of performance. There was a significant interac¬
tion between consistency and markedness. Linguistically marked com¬

Consistent data

•O- Inconsistent data

—•O'— Consistent predictions

—a— Inconsistent predictions

ist Sentence 2nd Sentence 3rd Sentence

Query Points

Figure 10.4 Percentage correct hypothesis choice for addition problems at

three query points.
350 Models of comprehension

Consistent data

O Inconsistent data

—o— Consistent predictions

—a— Inconsistent predictions

1st Sentence 2nd Sentence 3rd Sentence

Query Points

Figure 10.5 Percentage correct hypothesis choice for subtraction problems

at three query points.

parison terms dramatically interfered with performance on the addition

problems, whereas the unmarked term had a lesser, though still significant
effect on the subtraction problems. Furthermore, guessing after the ini¬
tial sentence showed a distinct bias in favor of addition, presumably
because of the prevalence of addition problems in U.S. schools.
To simulate these results, we assume that there are two procedures,
ADD and SUBTRACT, that will be chosen with probabilities according
to the ratio rule,

s(ADD)
Pr (ADD) =
s(ADD) + s(SUBTRACT)

where s(ADD) and s(SUBTRACT) are the strength values of these pro¬
cedures determined by the spreading activation mechanism of the Cl
model. The strengths of these procedures depend on how much support
they receive from their associated schemas, ADD-SCHEMA and SUB¬
SCHEMA, as well as from the more- or less- keywords. The schemas in
turn are supported by the text. Text propositions favoring either the
ADD-SCHEMA or the SUB-SCHEMA are linked to the respective
schema nodes. Text propositions, such as those derived from the first sen-
 Word problems 351

tence of these problems, which do not indicate the problem structure, are
linked equally to both hypotheses, although more weakly (Figures 10.6a,
10.6b, and 10.6c).
Figures 10.6a, 10.6b, and 10.6c provide a processing trace of the model
for the inconsistent addition problem. The parameter values used were
estimated by informal exploration to maximize the goodness of fit of the
model to the data in Figures 10.4 and 10.5. Specifically, links among text
propositions and and links among schemas and procedures were assigned
a value of 1. Links between text propositions and the schema they sup¬
ported were given a value of .8. Neutral propositions were linked to both
schemas w ith a value of .4. Marked key words were linked directly to the
procedure with a value of .4. All other links were zero. Each node was
linked to itself w ith a value of 1, except for the ADD procedure, which
was assigned a value of 1.2, which is the model’s way of accounting for the
observed response bias in favor of addition. Intuitively, these parameter
estimates seem reasonable in this problem context.
The first sentence, shown in Figure 10.6a, expresses the proposition
TALL[TOM] with the modifier 175CM. These nodes are linked to both
schemas, which in turn are linked to their corresponding procedures. The
network settles with the activation values shown. The probability of
choosing the correct hypothesis (ADD) at this query point is .58/(.58 +
.47) = .55; that is, the model starts out with a modest addition bias.
The second sentence, in Figure 10.6b, adds another proposition with
a modifier,

SHORTER[TOM,JEFF]
I_ 12CM

which unequivocally specifies the problem as an addition problem; these

propositions are therefore linked to the ADD-SCHEMA only. However,
the key word shorter suggests subtraction and is therefore linked to the
SUBTRACT procedure. The activation values that result from the set¬
tling of the network are also shown. SHORTER[TOMJEFF] becomes
the strongest proposition, which seems right intuitively. The probability
of choosing the correct procedure becomes ,42/(.42 + .31) = .58, thus
barely increasing over the prior value. This is comparable to the results
observed for inconsistent addition problems.
Figure 10.6c shows the final proposition - [HOW[TALL[JEFF]]] -
352 Models of comprehension

T ALL [TOM] 1.0000

1 75CM 1.0000
addition schema 0.8397
subtraction schema 0.7748
add 0.5814
subtract 0.4725

"Tom is 175 cm tall."

(a)

TALL [TOM] 0.3493

175CM 0.5603
SHORTER [TOM, JEFF] 1 .0000
12CM 0.7134
addition schema 0.9603
subtraction schema 0.3562
add 0.4244
subtract 0.3087

subtraction schema

"He is 12 cm taller than Jeff."

(b)

Figure 10.6 A processing trace for an inconsistent addition problem at three

query points. Final activation values of the nodes are shown next to each
graph.

which further strengthens the add procedure (it supports the ADD-
SCHEMA in the same way as the previous sentence) but succeeds in lift¬
ing the probability of choosing the correct procedure to only .65. This far
from perfect performance level attests to the strong interference pro¬
duced by the marked keyword and corresponds to the actual performance
of Lewis’s subjects.
 Word problems 353

TALL [TOM] 0.5975

175CM 0.3608
SHORTER [TOM, JEFF] 0.9432
12CM 0.5837
TALL [JEFF] 0.7579
HOWlTALL] 0.5212
addition schema 1.0000
subtraction schema 0.1937
add 0.3587
subtract 0.1920

"How tall is Jeff?"

(c)

Figure 10.6 (cont.).

Figures 10.4 and 10.5 show the predictions for all four problem types
(Add, Subtract, Consistent, & Inconsistent) used in the Lewis experi¬
ment. It is obvious that the predictions agree rather well with the data.
Indeed, a chi square goodness of fit test yields a nonsignificant value, x2
(9) = 18.75, which is quite impressive considering (1) that each of the 12
data points was based on 576 observations so that the goodness of fit test
is very sensitive, and (2) that we have estimated only three parameters.
The success of the Cl model in accounting for these data does not
mean that alternative models are wrong. However, it is worth noting that
time course predictions for the formation of hypotheses could be
obtained with the Cl model without any additional assumptions, just by
running the model sentence by sentence. Also, key word effects and bias
parameters could be incorporated in a natural way. That is not to say that
alternative models could not be similarly elaborated, although for single¬
schema, top-down control models some rather arbitrary assumptions
would be needed to account for the time course data as well as the key¬
word effects.
What intellectual gain has been achieved by modeling these data? If
one wants to know about the time course of hypothesis formation, key¬
word effects, or response biases, one merely has to look at the data to find
354 Models of comprehension

out. The simulation does not tell us anything about these matters that the
data did not already reveal. The primary value of the simulation is that it
allows us to understand the particular results of this study within a broad,
overarching framework that ties together many different observations
about language understanding and problem solving. It lets us see these
results as pieces of a much larger puzzle, as systematic observations
related to a host of others.
There may also be some more immediate gain. As Kintsch and Lewis
(1993) have suggested^if we want to find out about inconsistency effects
in Change problems, we would have to run another expensive experi¬
ment. Or we can boldly calculate. According to our simulation, the fol¬
lowing problem should have a solution probability of .88.

(9) Tom has 12 marbles.

Jeff gives Tom 3 more.
How many . . .?

If the problem is reformulated with an inconsistent marked term: Tom has

12 marbles. He takes away 3 marbles from Jeff the probability of a correct
solution ought to drop to about .63. Thus, we don’t have to do a new
experiment for every new problem - we can use the model to simulate
outcomes. How successful such an engineering approach would be
remains to be seen, but it certainly should be tried.

10.1.5 The emergence of schemas

The Cl model for word arithmetic problem solving as formulated in

Kintsch (1988) and Kintsch and Lewis (1993) differs from the earlier
generation of models primarily in that the Cl model considers alternative
schemas in parallel as candidates for forming a problem model. In con¬
trast, in Kintsch and Greeno (1985) and Cummins et al. (1988) smart
rules were employed to select the right schema in the first place. The Cl
model undoubtedly represents progress over the earlier formulation, but
it is not the end of the story. Arithmetic problem schemas must be
learned, whereas Kintsch (1988) assumes that the children already have
full-fledged schemas for arithmetic that need only to be applied correctly.
Stern (1993b) has shown that this is by no means the case and that acquir¬
ing the right kind of arithmetic problem schemas is a major facet of learn¬
ing to solve word problems. I sketch here Stern’s findings and outline
 Word problems 355

what the next generation model for word problem solving should look
like. It incorporates the work discussed so far but turns further away from
the abstract schema concept and toward the notion of situated cognition.
The Kintsch (1988) formulation corresponds to the performance of
children who have had a great deal of experience with these particular
word problems. However, a schema-based problem model is the end
product of learning to solve word problems in school; it does not charac¬
terize the way children learn to solve these problems. Stern (1993b)
showed convincingly that children in the early elementary school years do
not have an abstract schema that they use to form problem models in the
way we have surmised. Instead, their knowledge is more accurately char¬
acterized as a set of concrete, situation-bound principles about arithmetic
that allow them to solve many problems and out of which abstract general
schemas will eventually be constructed.
Stern’s data come from an eight-year longitudinal study that assessed
the cognitive and social development of a large group of school children
in Munich, Germany - the LOGIK project of Weinert and Schneider (in
press). Two observations from this study form the core of Stern’s argu¬
ment. First, in a large number of cases, students can provide the correct
answer to a word problem but are unable to explain how they arrived at
their answer. Specifically, they are unable to write a correct equation for
their solution. This happens to all students, both high and low achievers,
and for all kinds of problems, as long as the problem is challenging for a
particular student. The student is able to figure out the problem by what
he or she knows about numbers and arithmetic operations, taking into
account the situational constraints imposed by the problem, but without
being able to form a metalinguistically explicit problem model. Figure
10.7 shows some representative results. Percentage of errors, correct
solutions with equation, and correct solutions without equation are
shown for a three-year period for the following problem:

(10) Three children are celebrating a birthday.

Mother bought 10 apple tarts.
Each child eats 2 apple tarts.
How many tarts will be left over?

This problem can be understood in terms of concrete operations and can

be solved correctly without forming an abstract problem model. Across
three years of schooling, the percentage of errors is approximately halved,
356 Models of comprehension

Errors

Correct&Equation

Correct&No-Equation

Figure 10.7 The percentage error, percentage correct solutions with correct
equation, and percentage of correct solutions without correct equation for the
“apple tart” problem in grades 2, 3, and 4. After Stern (1993b).

but the proportion of correctly solved problems for which no equation or

an incorrect equation was given remained around 35%.
The second observation that, according to Stern, indicates that chil¬
dren do not have fully developed abstract arithmetic schemas is based on
an analysis of children’s attempted solutions of various types of problems.
If solving a certain problem is evidence that the student has formed the
correct problem model for this type of problem, then the student should
also be able to solve other problems requiring the same problem model.
This is often not the case. As an example, consider a student who suc¬
cessfully solves

(11) Fred had 4 balls.

He has 3 more balls than Elly.
How many balls does Elly have?

and

(12) Tom has 7 balls.

Jane has 11 balls.
How many balls have they altogether?
 Word problems 357

According to Stern’s argument, the first problem provides evidence that

the student has a Compare schema for unknown referents; the second
provides evidence that the student has a Combine schema. The student
therefore should be able to solve

(13) Jan has 7 rabbits.

He has 4 more rabbits than Mike.
How many rabbits do Jan and Mike have together?

In this problem the Compare schema must be used first to determine how
many rabbits Mike has, but that is not specifically asked for. About one
third of the children in the two grades where all these problems were
tested solved the first two problems correctly but could not solve the
third problem. (The reverse, that the third problem was solved correctly
but errors were made on either of the first two, was very rarely observed,
indicating that we are not dealing here with random variability but with
a true asymmetry.)
Problem models, therefore, are not derived from metalinguistically
available schemas, except at a very late stage of schooling. Stern (1993b)
discusses some evidence that suggests such schemas are in fact the end
result of learning how to solve word problems.
Stern also points out that the assumption that children use arithmetic
schemas to solve word problems is somewhat inconsistent with the
essence of the Cl models. In other applications of the model, schemas are
not treated as full-fledged structures, ready to be retrieved and used, but
as emergent structures that are put together in the context of a particular
task from smaller knowledge units. In the case of word arithmetic prob¬
lem solving, the knowledge units used to construct problem models are
the arithmetic principles as they are learned in concrete task contexts by
the child. Only later will schemas be abstracted from these principles.
One principle will come to be associated with another one, thus eventu¬
ally becoming a unit and a schema. But while the children are learning to
solve word problems, their knowledge is less orderly, less abstract, and
more situated. For instance, commutativity may be contextually
restricted to numbers below 20; the possibility of reformulating more
statements by less statements may be realized in some contexts but not
others; the term more than may be understood in the context of an action
but not in the context of a static comparison. Instead of a few powerful
358 Models of comprehension

schemas, children appear to have a lot of disjointed, situationally

restricted knowledge. So far, no one has systematically analyzed what
these bits and pieces are and how they are interrelated. Once this is done,
a simulation within the framework of the Cl model could be constructed
to account more accurately for the learning that children engage in when
they solve word problems. The Kintsch (1988) formulation might be the
final, most sophisticated version of a sequence of models characterized by
increasingly general and more powerful arithmetic schemas, but starting
with quite restricted and context-bound knowledge structures. This
would be by no means an idle exercise but would have considerable the¬
oretical and practical interest. It would be a good environment in which
to study how schemas emerge, and it could be used to explore various
instructional scenarios. Instructors and tutors can no longer assume that
their students merely need to apply their arithmetic schemas in solving
word problems; they need to be concerned how these schemas emerge in
the first place.

10.2 Word algebra problems

There are some good reasons why one should study word algebra prob¬
lems. The domain of algebra is so much richer and more complex than
that of arithmetic, and hence the texts of word problems are longer and
more interesting, placing greater demands on a model of discourse com¬
prehension. Furthermore, the central issues in algebra word problem
solving appear to be the same as in the arithmetic domain. In particular,
the importance of situation models in algebra word problem solving is at
least as high as in arithmetic word problem solving. In a pioneering study,
using extensive protocol analyses, Hall, Kibler, Wenger, and Truxaw
(1989) have shown that competent college students reason within the sit¬
uational context of a story problem to identify the quantitative con¬
straints required for a solution. They use the text to build, elaborate, and
verify a situation model from which they derive their solution. A variety
of reasoning strategies are used by students to develop these situation
models, which are by no means restricted to the algebraically relevant
aspects of a problem. Integrating the dual representations of a problem at
the situational and mathematical level appears to be the central aspect of
competence according to Hall and his co-workers, as indeed it was for
children working on arithmetic word problems.
 Word problems 359

10.2.1 Algebraic schemas

The basic components of a model of word algebra problem solving might

therefore be quite similar to those of our model for arithmetic word prob¬
lem solving. However, the very richness and complexity of the algebra
domain that attracted us in the first place might seem to make this
approach impossible. We had to deal with only a few arithmetic schemas -
those used by first-graders discussed earlier, the Part-Whole schema,
which subsumes these at a more advanced stage of development, plus a few
more primitive versions characteristic of earlier levels of development, as
described by Riley et al. (1983). But how many algebra schemas are there?
Mayer (1981) compiled a list of 1,097 problem types from current algebra
texts. How can we deal with such complexity?
It turns out that a complete simulation of algebra word problem solv¬
ing is indeed impossible. It is not the complexity of algebra that is the
decisive obstacle, but linguistic considerations. The algebra is really quite
straightforward. No more than a few schemas are needed to solve the vast
majority of problems, plus some knowledge of other domains such as
geometry, physics, statistics, and so on. Thus, a relatively small number of
algebraic schemas can do all the work that is needed for the construction
of the necessary formal problem models. The difficulty is in the con¬
struction of the situation model, without which the student cannot get to
the right problem model. The demands on the construction of situation
models in college and high-school algebra problems are open-ended,
employing all kinds of linguistic and general world knowledge that might
be needed for that purpose. For the arithmetic simulation of Cummins et
al. (1988), the general knowledge that had to be given to the simulation
was quite limited. If a new concept was encountered in a problem that
was not already known to the simulation, it asked the experimenter, and
the experimenter could give simple answers: steal is like take away, deer
and pencil are countable objects; cars and trains belong to the superset
vehicle; cars and dolls to the superset toys; yellow is a modifier. The sim¬
ulation needs to know no more. This strategy does not work for algebra
word problems. One really must know what words mean, what their
implications are, how various objects move in the world, what their
shapes are, and so on, endlessly. As we have regretted in other places in
this book, there is no knowledge base that contains all that information in
a form that could be used by a simulation program.
360 Models of comprehension

The only way one could build a simulation of word algebra problem
solving would be to restrict the domain radically, say to a blocks world, so
that the simulation could know everything there is to know in that
domain. Further restrictions would have to be placed on the form of the
linguistic input. This did not seem an attractive alternative to us. We
decided, therefore, in favor of a hand simulation in which the problem of
general world knowledge is avoided by using our own intuitions as
required. There is a great deal this approach cannot deliver in compari¬
son with a real simulation, but some useful implications can still be gath¬
ered from it.

10.2.2 A model for word algebra problems

The model we have constructed is described in Nathan, Kintsch, and

Young (1992). It is a direct development from the models for arithmetic,
in particular Reusser (1989), because the interplay between the situation
model and the problem model is central to it. A major innovation is the
algebraic schemas on which the more than 1,000 problem types of alge¬
bra are based. Four Rate schemas are needed to deal with the largest class
of algebra problems, rate problems, which I describe presently. Another
class of problems involves physics, geometry, and schemas from other
domains. This is essentially an open class. Examples are Newton’s second
law (the sum of all forces for a system in equilibrium must be zero);
Ohm’s law; the Pythagorean theorem; and the formulas for the area and
circumference of a circle and other geometric shapes. Finally, there is a
third class of word algebra problems — number problems - which are not
schema based but must be constructed from the text directly without the
use of a schema. We focus here on rate problems.
The four Rate schemas are all of the same form.

Unit! = rate-of-Unit[-per-Unit2 x Unit,

The units may be amount per time (a boat travels 4 km per hour), cost per
unit (a pound of almonds costs $5.99), portion to total cost (7% of the cost
of a car), and amount to amount (5% acid in 3 gallons of solution). These
Rate schemas are the building blocks of algebraic problem models. The
story text usually specifies one or two of the members of a schema (Unitj,
Unit2, rate). If only one member is missing, it can be computed from the
 Word problems 361

formula. Algebra problems usually require the instantiation of more than

one schema, together with the relation between them. Thus, we might
have two kinds of nuts with different costs, and mix the two according to
certain specifications, requiring three cost-per-unit schemas. To construct
a problem model, the problem solver must (1) pick the right schema,
which can be done on the basis of textual cues, just as in the case of the
arithmetic schemas; (2) specify as many elements for each schema as the
text allows; and (3) find or infer from the text the relationships among
the schemas used. Once a problem model has been constructed, an equa¬
tion for the problem can be found by constraint propagation within the
problem model.
The working of the model is best illustrated by an example. Figure
10.8 shows a common distance-rate-time problem. Its textbase contains
three propositions. The time slot of the first proposition is unspecified,
but the reader must assume some initial point in time to build a situation
model. Both planes leave from the same place but at different times and
at different speeds. At some point in time, they will have covered the same
distance and the second plane will overtake the first. It is easiest to think
of the model that students construct for this situation as an animation:
Two planes start from the same place on the some route and continue
until they meet. The rather clumsy two-dimensional graph in Figure
10.8 is merely a substitute for this scene. The actual situation model rep¬
resents directly what happens in the real world.
Nathan et al. (1992) used some simple graphical conventions to con¬
struct a problem model. Every Rate schema is represented by a vertical
arrangement of three circles, the two units, and the rate that relates them.
Quantitative values from the problem text that specify either a unit or a
rate can be written inside the appropriate circles. Thus, the schema from
Plane 1 consists of a time slot and a distance slot, both unspecified because
nothing was said about either in the problem, and a rate, which the prob¬
lem identified as 200 miles per hour. This particular schema was selected
from all algebraic schemas the reader knows on the basis of such keywords
as distance, rate, time, and moving objects. A second schema is similarly
constructed for Plane 2.
To complete the problem model, the student must find out what infor¬
mation the text provides that specifies the relation between the two Rate
schemas that have been constructed. First, there is a relation between the
flying times of the two airplanes: one leaves 3 hours after the other, lo
362 Models of comprehension

Text:
A plane leaves Denver and travels east at 200 miles per hour. Three
hours later, a second plane leaves Denver on a parallel course and
travels east at 350 miles per hour. How long will the second plane
take to overtake the first plane?
Textbase:
LEAVE[PLANE1, DENVER] LEAVE[PLANE2,DENVER] OVERTAKE[P1,P2]
rate: 200MPH rate: 350MPH
direction: EAST direction: EAST
time:3 H LATER time: ?
loc: PARALLEL COURSE

Equation:
200(t+3) = 350t

Figure 10.8 A word algebra problem: the text, textbase, situation model,
problem model, and equation.
 Word problems 363

enter this information correctly into the problem model, the student
must realize that this means that the flying time of Plane 1 equals the fly¬
ing time of Plane 2 plus the 3 hours. The other piece of information
needed is not explicit and must be inferred on the basis of the student’s
world knowledge: If the planes start from the same place and fly along the
same route, they will have covered the same distance at the point at which
the faster one overtakes the slower plane. Students have a great deal of
experience with races between various kinds of vehicles and are quite
familiar with this situation. Therefore, this knowledge is available, but it
must be activated at this point and used to complete the problem model.
An equation is easily derived from the problem model. The unknown
in the problem is the time t that the second plane needs to overtake the
first. This can be entered in the problem schema as shown. Next, t + 3 can
be computed as the travel time for Plane 1. Finally, the distances traveled
by each plane can be expressed as the product of time multiplied by
speed, and set equal, resulting in the equation shown in Figure 10.8. The
equation can then be solved as in any numerical problem.
This is indeed a complex model for solving a simple problem! Similar
problem models can be constructed for all rate problems and, with some
obvious extensions, also physics and geometry problems. But is this com¬
plexity necessary for these rather simple problems? First, algebra word
problems are not simple for most people and most students. Second,
there is some evidence that the model presented here is by no means
unnecessarily complex. What we lack and cannot have for the reasons we
have discussed is a full simulation of the model and direct empirical tests
of it, as in the case of word arithmetic problems. (The subtlety of infer¬
ring equal distance from overtake is a good example of the kind of general
world knowledge that is necessary for these problems that we do not know
how to represent in a knowledge base.) We can argue for the present
model on three grounds.
First, the components of the model are directly motivated by our
research on word arithmetic problems. We had to introduce a situation
model and problem model representation between the textbase and equa¬
tion to understand how people solve word arithmetic problems. It would
be surprising if we could do with less in the case of woid algebra. Second,
in their study of word algebra problem solving, Hall et al. (1989) obtained
strong empirical evidence that good students engage in extensive leason-
ing at the level of the situation model and that the major difficulty they
364 Models of comprehension

have is in coordinating their formal model with their situational under¬

standing. Finally, Nathan et al. (1992) and Nathan (1991) have collected
some experimental evidence that supports the psychological reality of the
structures and processes assumed here. For this research they adopted
the strategy of developing an algebra word problem tutor from the prin¬
ciples about word problem solving that have been discussed here.

10.2.3 The ANIMATE tutor for solving word algebra problems

Intelligent tutoring is based on the premise that the system solves a prob¬
lem in a way that simulates what human experts do. The system also has
a model of the student and is able to interpret the student’s behavior in
terms of that model and of its own understanding of the problem. Flence,
it can diagnose errors and take remedial steps. Because we cannot simu¬
late human word algebra problem solving, this approach is not open to us.
Flowever, suppose the model described here is essentially correct. If
that were the case, what instructional implications would it have? What
are the difficult steps in solving word algebra problems where students
might need help? Because reading is not problematic for college students,
they are perfectly capable of forming a textbase. Also, because our initial
goal is to bring their performance on word problems up to their level of
performance on numerical problems, we will not bother helping them
with the equations. Furthermore, we know that college students have no
trouble understanding the situations described in algebra word problems.
There are some exceptions, especially in the areas of statistics and prob¬
ability, and we have also encountered some rather fuzzy notions about the
concept of “interest.” However, these exceptions are neglected here. In
general, students do not need help constructing situation models from
the simple contexts depicted in word problems. They do need help with
the formation of the problem model, because they often skip this stage
and jump to equations without a clear understanding of the problem
structure. Thus, we must force them to build explicit problem models,
and we must provide them with the means to do so. Some students and
experts successfully use notes and graphs for this purpose, but in general
students do not have available a systematic, easy system for designing
problem models. Nathan et al. (1992) decided to teach students the
graphic system we used to represent problem models in our theory as a
systematic way to construct such models while attempting to solve them.
 Word problems 365

This has proven to be quite feasible. With minimal instruction, students

learn to use this way of representing a problem and most find it useful.
Good students sometimes prefer not to be bothered because they have
their own ways of doing these things, but average or poor students are
often able to gain insights by using this simple representational scheme
that they have been unable to obtain in years of traditional instruction.
However, students are even more in need of help with another diffi¬
culty in word algebra problem solving. It is not enough to give students a
means to build formal problem models. They must be able to tell whether
they have built the right one, the one that corresponds to their situation
model.
Problem models are a necessary step in problem solving because we are
unable to compute the solution of a problem directly in our situation
model. This is not the case for all problems. Children need not compute
symbolically a problem like 9 = 7 + 2, for they have the answer available
directly. We need not formalize the problem of how to make coffee with
an unfamiliar coffee machine, for we can solve the puzzle by responding
directly to the pot’s affordances (Larkin, 1989). Many problems, however,
cannot be solved without recourse to some formal computational system.
Building a highway bridge and solving differential equations are some
examples. Algebra word problems generally are in that class, too. We have
a perfectly good situation model for the problem in Figure 10.7, but it will
not help us to find a solution. For that, we need to reformulate the prob¬
lem mathematically. In using mathematics, however, there is the danger
of losing the correspondence between the mathematical formulation of
the problem and its real-world structure (Greeno, 1989; Kintsch, 1991).
Situation model and problem model must match. At all levels of educa¬
tion in science and engineering this match is often difficult to ensure and
is frequently not achieved. The difficulty is not the formal system per se,
nor is it situational understanding per se; it lies in the correspondence
between the formalization and the situation.
This is a notoriously difficult issue in engineering or software design.
But eventually the real world gives the formalist some feedback as to the
adequacy of the formalization: Bridges last for centuries, or they collapse;
the software is perfect but does not do what the customer imagined it
would. Experienced algebra word problem solvers provide their own feed¬
back by checking their answers for reasonableness against their situation
model. That is exactly where students need help; otherwise they come up
366 Models of comprehension

with negative values for the length of a board and similar absurdities (Paige
& Simon, 1966). Because algebra students cannot operate in the real world
and receive feedback from it, the next best thing is to supply them with a
substitute world in the form of an animation that acts out what is implied
by the problem model they have constructed. Thus, in the case of the
problem in Figure 10.8, the animation would show Plane 1 leaving; after a
delay, Plane 2 leaves at a greater speed. After some time, t it overtakes the
first plane and the animation stops. On the other hand, if the student
makes an error and constructs a problem model in which the slow plane
leaves after the fast plane, this is exactly what happens in the simulation,
and the student invariably realizes the error. Similarly, if the student for¬
gets to specify the distances the planes travel, they will simply go on fly¬
ing, even after one overtakes the other. The student notices that something
is wrong, and now has a chance to figure out what.
Nathan et al. (1992) have built such a tutor, called ANIMATE. It is not
an intelligent tutor, because the system does not understand the problem
at all. It merely executes an animation as instructed by the student’s prob¬
lem model (it paints fences, mixes solutions, stacks up piles of money, or
lets objects move). If the right thing happens, the student realizes that the
solution is correct. If the wrong thing happens, the student must figure
out how to correct it. If nothing happens because the simulation was not
given enough information to execute anything, the student knows that
one or more pieces are missing from his formal model and can try to com¬
plete it.
Figure 10.9 provides a graphical description of word problem solving.
Both problem-relevant information and situation-relevant information
are extracted from the problem text, forming potentially isolated problem
representations that must be coordinated. This link is what the ANI¬
MATE tutor focuses on. ANIMATE has two components, Network,
which is a system that allows problem models to be constructed accord¬
ing to the graphical conventions we have sketched, and the Animation,
which runs whatever Network tells it to. The purpose of ANIMATE is to
enable the student with the two links in the figure labeled Is it right? and
What went wrong? by making explicit the correspondence between the
problem model (in Network) and the situation model (the Animation).
ANIMAl E works quite well. Although it has never received a field test
in a classroom study, it has passed a number of laboratory tests with good
grades (Nathan et al., 1992). Nathan showed that with only minimal
 Word problems 367

Situation Conceptual
Model Problem Model

WORD PROBLEM
Figure 10.9 The role of the ANIMATE tutor in helping students to solve
world algebra problems. From Nathan et al. (1992).
368 Models of comprehension

training on ANIMATE, students outperformed suitable control groups

who worked on new word problems without the system. The various
components of ANIMATE were tested separately in this study, which
clearly established the significant role of the animation. For instance, in a
pre- and posttraining comparison, students who were trained with Net¬
work only, without the animation, showed an improvement of 30%,
whereas students who were trained with the whole system, including the
animation, improved by 57.5%. For comparison, a conventionally
instructed control group improved by 9%. Such results are impressive if
one remembers that these students had many years of conventional train¬
ing and were exposed to the ANIMATE environment for only a single
brief session.
Nathan (1991) performed an extensive protocol study of a small group
of students using ANIMATE with Work, Mixture, and Investment prob¬
lems as well as Motion problems. He gave his subjects a pre- and posttest
of comparable difficulty, so that their improvement as a function of train¬
ing with the ANIMATE system could be accurately determined. ANI¬
MATE was by no means equally effective for everyone. The two subjects
who scored highest on the pretest performed very well on the posttest. A
second group of three subjects performed almost equally well on the
posttest, in spite of error scores in excess of 50% on the pretest. The two
remaining subjects scored poorly on the posttest, although for one of
these subjects that was a great improvement over the pretest, where he
got almost nothing right. Nathan also gave the subjects in this study an
impossible problem (a problem that had no solution). Only the three mid¬
dle subjects who showed the greatest improvement from pre- to posttest
recognized that this was an impossible problem. It is not surprising that
the two poor subjects failed to do so, but the two subjects who were very
good from the beginning also did not recognize that the problem was not
well specified. Even though they knew how to solve these problems, their
procedures were mechanical rather than based on understanding.
The protocols Nathan collected suggest that ANIMATE did not pro¬
vide enough help for the two poorest subjects, who were lost and reduced
to making random adjustments to the animation. These learners needed
more guidance than ANIMATE could give them. If they had been given
more elementary problems to solve, they could perhaps have made better
use of ANIMATE and learned successfully. The two subjects who already
knew how to solve these problems also did not learn much. They relied on
 Word problems 369

what they knew and failed to use ANIMATE to deepen their understand¬
ing. If they had been given more challenging problems, they might have
found more use for ANIMATE and learned to engage in situational rea¬
soning in the process. For the other three subjects, ANIMATE seemed
just right. The problems were difficult for them but not too difficult, and
the kind of support ANIMATE provided enabled them to solve these
problems and learn a great deal in the process. My hunch is not that this
style of tutoring does not work with poor students or with very good stu¬
dents. Rather, in order to learn students must receive problems that are
adjusted to their level of skill: problems they cannot quite solve on their
own but which they can just manage with the scaffolding provided by the
tutor.
It is unfortunate that ANIMATE was never used in a real educational
environment. As it is, it is impossible to judge its value as a classroom
tool. But from a theoretical perspective, the kind of work that has been
done with ANIMATE has proven to be quite interesting. Discourse
comprehension theory can be used for the analysis of word problem solv¬
ing, and such an analysis allows us to significantly expand and elaborate
the theory.

How people understand and solve word problems in algebra or

arithmetic is of particular interest for the theory of text compre¬
hension because for these kinds of tasks comprehension does not
merely result in the formation of a mental representation that is
to be tested or evaluated later by some indirect measure; instead,
comprehension directly results in an action. 1 hus, the empincal
data constrain the simulations of mental processing more tightly.
In addition to the usual predictions about comprehensibility and
memory, the model can also predict whether a solution will be
correct or what sort of error might occur. The simulations of how
children solve word arithmetic problems first focused on the
arithmetic, then on the language, and then more and more on the
child’s understanding of the situation and the mathematization
of that understanding. Much the same thing can be said about
college students solving algebra problems. 1 he crucial step is
alwavs the formalization of the student’s situational understand¬
ing in terms of a mathematical problem model. What is achieved
by the simulations that were developed and tested here is to spell
Models of comprehension

out the components involved in this formalization: the mathe¬

matical concepts and operations themselves, the informal situa¬
tional understanding, and the knowledge of the linguistic con¬
ventions used in word problems. All three can be sources of
errors and difficulties for a student that can be studied and
understood by tracing their sources in a simulation. We cannot
observe directly the students’ thinking, but we can simulate it,
manipulate conditions in the simulation, and try to reproduce
behavior that we see in our students. Thus, we can understand
what students -are faced with and help them more effectively
when needed.
 11

Beyond text

The model for word problem solving that was developed in the previous
chapter can provide a framework for a broad range of related compre¬
hension tasks whenever a verbal instruction — a text or an oral communi¬
cation - must be understood in the context of a particular task and situa¬
tion to guide the actions of the comprehender. In the previous chapter the
action required was to calculate a numerical answer to a mathematical
word problem; the text was used to construct a situation model and a cor¬
responding problem model. The formal problem model provided the
mathematization of the problem, which made it possible to employ the
tools of arithmetic or algebra for the solution of the problem. An analo¬
gous situation is encountered in many other tasks, in particular tasks that
require the use of a computer for the solution of a problem. Instead of a
mathematical problem model, we are dealing in this case with a formal
system model. To use the computer, the situational understanding of a
task must be reformulated into what we have called a system model
(Fischer, Henninger, & Redmiles, 1991).
The formal specification of a computing task in terms of a system
model is analogous to the formulation of a word problem in terms of a
mathematical problem model. It requires a situational understanding of
the task at hand, as well as a knowledge of the formal constraints of the
computer system that must be respected in generating the system model.
The required situational understanding can be more complex than in the
case of word problems. In a word problem the question (usually) specifies
precisely the goal of the computation. When people perform tasks on a
computer system, an explicit specification of the goal state is not neces¬
sarily available. The goal of the action is frequently specified only in
general, situational terms and must first be translated into system terms.
372 Models of comprehension

For instance, suppose my goal is to send a certain manuscript to a col¬

league. In system terms, this might mean that I want a file containing this
manuscript to be available to my colleague at her computer. Alternatively,
in a different system, it might mean that I print out the manuscript and
mail it to her. Thus, the formulation of a sufficiently precise goal schema
in system terms may in itself be nontrivial - for instance, if a user has an
inadequate understanding of the system characteristics.
A goal schema, formulated within a particular system model, guides
action through the mechanism of pattern completion. The term pattern
completion or redintegration - Komplexergdnzung — was introduced by
Selz (1922; see also the discussion in Kintsch, 1974). The basic notion is
that, given a portion of some pattern, the whole pattern completes itself.
Thus, in action planning an action goal anticipates the outcome of the
action and thereby guides the action toward that outcome. In the follow¬
ing sections, I show how this kind of pattern completion process func¬
tions to allow experienced users to perform routine computing tasks.

11.1 Action planning

Action planning on the basis of verbal instructions can be understood

within the framework of the construction-integration model. The kinds
of actions considered here are routine tasks in familiar situations - in par¬
ticular, computing tasks. The tasks concerned are not so simple that they
are entirely scripted, that is, require merely a fixed action sequence. But
neither are they so complex that they require intentional, active search
and problem solving. They are the kind of task that experts spend much
of their time with. They are performed semiautomatically, with little con¬
scious effort, on a regular basis, but always in somewhat different ways,
because the situational context and the task requirements are complex
and conflicting and never quite the same. Specifically, in Mannes and
Kintsch (1991) we studied routine computing tasks involving file main¬
tenance and editing as well as using the mail system.

11.1.1 Routine actions

We started by observing how experienced computer users perform such

tasks while providing concurrent think-aloud protocols. Our subjects
knew immediately what to do in response to the instructions we gave
 Beyond text 373

them and provided rich verbal protocols. These protocols were proposi-
tionalized and formed the core of the knowledge for our computer simu¬
lation. In addition, this knowledge base was augmented in three ways:

1. Six other experienced computer users were shown each of the proposi¬
tions generated in the concurrent verbal protocols and asked to provide
a free association - the first thing that came to mind. These subjects
were not told anything about our purposes, so that the associations they
produced (such as as post office and stamp in response to mail) were fre¬
quently irrelevant to the computer context.
2. For all propositions that stated a request (e.g., to enter mail, edit a file),
a second proposition was added that stated what the outcome of the
action would be if the request were acted on. This information was
sometimes generated spontaneously in the protocols or the free associ¬
ation task, but because it was crucial for our purposes we had to make
sure that it was available for every request.
3. We added a few action plans that were needed to perform the tasks we
were concerned with that were not generated spontaneously by our
subjects.

Three types of nodes in the long-term memory network were thus

generated: general knowledge about computers (e.g., to use the mail, you
must be at the mail-level, not the system-level), general knowledge about
the tasks to be done (e.g., to write a letter to someone, you need an
address), and specific action plans called plan elements, which are the
commands needed to execute the task.1 Plan elements are formally
propositions, like all other elements in the network. They take thiee argu¬
ments: a plan name (e.g., REPLY to a message), a set of conditions that
must exist in the world for this plan element to be executable (e.g., a mes¬
sage must exist in order to reply to it), and a set of outcomes of the execu¬
tion of the plan elements (e.g., the sender of the message receives the
reply). Figure 11.1 shows an example of a plan element, and the follow¬
ing are some examples of plan-element names (out of a total of 26 used in
the simulation). A question mark indicates an unbound variable.

(EDIT FILE?)

1 Plan elements are analogous to productions in AC. 1 * (Anderson, 1983).

 374 Models of comprehension

Plan-element Name » PRINT FILE?

EXIST FILE?
Plan-element Precondition
KNOW FILE? LOCATION

Plan-element Outcome * EXIST FILE? HARDCOPY

F igure 11.1 A sample plan element for printing a file.

(PASTE TEXT? FILE?)

(FIND FILE?)
(PRINT FILE?)
(SEND FILE? MAIL)
(SEND FILE? SYSTEM)
(COPY MESSAGE)
(RENAME FILE? FILEANEWNAME)

The nodes thus generated were connected via argument overlap, form¬
ing an 81 x 81 long-term memory matrix that was fixed and did not
change during the course of a simulation. This long-term memory net¬
work was then used in generating dynamically a task network in working
memory when the system was instructed to perform some particular task.
Generating the task network involved the following sequence of steps:

1. T he instruction was propositionalized, yielding two types of propo¬

sitions: REQUESTS to do something (e.g., “print the file called ‘letter’”
(REQUEST ((PRINT FILEALETTER))), and INTHEWORLD
propositions describing the task environment (e.g., “there is a file called
‘Letter’ in the world” -»• (EXIST(FILEALETTER INTHEWORLD))).
2. For every REQUEST proposition, the anticipated OUTCOME of
that request was retrieved from long-term memory and added to the task
net. Thus, for the request to print the file “letter” the outcome (EXIST
(FILEALET TER HARDCOPY)) was included.
3. Every REQUEST, INTHEWORLD, and OUTCOME proposition
in the task network served as a retrieval cue to sample n propositions from
the long-term memory network. Thus, the task network was enriched by
related information from long-term memory. For large values of n, the
 Beyond text 375

system used practically everything it knew that was related to the task at
hand; for small values of n, on the other hand, the system might some¬
times fail to use critical information it knew about but did not sample on
a particular occasion, leading it to make an error (see section 11.3 for fur¬
ther discussion of this aspect of the model).
4. All plan elements in long-term memory were bound to existing
objects in the world and added to the proposition list. Thus, if two
files are mentioned in the instructions, two copies of all plan elements
with the variable FILE?, one bound to each INTHEWORLD file, were
generated.

5. The next step was to link the nodes that had been generated in
working memory into a network:

• Positive links were established on the basis of argument overlap.

This assures, for instance, that a (REQUEST SEND X) is linked to
all plan elements with the name SEND, or that (OUTCOME Y) is
linked to all elements that have the outcome Y.
• Inhibitory links were established (1) between a requested outcome Z
and a plan element that had outcome —Z; (2) between a plan ele¬
ment that had precondition X and a plan element that had the out¬
come ~X; and (3) between an INTHEWORLD proposition and a
plan element that has the same INTHEWORLD proposition as its
outcome; thus, (KNOW (FILEALETTER LOCATION INTHE¬
WORLD)) inhibits the plan element (FIND (FILEALETTER)).

Figure 11.2 illustrates how inhibitory links function in this system.

They allow a requested outcome to inhibit incompatible plans; they allow
each plan element to protect its preconditions; and they deactivate plan
elements that are no longer necessary because their outcomes already
exist INTHEWORLD.
Thus, a task network was constructed from two sources: (1) the
instructions given to the system, which included information about the
state of the world, and (2) what the system knew about the task and itself.
The task network thus constructed was then integrated by spreading
activation around until the activation pattern stabilized. This pattern
indicates what the system wants to do: which plan elements are strong!)
activated, which are less activated, which are altogether deactivated.
 376 Models of comprehension

(KNOW FILEALETTER LOCATION INTHEWORLD)

/
/

Figure 11.2 Excitatory (solid lines) and inhibitory (broken lines) relations
among plan elements and INTHEWORLD propositions.

However, it is usually not possible to carry out the action one wants to do
most, because the preconditions for that action might not be met by the
existing INTHEWORLD environment. Thus, we cannot send a letter
without first writing it, entering the mail system, and so on. We therefore
introduced a distinction between can-do and want-to nodes.
Accordingly, all computations were done in the network of want-to
nodes, but the can-do nodes determined what action was taken. Each
want-to node activated a corresponding can-do node in proportion to its
own strength. However, can-do nodes were inhibited if their precondi¬
tions did not exist INTHEWORLD. Thus, although the system wants
most strongly to send a letter, it cannot do so because there is no letter. It
therefore executes the most highly activated can-do plan element. As we
shall see, this tends to be an action that produces the missing precondi¬
tions and starts a chain of intermediary actions that change the state of
the world step by step and eventually result in the execution of the
desired final action.
Mannes and Kintsch (1991) give a formal and detailed description of
how task networks are constructed and integrated and describe a com-
putei program called NETWORK that performs these operations. I
must refer the reader to this original paper for the details of how this sys¬
tem works and focus here instead on the major conceptual issues
involved. The details of such a program are, of course, all important, but
they are available elsewhere, which allows me to discuss the theoretically
significant points at a more general level.
 Beyond text 377

First, we work through a very simple task: “Include an address you

know in a letter that is in a file.” This text, translated into propositional
form, becomes the core of the task net: There exists a text called “letter”
in a file called “letter,” as well as a known text called “address,” and a
request to include the address in the “letter” file. Next, the system
searches its long-term memory for the outcome of the request to include
a text in a file. It retrieves the outcome proposition (IN TEXT A
ADDRESS FILEALETTER) and adds it to the task net. Then, each of
the propositions already generated is used to retrieve further information
from long-term memory, resulting in the inclusion of another set of
propositions in the task network. In this particular case, however, these
enrichment propositions play no role whatever, so we can neglect them.
Finally, the variables in the 26 plan elements in long-term memory are
bound to the objects existing INTHEWORLD, yielding 33 bound plan
elements for the task net. Note that all plan elements containing the vari¬
able TEXT? will be duplicated in this process, the variable being bound
once to TEXT A LETTER and once to TEXT A ADDDRESS.
The task network thus constructed was a 78 x 78 matrix. 1 he following
link strengths were used: links based on argument overlap were assigned
a value of .4, whereas positive links based on causal relations (as in Figure
11.2), as well as request and outcome links, were emphasized by assigning
them a higher value (.7). Inhibitory links were set at —10 to assure theii
effectiveness in a network consisting mainly of positive links. These link
strengths were obtained on the basis of informal trial-and-error explo¬
rations for a workable parameter set (which remained unchanged for all
simulations that were performed).
The task network was now ready for integrating. The activation values
for the four most highly activated plan elements after the first integration
cycle are illustrated in Figure 11.3 (complete results are in the original
paper). The system wants most strongly to paste the address into the let¬
ter or to type it, but because the preconditions for these actions aie not
met INTHEWORLD, only the want-to but not the can-do nodes are
activated. The most strongly activated can-do node is FINDALETTER.
Hence, this action is performed, and the outcome of this action - the
location of the “letter” file is now known - is added to the list of INTHE¬
WORLD propositions.
Thus, a new, changed task net is obtained, which is now integrated in
turn. As shown in Figure 11.3, after the second integration cycle, the sys-
 378 Models of comprehension

Cycle 1 Cycle 2 Cycle 3

Figure 11.3 Relative activation values for selected plan elements on the first
three cycles of the “Include” task. AD abbreviates ADDRESS and LET
abbreviates LETTER. Want-to nodes are shown as open circles, can-do nodes
are shown as filled circles. A star indicates the action taken on each cycle.

tem still wants most strongly to paste in the address or type it, but cannot
do so. The strongest can-do node now is EDIT FILEALETTER. This
action is executed, it once again changes the state of the world - the file
lettei is now in a buffer, ready for typing — and another processing cycle
begins. This time TAPE TEXTAADDRESS can actually be performed;
it generates the outcome (IN TEXTAADDRESS FILEALETTER),
which completes the pattern established by the original request. As a final
step (not shown in Figure 11.3), the system exits the newly created
FILEALETTER/ADDRESS.
We have here an example of situated cognition. The system does not
plan ahead by setting up a search space and finding a goal path by means-
end analysis or some other heuristic. Instead, it looks at the situation,
comprehends it in terms of its goals, figures out what it wants to do, and
does what it can do, thereby creating a new situation. It then compre¬
hends the changed situation and acts again, producing another change in
its world. In the example just described, each of these^changes was a step
in the right direction, eventually allowing the system to reach its goal.
There is no guarantee that this would have been the case, but in an
 Beyond text 379

orderly world about which the system has sufficient knowledge, that is
exactly what should happen.
This action without planning, but responding in a goal-directed man¬
ner to an ever changing situation, has considerable intuitive appeal. Con¬
sider what happens when I decide to leave the building: I do not plan to
open the door, go out in the hallway, turn right, descend the stairs, and so
on. Instead, I simply respond to the open door by walking through it,
respond to being in the hallway by walking toward the stairs, and so on.
Such routine tasks are different from real problem solving. For example,
when I design an experiment, I do try to think of all the possibilities in
advance and plan very far ahead. Not all thinking consists in understand¬
ing situations in terms of particular goals. More complex, intentional
problem-solving processes certainly are required in many cases, but not
all planning is problem solving, and much that we have treated in cogni¬
tive science as problem solving might be better understood within a com¬
prehension paradigm.
Mannes and Kintsch (1991) work through several more illustrative
tasks in their paper. One is a version of the notorious conflicting subgoal
task: “Print and delete the file.” To handle the sequencing in this task is
not trivial for most AI programs and may require special mechanisms.
NETWORK solves the task without problems. The spreading activation
process favors PRINT over DELETE because of the causal chaining built
into the plan elements (Figure 11.2). PRINT inhibits DELETE because
it would destroy a precondition necessary for printing, but not vice versa.
NETWORK also handles more complex, two-step tasks, such as
“Revise a manuscript you are working on with a colleague by removing a
paragraph, and then send the revised manuscript to that person.” Quite
a few steps are involved in executing these instructions, all in the propel
sequence. There is much that could go wrong, therefore, but NET¬
WORK manages not to lose its way.
I have already commented on one characteristic of NETWORK,
namely, that it does not plan ahead, but responds to each situation as well
as it can in a goal-directed way. Another aspect of NETWORK worth
noting is that it is nondeterministic. In generating a task net, it does not
include all information it has available in long-term memory but only a
random sample - whatever the current propositions in working memory
are able to retrieve. This means that it has different information available
 380 Models of comprehension

on different trials with the same task, and hence it may take a different
course of action. If by accident a crucial piece of information is missing
(has not been sampled from long-term memory), it may fail at a task that
it could do otherwise (see section 11.1.3 for further discussion) or per¬
form it in a different way. That, of course, happens to people, too. Mannes
and Kintsch (1991) asked one of their subjects, an experienced computer
user who had just performed one of the tasks we had given him in a rather
clumsy, inelegant way, why he had not chosen a more optimal solution.
He immediately understood the optimal solution but commented that he
just had not thought of it. Because of its random sampling mechanism,
NETWORK also might “not think” of a possible solution path.
We, the theorists, have built into NETWORK all the knowledge it has.
Thus, NETWORK does not address the question of how knowledge is
acquired; yet once constructed, NETWORK can learn from experience.
Mannes and Kintsch (1991) reported some initial observations of how
such a system could learn from its own experience by remembering par¬
ticular cases it had worked on. Case-based reasoning is a large and impor¬
tant topic in cognitive science, and our explorations with NETWORK
are certainly not the last word on this topic, but they are worth noting
here because of the extreme simplicity of our approach. We simply add a
node that encodes a few essential features of a case NETWORK had
already solved to its long-term memory network. If retrieved when
NETWORK works on a new case (e.g., because there is some overlap
between the current task and the prior task), this case node participates
in the integration process as part of the task network, often playing a
major role in redirecting the activation flow.
Io illustrate how NETWORK reasons with cases, Mannes and
Kintsch descnbe how NE TWORK solved the “Revise manuscript” task
aftei having previously worked on three related tasks. For each prior task,
we created a memory node consisting of the most strongly activated argu¬
ments of the nonplan-element propositions in the network, plus the
names of the plan elements involved in performing the task. Thus, for the
“Include” task shown in Figure 11.3, the arguments FILEALETTER,
TEXT ADDRESS, and so on would be included, together with the plan
elements that were actually executed (FIND, EDIT, TYPE, EXIT).
Given a new problem involving a FILEALETTER, this case memory
might be retrieved from long-term memory and strengthen the plan ele-
 Beyond text 381

ments associated with it. If the two tasks are sufficiently similar, this
might be a good thing: Relevant plan elements would be strengthened,
and irrelevant, potentially interfering elements would be weakened, lead¬
ing to a quicker solution of the problem. On the other hand, case mem¬
ory can equally well interfere with the solution of new problems - for
instance, if the new task requires some entirely different action involving
FILEALETTER. Both positive and negative effects of case memory are
well demonstrated, and it remains to be seen whether NETWORK can
simulate these results in detail.
Considerable work needs to be done on case-based reasoning in NET¬
WORK. For instance, just what information, and how much, does a case
memory contain? Do we need a special forgetting mechanism (a more
recently performed task probably has a stronger influence than one done
a long time ago)? Thus, there are many open questions at this point, but
they seem worth studying, for NETWORK has some potential advan¬
tages over other case-based reasoning systems. It requires neither a spe¬
cial selection mechanism nor special evaluation procedures nor analogi¬
cal reasoning. All we need is to add properly defined memory nodes to our
network - the retrieval and integration mechanisms are already there. It
would be interesting to see how far such a simple-minded model can be
pushed.

11.1.2 Novices and experts

Expert-novice differences have received much attention in the literature

on skill acquisition (Chi, Feltovich, & Glaser, 1981; Larkin, 1983). Expei ts
not only know more than novices; they perform tasks in a qualitatively dif¬
ferent way, seemingly unaffected by the resource and working memory
demands that are such a burden for novices and intermediates (Ericsson
& Charness, 1994). Doane and her colleagues (Doane, Kintsch, & Poison,
1989; Doane et al., 1992) have investigated this phenomenon within
the context of the construction-integration model. The skill that they
examined is particularly interesting and challenging: UNIX command
production.
It is widely known that UNIX is an extremely powerful and useful pro¬
gramming language but very difficult to master. Authors such as Norman
(1981) have pointed to the arcane command names employed in UNIX as
382 Models of comprehension

one source of this difficulty, but as we show later, this is no more than a
minor problem among many more serious ones that UNIX poses for its
users.
Doane, Pellegrino, and Klatzky (1990) observed UNIX users of vari¬
ous levels of experience, some of them over extended periods: novices
(average experience with UNIX 8 months), intermediates (average expe¬
rience 2 years) and experts (5 years’ experience). All subjects had taken
classes in the advanced features of UNIX and were active users of UNIX.
Thus, even the novices in this study were by no means inexperienced.
Subjects were given brief verbal instructions to produce legal UNIX
commands, using as few keystrokes as possible. There were three types of
tasks: single commands-(e.g., “display a file”), multiple commands (e.g.,
“arrange the contents of a file alphabetically,” “display the file names of
the current directory”), and composite commands (e.g., “sort the con¬
tents of a file alphabetically and print the first ten elements”). Composite
commands are thus little custom-made programs to perform specific
tasks and made up of elementary command sequences using special
UNIX facilities such as pipes or redirection symbols.
Doane et al. (1990) made a number of interesting observations. Both
novice and intermediate users knew the elements of UNIX; that is, they
could successfully execute the single and multiple commands, but they
could not put these elements together to form composite commands. The
novices failed entirely at this task, and the intermediates often found
nonoptimal solutions by using unnecessarily complex redirection meth¬
ods rather than directly piping the output of one command into the next
one. Only the experts could effectively use the input-output (I/O) redi¬
rection facilities of UNIX — one of its central and most powerful design
features!
Doane et al. (1989) simulated these results in a system called UNI¬
COM, which was modeled after the NETWORK system described in the
previous section. A knowledge base for UNICOM was constructed on
the basis of the empirical results of Doane et al. (1990). It consisted of
general, declarative knowledge and plan elements (procedural knowledge
or productions). General knowledge contained both specific information
about the syntax of UNIX commands and the I/O redirection syntax, as
well as more abstract conceptual knowledge about commands and redi¬
rection in UNIX. The plan elements were formally equivalent to those in
NETWORK (Figure 11.1): a name, preconditions, and outcomes. As in
 Beyond text 383

Figure 11.2, each plan element activated those plan elements that gener¬
ate its preconditions and inhibited those elements that interfere with its
preconditions. Thus, the model is designed so that the interconnections
among plan elements represent the user’s knowledge about causal rela¬
tionships within the UNIX operating system.
The verbal instructions were again propositionalized into REQUEST,
OUTCOME, and INTHEWORLD propositions, as in the previous sec¬
tion. UNICOM integrated these instructions together with its knowl¬
edge and responded by performing the most activated among its plan ele¬
ments whose preconditions were met, as was described in the previous
section for NETWORK. In this way, the system performs correctly all
the tasks used in the Doane et al. (1990) experiment. Thus, it simulates
the behavior of the expert subjects in that study, roughly speaking, for
even experts sometimes make errors.
To simulate the behavior of the intermediate and novice subjects, we
(Doane et al., 1989) lesioned the knowledge base of UNICOM. We took
out all the knowledge, general and specific, about redirecting input and
output of commands to simulate the novice subjects. To simulate the
intermediate subjects, we gave the system knowledge about rediiection in
general and information about the use of redirection symbols, but we
deleted more advanced features, such as knowledge about the use of
pipes. With these lesioned knowledge bases, UNICOM reproduced the
major features of the Doane, Pellegrino, Klatzky data. The system cor¬
rectly performed single and multiple commands, unless know ledge about
a specific command was deleted, but it was unable to perform composite
tasks. It failed completely in the novice version and succeeded only occa¬
sionally, and without finding an optimal solution, in performing a com¬
posite command in the intermediate version.
We also were able to simulate the performance of particular users. Our
strategy was to give UNICOM all the knowledge a specific user had
employed in solving the task. This knowledge was determined from the
explicit evidence the user had provided in a verbal report. We then ran
UNICOM with this knowledge base on the problems that we had previ¬
ously given to the subject and noted whether or not UNICOM s behav¬
ior matched that of the subject. If we had everything right - the theory as
well as the knowledge - the behavior of the subject and our simulation
should match closely.
For instance, one of our tasks was “Display the first 10 elements ot the
 384 Models of comprehension

file ‘eggplant,’ sorted alphabetically.” The correct command is ls/sort/

head (where / is a pipe, Is means list, sort means alphabetize and head
limits the action to the top 10 elements). A novice subject we studied had
shown evidence that he knew the command Is, but not sort and head.
With this kind of knowledge, both the novice and the system failed at the
task, both entering “unknown” as their response. Two years later, the sub¬
ject, now an intermediate, entered Is as his only response. As we assessed
his knowledge base, we found that he now also knew the command sort.
Given this addition to its knowledge base, UNICOM matched this
behavior. Although the Is command was available on the first test, it was
not differently activated. On the second test, however, Is became more
strongly activated. Because the outcome of Is is an unmet precondition of
sort, additional activation was generated by the sort command. For both
the user and UNICOM, then, knowing only Is resulted in no response,
whereas knowing both Is and sort resulted in the response Is. Thus, if
UNICOM is given the same knowledge that individual subjects appear to
have at different levels of skill, UNICOM will perform the same way the
subjects did.
We further investigated expert-novice differences in a study of prompt
comprehension (Doane et al., 1992). In that study we gave subjects diffi¬
cult UNIX tasks to perform, while providing them with various types of
prompts as needed. The fact that quite different prompts were helpful to
novices and experts reveals qualitative differences that exist in the pro¬
gramming skills of novices and experts.
In Figure 11.4 a sample task is shown together with a series of help
prompts that were shown to the subjects whenever they made an error.
The prompts were shown one at a time in a fixed order, regardless of the
nature of the error. There were four types of prompts. Some prompts,
such as P2, provide specific command syntax knowledge. On the other
hand, P3 merely provides general conceptual information rather than a
specific command syntax. Other prompts are merely reminders, remind¬
ing the user of information already made available in earlier prompts (P5,
for example). Finally, P6 provides order information. If the user is unable
to solve the problem with all these prompts, the correct solution is
revealed in the last prompt.
Natuially, experts outperformed novices and intermediates, but there
was a clear difference in the kind of prompts experts and novices found
most useful, as shown in Figure 11.5. Novices needed a lot of specific
 Beyond text 385

Task Description

Format the text in ATT2 using the -ms macro package and
store the formatted version in ATT 1

Eiomols

Prompt 5 You will need to use the

Prompt 1 You will need to use the
arrow symbol" if and the command
following command
nroff -ms
One that will format the contents
of a file using the -ms macro package Prompt 6 You'll need to use an
nroff -ms on ATT2 (which will output

Prompt 2 You will need to use this the formatted contents of ATT2). and

command youll need to redirect this output as

input to ATT 1
nroff -ms will format the contents
Prompt 7 You will need to use
of a file using the jns macro package
exactly the following command
elements (though not necessarily
in this order):
nroff -ms
Prompt 3 You will need to use a
special symbol that redirects command Prompt 8 You'll need to use the
output to a file oommand nroff -ms followed by the
arrow symbol' j'

Prompt 9 The correct production is

Prompt 4 You will need to use the
arrow symbol' / that redirects output nroff-ms ATT2>ATT1
from a command to a file Please enter this production now

Figure 11.4 Example of task description and prompts for the problem
“nroff-ms ATT2>ATT1.” From Doane et al. (1992).
386 Models of comprehension

[x] Novice

□ Intermediate

[~] Expert

Figure 11.5 Changes in percentage correct after four types of prompts for
novice, intermediate, and expert UNIX users. After Doane et al. (1992).

information about command syntax and ordering. Experts profited lit¬

tle from this type of information, presumably because they knew it
already; however, more general prompts - conceptual information and
reminders - were most helpful in their case. This finding suggests that the
errors made by experts have a different source than novice errors. Even
though experts have the requisite knowledge, they may forget something,
or it simply does not occur to them. Hence, a general reminder will nudge
them back onto the right path. Does this mean that experts are simply
being careless when they make an error? We investigate this interesting
question further in the following section.
So far, the only difference between experts and novices that we have
examined has to do with the amount of knowledge. However, real users
have other problems, too; they may have misconceptions, and they may
make errors because their working memory becomes overloaded. Doane
et al. (1992) have not studied the misconceptions UNIX users have, nor
have we simulated them as was done, for instance, in the Buggy system of
Biown and vanLehn (1980). We did, however, obtain evidence for mem¬
ory load problems in our study of prompt comprehension study. Mem-
 Beyond text 387

ory losses may involve both item and order information. In our study,
novices and intermediates left out five or six times as many items initially
as experts (very few deletion errors were made by anyone after the third
prompt). Loss of order information as indexed by substitution errors
showed a similar pattern. Such errors were two to three times more fre¬
quent among novices and intermediates than among experts, and they
continued to be a serious problem up to the sixth prompt, which first pro¬
vided the order information the novices and intermediates were lacking.
Thus, memory loss appeared to be a serious problem for the novices and
intermediates, whereas the experts managed working memory much
more successfully.
Novices and intermediates need help retrieving the elements that go
into making a composite command, and they need help with ordering
these commands. The task of command construction severely taxes their
working memory, so that memory errors become a serious problem. This
is just what one would expect from the UNICOM simulations reported
earlier. The number of elements that go into the construction of a com¬
posite command, together with their ordering, imposes a significant bur¬
den on working memory. Furthermore, the user must also represent and
maintain in working memory the invisible intermediate results of the
process. Nowhere does UNIX tell the user what the outcome of the first
computational step is or the input to the next step. The user must antici¬
pate and remember such information without help from the system. Only
experts with well-established retrieval structures are up to this task.
Novices and intermediates not only lack knowledge; they can become
confused even when they have it.

11.1.3 Simulation of errors

One puzzling fact about skilled performance is that even experts, who
certainly know better, make errors. Indeed, error rates for experts are
often surprisingly high. For instance, in the prompting study desciibed
in the previous section, experts made many fewer errors than intermedi¬
ates and novices in producing composite UNIX commands, but their ini¬
tial solution attempts were by no means always successful, with error
rates varying for different problems between about 13% and 65%. Card,
Moran, and Newell (1983) have studied a skilled user performing two
tasks, manuscript editing and electronic circuit design editing. Eriors
 388 Models of comprehension

were made on 37% of the editing operations. Most of these were detected
during the generation of the command sequence, but 21% of the com¬
mands issued by this highly skilled user contained an error that had to be
corrected later. Similarly, Hanson, Kraut, and Farber (1984) collected a
large database from 16 experienced UNIX users performing document
preparation and e-mail tasks. They observed an overall error rate of 10%,
with a range for different commands from 3% to 50%. Why do experts
make errors, and can simulations with the construction-integration
model account for this phenomenon?
Kitajima and Poison (1995) have developed a simulation of a human-
computer interaction task that suggests an answer to this puzzle. The
sample task that they analyze involves the use of the popular graphics
application, Cricket Graph. A user is given a set of data and instructed to
produce a graph with certain characteristics. This is a much more com¬
plex problem than the routine tasks studied by Mannes and Kintsch
(1991) or even the production of composite UNIX commands simulated
by Doane et al. (1992). Nevertheless, Kitajima and Poison successfully
simulated performance on this task with a model designed in much the
same way as these simpler models. That is, they used a propositional rep¬
resentation for the various knowledge sources involved: the display the
user is confronted with, the user’s goals, the user’s knowledge about
graphs and computer systems, and finally the objects and actions involved
(what we called the plan elements in Figure 11.1). A task network was con¬
structed fiom these elements in much the same way as in Mannes and
Kintsch (1991), and an integration process then selected the strongest
action candidate for execution. This action changed the display in one way
or another, which led to the selection of a new action, until the desired goal
state was reached.
However, to achieve their goal, Kitajima and Poison had to elaborate
this basic model in several ways. First, the practice of binding all objects
in the environment to all possible actions had to be abandoned — there were
simply too many possibilities, leading to a combinatorial explosion. Inter¬
views with human users suggested that these users simply were not pay¬
ing attention to most of these possibilities. Selective attention is of course
something familiar to psychologists, and Kitajima and Poison found a
ready way to incorporate such a mechanism into the Cl model. First, the
model does a construction-integration cycle with just the objects (e.g.,
menu items, whatever else there is on the screen) in view and the task goals
 Beyond text 389

to decide which objects to pay attention to. Then, in a second construc¬

tion-integration cycle, it selects an appropriate action, restricting its
attention to the objects selected in the first cycle. On the basis of the ini¬
tial integration process, the model focuses specifically on three candidate
objects, constructs action-object pairs for these objects, and then goes
through another integration cycle to select the strongest action-object
pair from this set. Thus, the model did not have to consider all possible
actions at each step, only those involving the objects it was currently pay¬
ing attention to.2
Kitajima and Poison’s first and by no means trivial result was that
this system actually could draw the graph it was supposed to. This is a
complex task, involving several subtasks and subgoals, and it is of some
interest to note that a comprehension-based planner is still able to find its
way through such a complex task. It was not at all clear that problem¬
solving tasks of this complexity could be simulated within the Cl frame¬
work. Kitajima and Poison’s work thus further obscures the border
between comprehension processes and problem solving and suggests that
the domain of the comprehension paradigm may be wider than anyone
suspected.
Even more important than the mere fact that it is possible to simulate
such complex problem-solving behavior with a comprehension model are
some of the implications of this simulation. Each action cycle in the Kita¬
jima and Poison (1995) model has two parts: an attention cycle and an
action selection cycle. The attention cycle must focus on the light objects
for a correct action to occur, but what basis do we have for making this
selection? Activation spreads from the task goal and anticipated outcome
to the objects on the screen in proportion to their relatedness to the task
goal and outcome. Sometimes this relatedness is established because rele¬
vant knowledge of the user is sampled in the elaboration phase, thus pro¬
viding a link between a screen object and the task goal even if no obvious
and external relation exists. But often there is an obvious relationship
between goals and screen objects, because screen objects tend to be labeled
in ways that indicate their goal relevance. Thus, if a user wants to delete an
object, there might be a screen object labeled “delete or erase , if some¬
thing needs to be copied, the user will look for “copy or duplicate, 01

some similar term. Labels often, but not always, indicate the action to

2 The details of this simulation are described in Kitajima and Poison (1995).
390 Models of comprehension

which an object is relevant. The Kitajima and Poison model strongly relies
on such labels. If the labels do not indicate which objects are relevant, the
model might still make a correct decision if it has a great deal of knowl¬
edge about the task at hand and if it uses that knowledge, but for the most
part it will be lost. So are human users. Novices rely heavily on a label¬
following strategy, failing when that strategy fails. Experts are more flexi¬
ble and often can figure things out on the basis of their knowledge and
experience, but even experts do better when they can follow the labels.
Figure 11.6 shows data on Cricket Graph users from a dissertation by
Franzke (1994). On the first trial, novice users of Cricket Graph are able
to act quickly if the action is labeled in the same way as in the instructions
or by means of a synonym. The action time increases sharply when an
inference is required, that is, when there is no menu item that is linked to
the instruction either directly or by a synonym. Action times are even
longer when no verbal link exists at all (e.g., to change an axis in a graph
you must double click on it in Cricket Graph — even longtime Macintosh
users have a hard time figuring this out!). On a second trial, the time dif-

Figure 11.6 Time required for using the Cricket Graph software for actions
that are labeled in the same way as the instruction, actions that are labeled
with a synonym, actions that are labeled in such a way that the user must
make an inference, and actions that are not labeled at all. The data are for the
first and second trial of novice users. After Franzke (1994).
 Beyond text 391

ferences are greatly reduced, but they are still apparent. Indeed, some
recent unpublished data of Poison show that even experts are faster when
they can rely on the label-following strategy than when they cannot.
Thus, the label-following strategy appears to be the choice of novices in
a wide variety of situations and remains important even when a user has
become familiar with an application. The Kitajima and Poison (1995)
model correctly implies the importance of this label-following strategy in
human—computer interaction because labels play an often crucial role in
directing attention to the task-relevant objects on the screen in the inte¬
gration cycle that precedes the actual action selection.
The other aspect of Kitajima and Poison’s (1995) model that is of par¬
ticular interest here is their error simulation. A simulation that can solve
all the problems is not a good simulation of human behavior because
humans, even experts, make errors. Kitajima and Poison found a natural
way to account for these errors. One of the parameters of the model is the
rate of sampling knowledge that is related to the current display, as in
Mannes and Kintsch (1991). Nsample is the number of times each element
in the representation that has been constructed is used to retrieve a
related knowledge element, where the probability of sampling a particu¬
lar knowledge element depends on the strength of its relationship to the
cue, relative to all other elements. Kitajima and Poison ran 50 simulations
each for values of Nsample between 4 and 20. Some of their results are
shown in Figure 11.7, where the error rate for various component actions
is shown as a function of how much long-term memory is retrieved.
Although some actions are much more knowledge-dependent than oth¬
ers, it is clear that the sampling rate is a major determinant of success. II
an item is allowed to retrieve no more than four associated knowledge ele¬
ments from long-term memory, the error rates vary between 8% and
75%. When up to 20 elements can be sampled, the error rates drop to
between 0%> an d 15*%. Thus, the model has a natuial explanation for
such errors. If users try hard to retrieve whatever knowledge they have,
error rates will be low. But if users are sloppy and do not use the knowl¬
edge they have, large error rates will result.
The three simulations discussed here - routine actions by Mannes and
Kintsch (1991), UNIX commands by Doane et al. (1992), and Cricket
Graph by Kitajima and Poison (1995) — demonstrate that the Cl model of
comprehension can be extended beyond text comprehension. The same
392 Models of comprehension

—□— Task 1

.O. Task 2

—0— Task 3

Figure 11.7 Predicted error rates for a difficult (Task 1), an intermediate
(Task 2), and an easy task (Task 3) as a function of the amount of elaboration,
the Nsampie parameter in the model. After Kitajima and Poison (1995).

basic processes of bottom-up, relatively context-free construction fol¬

lowed by contextual integration can account for text comprehension as
well as for at least some ways of comprehending a situation and taking
action on that basis. Just where the limits of the Cl model are in this
respect is still an open question. It seems clear that there are limits and
that more complex forms of planning and problem solving require the
construction of problem spaces and goal hierarchies in the manner envis¬
aged originally by Newell and Simon (1972). But at this point it certainly
appears worthwhile to investigate other possible fields of inquiry that tra¬
ditionally have not been considered in terms of comprehension, to try to
delineate further the scope of the comprehension paradigm.

Where is the boundary between comprehension and problem

solving? Action planning and decision making, at least in some
cases, can be considered within the framework of comprehension
theory. Just how successful the comprehension account of these
phenomena will be in the long run remains to be seen, but we
certainly have here a reversal from the days when text researchers
Beyond text 393

were enjoined to look at text understanding as a process of prob¬

lem solving. Instead, we can now look at planning and decision
making as a comprehension process!
What are the essential, distinctive psychological assumptions
about action planning that characterize the models presented in
this chapter and differentiate them from alternative accounts?
There is first the representation of knowledge as a network of
propositions. Perhaps alternative forms of knowledge represen¬
tation could also be made to work, but the work discussed in this
chapter illustrates how smoothly the knowledge net representa¬
tion chosen here meshes with the mechanisms of the construc¬
tion-integration model. We also should note, however, that the
knowledge net is no longer strictly an associative net, for the pat¬
tern of interconnections among plan elements reflects causal
knowledge: plans that interfere with each other inhibit each
other, plans that require each other activate each other.
The assumptions involved in the construction of the task net
and its integration are in part dictated by external constraints
that are not specific to the present model. Any model must some¬
how represent the instructions and the state of the world. I he
psychologically important assumptions concern the process of
knowledge elaboration and variable binding. They are deter¬
mined by the philosophy of dumb construction rules that is basic
to the Cl model: Knowledge elaboration is associative and con¬
text independent, and its depth is variable; variable binding, too,
is promiscuous - every plan is bound to every object in sight.
Thus, the construction processes in action planning are simple
and context free, just as in the case of text comprehension. The
system obtains its power via the integration process in both cases.
A few more noteworthy assumptions in this model of action
planning need to be mentioned here. This is a planning model
without planning in the strict sense. The model just compre¬
hends a situation, takes an action that changes the situation, and
responds to the situation again until a solution or a dead end is
reached. There is no setting up of subgoals, no planning ahead;
no memory for earlier actions is necessary, except for the extei-
nal memory provided by the changed situation itself.
The distinction between want-to and can-do nodes is crucial.
394 Models of comprehension

The constraint satisfaction process must consider all potential

actions, not just the ones currently possible. Reasoning must be
general, but at the same time action is limited by the possible.
We also note that the extreme simplicity of the model, which
was sufficient to model routine computing actions and UNIX
command production, had to be abandoned in the more complex
Cricket Graph domain. It was no longer feasible to bind all vari¬
ables in the plan elements to all objects present; instead, a selec¬
tion had to be made of those objects that were to be attended to
in the light of the task at hand and the current display on the
computer screen. This is clearly a gain in psychological accuracy,
for selective attention is a ubiquitous phenomenon, but it comes
at the cost of substantially complicating the model mechanisms.

11.2 Problem solving, decision, and judgment

In this and the following sections I discuss some possible extensions of

the Cl model to new areas. I have not really worked in these areas, and
other than suggestions on how the Cl theory might be applied there I
have nothing to offer. The only solid research results I discuss have been
obtained independently elsewhere (the work of Kunda & Thagard (1996)
on impression formation), and I am merely pointing out the relationship
of their work to the Cl model. So why discuss these topics at all? Would
not silence be more appropriate if one does not have anything definitive
to offer? I think not. In the coming years, research on the topics treated
here so superficially may yield the richest rewards and may perhaps trans¬
form cognitive science into a broader and richer discipline than it is now.

11.2.1 The role of comprehension in problem solving

and decision making

I hroughout these chapters I have insisted on a distinction between prob¬

lem solving and comprehension. Comprehension is automatic, bottom-
up, described by the mechanisms of the Cl theory. Problem solving is a
controlled, resource-demanding process involving the construction of
problem spaces and specialized search strategies, as originally envisaged
by Newell and Simon (1972).
Not everyone will want to follow me here. So let me try to explain and
 Beyond text 395

justify this distinction in a little more detail. Both comprehension and

problem solving involve a construction phase and a solution phase. The
main difference between them lies in the nature of the constructions
involved in generating a mental representation. For comprehension, the
Cl theory claims, the construction phase is essentially guided by the tex¬
tual (or other perceptual) input. Propositions are constructed more or
less closely, mirroring the input sentences of a text sentence by sentence.
This is a highly constrained process, at least for the ideal reader, because
actual readers sometimes deviate in unpredictable ways from the text (not
forming certain propositions invited by the text, or constructing propo¬
sitions that do not accurately reflect the text, usually simplifying it). The
next step in the construction process - knowledge elaboration — is less
predictable, because memory retrieval is probabilistic. However, because
we all live in the same world and therefore have similar knowledge struc¬
tures and memories, even this phase is not totally unpredictable, at least
for an ensemble of comprehenders. The construction operations them¬
selves are typically highly practiced and demand few mental resources.
This is the realm of long-term working memory, for mental representa¬
tions that later will also serve as retrieval structures are being generated
automatically and reliably.
A very different situation obtains for true problem-solving tasks. Typ¬
ically, the input vastly underconstrains the construction of a problem
representation. Instead of a text that almost dictates the kind of mental
representation that will be constructed, the problem statement itself
gives few hints as to what the problem space looks like. Conscious, com¬
plex, resource-demanding mental operations are required to set up a
problem space. Because this can be done in many different ways, it is dif¬
ficult or impossible to predict how a given individual will approach a
problem. Often, the operators that must be used for the construction of
the problem space are unfamiliar. T he problem solver in this situation
cannot rely on a few well-practiced, highly overlearned, automatic oper¬
ations, and hence not on available retrieval structures in long-term work¬
ing memory. Thus, the resource demands of problem solving are much
greater than in comprehending a reasonably familiar text.
However, the boundaries between comprehension and problem solv¬
ing are obscured when we are dealing with expert problem solvers, whose
behavior is in some ways more like comprehension. Experts in some
problem-solving domains rely on their long-term working memory in
396 Models of comprehension

much the same way as text comprehenders do (Ericsson & Kintsch,

1995). The expert chess player, for instance, calculates moves almost
as automatically as a text comprehender understands a sentence, generat¬
ing a retrieval structure in the process that will support later memory
retrieval.
The solution phase in problem solving and the integration phase in
text comprehension are also different in nature. The Cl theory claims
that a spreading activation process is all that is needed for integrating the
kind of networks constructed in text comprehension. Spreading activa¬
tion almost certainly also plays a role in problem solving, but in addition
a variety of more directed processes also play a role. In many cases a prob¬
lem space is so huge that spreading activation processes become ineffec¬
tive, and problem solvers must employ other solution strategies (such as
the means-end analyses of Newell & Simon, 1972). As a consequence, a
relatively simple theory (like the Cl theory) may suffice to account for
orderly, predictable comprehension, whereas the theory of problem solv¬
ing (like SOAR, Newell, 1990) might need to be considerably more com¬
plex.
To indicate how problem solving might be analyzed within the Cl the¬
ory, I briefly discuss two classic examples: the Tower of Hanoi as the pro¬
totype of a whole class of laboratory problem-solving tasks, and a decision
task that illustrates the representativeness bias in decision making.
One of the most analyzed problem-solving tasks is the Tower of Hanoi:
a board with three sticks and three (or more) disks, graduated in size with
a hole in the center so that they can be placed on the sticks. Initially all
three disks are on the first stick, as shown in Figure 11.8. The task is to
move the three disks to the third stick 3, with two constraints: only the
top disk of a pile of sticks can be moved, and a larger disk can never be

3-2-1 0 0

Figure 11.8 Initial state of the Tower of Hanoi problem and formal notation.
 Beyond text 397

placed on top of a smaller disk. With just three disks it is not a difficult
problem, and most people solve it after some fumbling.
The strategies people use in solving the Tower of Hanoi problem have
been analyzed and documented by Kotovsky, Hayes, and Simon (1985).
Typically, novices initially adopt a simple difference-reduction strategy,
which proves ineffective. They then shift to a means-end strategy in
which subgoals are created (e.g., “clear everything off the largest disk, so
it can be moved”). Once they adopt the right subgoals, they solve the
problem quickly.
Now consider how one could solve the Tower of Hanoi task by relying
strictly on comprehension processes, that is, by constructing a network
representing this task and using spreading activation to integrate this net¬
work.
We first need a more compact notation, which I have already intro¬
duced in Figure 11.8: (3—2—1, 0, 0) means that the largest disk 3, the
medium disk 2, and the smallest disk 1 are all on the first stick, and that
no disks are on the second and third sticks; (3, 2-1, 0) means that the
largest disk is on the first stick and the other two disks are on the second
stick; (3, 1—2,0) is not a permissible state, because the medium disk 2 can¬
not be placed on top of the small disk 1.
Thus, the anticipated goal state is (0, 0, 3-2-1) - all disks on the third
stick. The task for the comprehender is to construct a network of moves
that link the initial state with this anticipated goal state. The minimum
such network is shown in Figure 11.9. The initial state is labeled
WORLD. There are two possible moves from here, labeled DO: move
disk 1 either to stick 2 or stick 3. Which is the right one? We need to con¬
struct a path linking the given state of the problem in the world with the
anticipated outcome. [0, 2—1, 3] is the first state that links up with the
anticipated goal state: Disk 3 is in the right place. If the network shown
in Figure 11.9 is integrated, the resulting activation values for the two
competing moves — first disk to stick 2 or 3 — clearly favor the coirect
move, and hence a decision that starts the solution process off on the right
path can be made.
But note that in order to link up the current state of the world and the
outcome state, a large and complex network had to be constructed,
including nine intervening nodes! 1 his may not look like much compared
with the networks we are familiar with from analyses of text comprehen-
 398 Models of comprehension

WORLD[3-2-1,0,0] 0.3469
DO[3-2,1,0) 0.0000
DO[3-2,0,1] 0.6836
[3,1,2] 0.0145
[3-1,0,2] 0.0103
[3,0,2-1j 0.0103
[3,2,1] 1.0000
[3-1,2,0] 0.5076
[3,2-1,0] 0.7787
[0,2-1,3] 0.5334
OUTCOME[0,0,3-2-1] 0.271 1

[3,2-1,0]

Figure 11.9 State network for the Tower of Hanoi - initial state. The activa¬
tion values ot the nodes after integration are shown in the left part of the
figure.

sion, but there is a crucial difference. In text comprehension, readers can

rely on well-practiced, automated encoding strategies and their long¬
term working memory, which makes it possible for them to handle large
networks. Here, the task is unfamiliar and the construction rules must be
laboriously applied, resulting in excessive demands on working memory.
This little Gedankenexperiment suggests that the Tower of Hanoi, even
in this simple three-disk version, involves more than mere comprehen¬
sion and requires something else, namely, the problem-solving strategies
Kotovsky et al. (1985) observed in their subjects. Here is a problem
beyond the Cl theory. However, imagine a problem formally equivalent to
the Tower of Hanoi but that does not involve the movement of meaning¬
less disks according to arbitrary rules. Imagine instead a problem in med¬
ical diagnosis and that we are dealing with an expert diagnostician. The
links in this case between the DO nodes and the anticipated OUTCOME
node might be retrieval structures in long-term working memory. What
is a difficult task for the subject in the laboratory might be trivial for the
expert in her or his domain. The Cl theory might possibly describe the
expert s behavior, even if it cannot deal with the laboratory problem.
There are boundaries between comprehension and problem solving, but
to determine just where they he will require further research.
The second illustration to be discussed here is different, in that it is
clearly a problem within the scope of the comprehension theory. It con-
 Beyond text 399

cerns a striking but common error in decision making that Tversky and
Kahneman (1983) have labeled the representativeness bias: Subjects judge
that a conjunction of two sets is more probable than one of the sets
involved if the other set is highly representative. This is a clear violation
of probability theory, for P(A), P(B) < P[AAB] for all sets A and B. Con¬
cretely, if subjects are given the text

(1) Linda is 31 years old, single, outspoken, and very bright. She
majored in philosophy. In college, she was involved in several
social issues, including the environment, the peace campaign,
and the antinuclear campaign.

and are asked to judge which of the following two statements is more
probable:

(2) Linda is a bank teller.

(3) Linda is a bank teller and is active in the feminist movement.

they frequently select the second statement, in violation of the axioms of

probability theory.
Pennington and Hastie (1993) have argued that decision makers do not
always base their decisions on a complete rational analysis of the situation
but often represent a situation as a story, favoring the decision that fits
best into the story context that they have constructed. That is, they form
a mental representation at the narrative level, rather than at an abstract,
formal level. The Cl theory allows us to model both of these alternatives.
First, let us assume that a subject in a decision experiment represents the
Linda text as a narrative. That is, the subject treats the Linda text just as
any other story. The question of interest is how well the test sentences (2)
and (3) will fit into the subject’s story representation. Intuitively, BANK
TELLER is going to be connected weakly to some of the nodes of the
Linda text, as shown in Figure 11.10, whereas the links between the text
propositions and BANK TELLER & FEMINIST ought to be stronger.
To estimate these link strengths objectively, latent semantic analysis
can be used. In the encyclopedia space, the cosines between sentence (2)
and the three sentences of (1) are .04, .02, and .04, respectively. I he
cosines between sentence (3) and the three text sentences aie .09, .06, and
.11, respectively. These cosine values can be used as link strengths in the
network shown in Figure 11.10. Multiplying by 10 puts them on a scale
400 Models of comprehension

q
si ngle

anti-nuclear

figure 11.10 Network corresponding to a story representation of text (1)

linked to the test statements (2) and (3) with link strengths estimates obtained
from LSA.

comparable to the link values between text propositions, which are set to
1. With these values, the spreading activation process yields final activa¬
tion values of .31 for BANK TELLER and .75 for BANK TELLER &
FEMINIST Hence, the model predicts that the more highly activated
BANK TELLER & FEMINIST should be the preferred response.
Does this little exercise in modeling a classical result tell us anything
worthwhile, or is it just that - a little exercise? I think it has significant
implications that ought to be explored. We might not need a representa¬
tiveness bias in order to understand the Linda results. If a subject repre¬
sents the text as a story and the decision is based on activation values in
that representation,Tversky and Kahneman’s result follows. However, a
subject with an elementary knowledge of probability theory may form a
higher-level, abstract, formal representation, as shown in Figure 11.11.
This representation includes the same net as before, with the same link
strengths estimated from LSA, but in addition there is another part of
the net that brings to bear the subject’s knowledge of probability theory
as it applies to this situation. I have indicated the general knowledge of
probability theory in terms of relations among the sets A and B and a the-
 Beyond text 401

single

Figure 11.11 Network corresponding to a story representation of text (1) as

well as a probability theory representation linked to the test statements (2)
and (3) w ith link strengths estimates obtained from LSA.

orem from probability theory, P(A&B) < P(A),P(B), which the decision
maker must know. Set A is linked to BANK4 ELLER, (BT), and set B is
linked to BANKTELLER & FEMINIST, (BT&FEM). The general the¬
orem supports the hypothesis that p(BT) is greater but contradicts the
hypothesis that p(BT&FEM) is greater. As a result, BT receives an acti¬
vation value of .34, but BT&FEM receives only .01. Therefore, a correct
response will be made.
If a decision maker represents the problem formally (which requires not
only a knowledge of probability theory but also that an encoding strategy
must be used to ensure that this knowledge will actually be applied), a
“correct” response will be made. If only a narrative-level repiesentation
is used instead, a response will be made that is incorrect in the formal sense
but consistent with the story representation: From what we know about
Linda, she surely looks like a feminist, even though she has a job as a bank
teller.
The Cl model thus allows us to reanalyze and reinterpret a classical
example from the literature on decision making. But obviously, analyzing
a single example is by no means the same as proposing a new theoiv of
402 Models of comprehension

decision making. Furthermore, our reanalysis is not even original, because

others have advocated this general approach before, and what has been
added here is merely the technical framework of the Cl model. Cheng and
Holyoak (1985), for instance, have distinguished between formal and
informal decision schemas in a manner compatible with our analysis.
More specifically, Pennington and Hastie (1993) have shown that the
informal decision schema often used to make decisions by jurors is the
story schema: Jurors try to make a coherent story out of what they are told.
Our analysis of the Linda story merely shows that the Cl architecture pro¬
vides a convenient framework for modeling such story-based decision
making. More generally, Hammond (1996) has argued for a distinction
between intuitive and formal decision procedures. The Cl model, espe¬
cially in combination with LSA, which permits us to use reasonably accu¬
rate and objective estimates of the semantic relations upon which intuitive
decisions are based, may provide a suitable framework to formalize ideas
that have been proposed within the field of decision making and to relate
these ideas to other comprehension phenomena that have hitherto been
regarded as separate and distinct.
That this is not an idle speculation but a real possibility is nicely illus¬
trated by the model of impression formation that has recently been pro¬
posed by Kunda and Thagard (1996). I describe their work in detail and
show how it can be translated into the Cl framework because it is such a
good example of theorizing in this vein in this field.

11.2.2 Impression formation

How do people judge how friendly, or aggressive, or clever someone is

whom they have just met? Much depends on how this particular person
looks and acts, but it has long been known that the stereotypes people
hold about the social class of the person modify and shape their impres¬
sion about the character of a person. Everyone has certain beliefs about
how sex, age, race, or profession are related to traits and behaviors, and
these expectations combine with what we actually perceive to determine
our impression. Social psychologists have compiled a huge empirical lit¬
erature on these topics, and a number of alternative theoretical view¬
points have been developed over the years. My purpose is not to review
this literature or to evaluate the theories but to suggest that the Cl model
provides a novel framework to view the phenomena of impression forma-
 Beyond text 403

tion and to relate them to other comprehension processes. Recently

Kunda and Thagard (1996) have proposed a parallel constraint satisfac¬
tion theory to account for many interesting findings in the field of
impression formation. Their proposal is closely related to the Cl model.
Like the Cl model, it demonstrates how this type of theory can account
for a rich and complex research area. Because impression formation is not
usually thought of in the same terms as text comprehension, I sketch
Kunda and Thagard’s proposal in some detail and show how it is related
to the Cl model. Both models provide similar accounts of how perceptual
data and information about stereotypes from long-term memory become
merged in impression formation.
Kunda and Thagard (1996) represent a person or event as a network of
interrelated nodes, much as a text is represented in the Cl model. The
nodes are in part based on perception, but there are also knowledge elab¬
orations - the perceiver’s beliefs, stereotypes, and prejudices. Actually,
Kunda and Thagard’s model is not a construction-integration theory,
because it does not deal with the construction of the network at all. The
perceptual factors that determine the saliency of various aspects of a per¬
son’s appearance or action — what makes people pay attention to particu¬
lar aspects and not others - as well as the belief structures and retrieval
processes that determine the knowledge elaborations that the perceiver
contributes to the perceptual network, are assumed as given. Kunda and
Thagard’s theory is concerned with the integration process that gener¬
ates a holistic impression out of these disparate and usually contradictory
elements. The authors can rely on a rich and well-tested expei imental lit¬
erature to help them to identify the salient aspects of a particular prob¬
lem and to set up a plausible network for it. Once such a network has been
generated, they integrate this network, using a mathematical algorithm
that is closely related to the one used here. To show just how closely
related it is, I describe a few of their examples, using their networks but
employing the integration mechanism of the Cl model instead of the one
they use. The results are qualitatively, though not numerically, identical
with the ones reported in Kunda and Thagard (1996).
The heart of their paper is a discussion of twelve phenomena concern¬
ing impression formation that have been studied extensively in the social
psychology literature. They discuss the rich but often confusing empiri¬
cal evidence concerning each one of these phenomena and then show how
a simple and straightforward analysis in terms of a constraint satisfaction
404 Models of comprehension

network provides a satisfactory account for each phenomenon. Further¬

more, the account they provide is consistent across the twelve phenom¬
ena they discuss. Thus, they develop an elegant and comprehensive the¬
ory of impression formation.
It is not my purpose either to describe this theory in detail or to evalu¬
ate it against alternative conceptions. Rather, my point is to demonstrate
its close resemblance to the Cl model in order to show how these phe¬
nomena, too, can be interpreted within the comprehension paradigm of
cognition. Thus, I take the first four impression formation phenomena
analyzed by Kunda and Thagard to illustrate their model, without, how¬
ever, discussing the empirical evidence motivating the networks they set
up or the controversies surrounding this evidence.

Phenomenon 1: Stereotypes color the meaning of behavior. The same

ambiguous action performed by different people will be interpreted dif¬
ferently, depending on stereotypes about the social class to which the
actor belongs. For instance, if a black man shoves someone, the act will be
perceived as more aggressive than when a white man shoves someone.
Kunda and Thagard (1996) conceptualize the situation as shown in Fig¬
ure 11.12. Suppose it has been observed that a black man pushes some¬
one. In addition, there are the two competing interpretation - that at the
shove was friendly or aggressive - plus the stereotype that links the nodes
black and aggressive. With all node and link strengths set to 1 or -1 for the
inhibitory links, the integration process yields the results shown in Fig¬
ure 11.12: The shove is clearly interpreted as violent rather than jovial. If
a white man is doing the shoving, and the observer does not think of white
as aggressive, the interpretation of the shove remains ambiguous, as shown
in Figure 11.13.

Phenomenon 2: Stereotypes color the meaning of traits. What a trait like

aggressive means depends on who is being characterized by that trait.
Lawyers and construction workers are perceived as being about equally
aggressive, but there is a difference between the refined aggression of the
lawyers and the crude aggression of the construction workers. The net¬
works that Kunda and Thagard (1996) set up are shown in Figures 11.14
and 11.15. Aggressive is linked to punch and argue; however, in the context
of lawyer, the punch node is inhibited by the upper-middle-class node char-
 Beyond text 405

observed 1.0000
black 0.9988
pushed-someone 0.9998 A.
aggressive 0.9978
violent-push 0.9984
jovial-push 0.0000

aggressive

Figure 11.12 Stereotypes affect the meaning of behavior. Violent push is

strongly activated by the stereotype that blacks are aggressive. Inhibitory links
are indicated by dashed lines.

observed 0.9269
white 0.5533
pushed-someone 1.0000
violent-push 0.3736 S observed.
jovial-push 0.3736
cr
white "a
pushed-someone

/-\
violent-push jovial-push

Figure 11.13 In the absence of a stereotype, an ambiguous act remains

ambiguous. Violent-push and jovial-push are both weakly but equally activated.

acteristic of lawyers, and the argue node is supported by the verbal node
also characteristic of lawyers. In contrast, construction workers are charac¬
terized by the nodes working class and unrefined, both of which support
punch and are unrelated to argue. The results are displayed in Figures
11.14 and 11.15. Aggressive lawyers argue rather than punch. Aggressive
construction workers argue just as much, but even more likely, they
punch.

Phenomenon 3: Individuating information can determine which oj a stereo¬

type's subtypes is used. Suppose that an observer has a more differentiated
stereotype about blacks than in the first example discussed. The blacks are
406 Models of comprehension

observed 0.4330
9
observed
lawyer 1.0000
aggressive 0.8110
upper-middle class 0.3434
verbal 0.7214
punch 0.2043
argue 0.6647

Figure 11.14 Stereotypes influence the meaning of a trait. Argue is more

activated than punch as characteristics of aggressive in the context of lawyer.

V
observed 0.3563 observed
construction worker 1.0000
aggressive 0.8919
working class 0.6663
unrefined 0.3919
punch 0.8725
argue 0.6351

punch argue

Figure 11.15 Stereotypes influence the meaning of a trait. Punch is more

activated than argue as characteristics of aggressive in the context of construc¬
tion worker.

aggressive stereotype is now reserved for ghetto-blacks, and there is now

another (weaker) subtype of black-businessman, which is not linked to
aggression. Being well dressed is the observable characteristic of a black
businessman in this scenario. Kunda and Thagard (1996) thus arrive at
the network shown in Figure 11.16, where heavy lines have been assigned
a value of 2 to reflect their greater strength. With this more differentiated
stereotype, obseixing a well-dressed black will not activate the aggressive
 Beyond text 407

A
observed
observed 0.6187
black 0.6182
veil-dressed 0.9991
ghetto-black 0.0000
businessman-black 1 .0000
aggressive 0.0000

aggressive

Figure 11.16 Subtypes of stereotypes and individuating information. A

well-dressed black is perceived as a businessman.

P
observed
observed 0.4142
black 1.0000
veil-dressed 0.0000
ghetto-black 0.9997
businessman-black 0.0000
aggressive 0.4141

aggressive

Figure 11.17 Subtypes of stereotypes and individuating information. If

well-dressed is not observed, a black is perceived as an aggressive ghetto black.

node. On the other hand, in the same network, if a black is observed who
is not well dressed, aggressive will become activated, as shown in Figure
11.17.

Phenomenon 4: Stereotypes typically do not affect trait latings in the pus

ence of unambiguous diagnostic information. Although stereotypes can colot
perceptions, they typically do not override it. Thus, il there is strong
unambiguous evidence that is perceptually available, stereotypes ha\e lit-
408 Models of comprehension

tie or no effect. If it is clearly seen that a person punches another adult,

that act will be interpreted as aggressive, whether the person is a house¬
wife, who normally isn’t expected to act that way, or a construction
worker, for whom punching someone is more expected. The calculations
according to the Cl model are shown in Figures 11.18 and 11.19.
The examples described here cannot do justice to Kunda and Tha-
gard’s (1996) interesting paper. By reworking them with the formalism of
the Cl model, I do not intend to claim that one formalism is better than
another but merely to demonstrate that this important area of research in
social psychology can be brought within the scope of a general model
of comprehension. Impression formation is thus seen as having much

observed 0.4477
punch adult 1.0000
housewife 0.0000
aggressive 0.3943

aggressive

Figure 11.18 Unambiguous diagnostic information overrides the stereotype.

A housewife’s punching someone is interpreted as aggressive.

observed 0.5700
punch adult 1.0000
construction worker 0.7850
aggressive 0.8891
SK

F igure 11.19 Unambiguous diagnostic information overrides the stereotype.

A consti uction worker s punching someone is no more aggressive than a
housewife’s punching someone.
 Beyond text 409

in common with text comprehension. In both cases there is a bottom-

up, data-driven construction process that yields a complex network in
which many factors are combined - important and unimportant ones,
mutually supportive and mutually contradictory ones, input data as well
as knowledge elaborations. Integrating this network yields a new, coher¬
ent whole - an impression, or a text representation.

11.3 The representation of the self in working memory

The Cl model focuses on the integration of knowledge and memory and

on texts or other nontextual sources in the process of comprehending a
text or situation. It deals with the environment to be comprehended and
the long-term memory that enables a person to comprehend it. The link¬
age between environment and long-term memory is effected in working
memory, as described by Ericsson and Kintsch (1995).
Long-term memory is conceptualized as a huge network the nodes of
which represent cognitive elements — knowledge, experience, beliefs. A
minute part of this network is activated, available in working memory, and
a larger part is connected by retrieval structures to the contents of work¬
ing memory and is directly retrievable on demand, thus forming a long¬
term working memory.
Suppose we admit to this network not only cognitive elements but also
noncognitive elements, such as emotions and body states. There is noth¬
ing very revolutionary about this proposal because it has already been
argued (in chapter 3 and elsewhere) that the cognitive elements of the
network, such as propositions, are not defined by their content or struc¬
ture — that is merely a crutch for the theorist — but by their position in the
network. Hence, adding noncognitive nodes that define and mutually
constrain each other, in the same way that propositions do, seems like a
natural extension of the original framework.
What I have in mind with noncognitive nodes are brain states that rep¬
resent information about one’s own body. Part of our netwoik is alieady
made up of brain states that represent sensations and perceptions from
the environment. In addition, we now also include feelings about the
body. Damasio (1994) has made some useful suggestions about how this
might work. According to Damasio, the brain represents the state of the
body by means of somatic markers. Certain parts of the brain (the amyg-
410 Models of comprehension

dala and prefrontal cortices) are involved in the control of the state of the
body, effecting changes in adjustments in the body. The body signals
these state changes back to other parts of the brain (the somatosensory
cortex), where these signals may be attended to and become conscious.
Moreover, it is not necessary for the body actually to assume a particular
state. The amygdala and prefrontal cortex may influence the somatosen¬
sory cortex directly, so that it acts as if the body were in a particular state.
Suppose that working memory contains, in addition to the cognitive
nodes that comprise short-term memory or consciousness, a set of
somatic markers that we experience as general body feelings. Introspec-
tively, this appears plausible. These somatic markers in working memory
can serve as cues that, via retrieval structures, can readily retrieve all
kinds of information about one’s own body. Thus, although I am not con¬
scious now of the position of a my right foot, this information is directly
available. People certainly have had the years of practice necessary to
acquire the retrieval structures that provide access to information about
their own body.
If somatic markers are always present in working memory, we must
then modify our view of what is involved in perception. Traditionally, we
think of perception as representing the environment, enriched by the
perceiver’s prior knowledge and beliefs. But if somatic markers are an
important part of working memory, body feelings become parts of the
perceptual experience, just as knowledge and beliefs do. Damasio (1994)
has pointed out that perception consists of more than simply receiving
sensory information. When I look out the window of my study on the
scene below, I see the city nestled in the trees, the flat curve of the hori¬
zon, and the sky above speckled with afternoon clouds. I move my head
as the eyes scan the horizon, the familiar view makes me feel good, and I
stretch my legs and take a deep breath. All of this contributes to the per¬
ception, not just the visual image. When I recreate the image later, traces
of my movements, the way I sat in my chair, and the somatic reactions that
occurred in the first place are regenerated with it. I experience a little of
that good feeling that went with the original perception. Cognition does
not occur in a vacuum or in a disembodied mind but in a perceiving, feel¬
ing, acting body. The image is reconstructed in the brain, and its content
includes the reactions of the body that occurred in response to the origi¬
nal scene - not just what happened on the retina, but bodily reactions in
the widest sense, motor as well as visceral. It is my perception and my
 Beyond text 411

memory image because its content is not only visual but also includes a
variety of somatic reactions of my body.
What we would gain from adding somatic markers as nodes to our net¬
work is the possibility of formulating an account of emotions, attention,
and motivation, as I suggest later. First, however, this would be a way of
representing the self in working memory.
The arena in which cognition takes place is working memory, alias
short-term memory, alias focus of attention, alias consciousness. The
whole model of cognition as it has been constructed here is based on the
contents and operations in working memory, but a moment’s considera¬
tion shows that we have omitted rather a lot from our working memory.
Take the typical reader as modeled by the Cl model: Working memory
contains a dynamically changing stream of about seven (atomic) proposi¬
tions. But there must be more than that. There is an awareness of the
physical environment, especially of changes in it. There is an awareness
not onlv of the goal set for this particular reading task but also of multi¬
ple other goals. There is the constant background feeling of one’s body -
its position, tone, and feeling. There is the self that is reading.
The self is important, for we know quite well from experience, obser¬
vation, as well as experiment, that everything self-related is special for an
organism and receives favored treatment. Yet experimental psychologists
and most cognitive scientists have refused to deal with the self concept.
In part, this refusal has been based on the prevalent view of the self,
which they rightly find unsatisfactory: that the self, in one way or another,
is some sort of homunculus that does the experiencing. However, think¬
ing about cognition in terms of shifting patterns of activation in a large
network suggests a more plausible alternative: that the self does not exist
but is continuously reconstructed as the activated part of the network.
The self nodes consist of some highly and more or less permanently
activated nodes in our long-term memory, including our name, the per¬
sons closest too us, our occupation, and other significant items. These per¬
sonal memory nodes would always be activated, though with some fluctu¬
ations. The memories that make up my self are probably not entirely the
same when I am at home with my family and when I am speaking at a pro¬
fessional meeting. Moreover, such essentially propositional nodes are not
the only ones that constitute the self. There is another group of nodes in
the network representing the body that is part of the self.
I am not only my memories but also my body. These nodes may repre-
412 Models of comprehension

sent limbs and muscles, the position of the body in space, visceral infor¬
mation, as well as the chemical or hormonal balance or imbalance of the
body. I make no attempt either to list or describe all that might be
involved but only point to the existence and significance of this part of
our being, which I lump together as the somatic nodes in the network.
Not all these somatic nodes are so strongly or permanently activated. Our
body sense depends on context: whether I am sitting in my chair or
climbing a mountain, whether I am hungry or satiated, excited or drowsy.
But at any time, there is a group of somatic nodes that is sufficiently acti¬
vated and that, together with the memory nodes, constitute my self — or
more precisely, that is how the Cl theory could model the self.
The two groups of self nodes form a single, tightly interrelated cluster.
In consequence, perception is not just a sensory event but also has an
inherent somatic component. Another way of saying this is that percep¬
tion is not only a cognitive process but also an emotional process. We react
to the world not only with our sense organs but also with gut-level feel¬
ings. The things that excite us, please us, scare us are the ones most
closely linked to the somatic level. Our most central memories are the
ones most intimately linked to our body.
All this is clearly speculation. Furthermore, at this point these specu¬
lations are not sufficiently well formulated to be empirically testable.
However, some support for notions like the ones advanced here exists.
Damasio (1994) has provided evidence from neuroscience for his the¬
ory of somatic markers. He specifically points to the importance of body
reactions for cognition. Feelings about the body — real or vicarious —
become conditioned to certain perceptions and actions and thus become
crucial to the cognitive processes involved in planning and decision mak¬
ing. Actions and perceptions that are tied to good body feelings are
favored over others, and those tied to negative feelings are avoided, thus
directly influencing planning and decision making. The search of prob¬
lem spaces and the rational calculation of outcomes is cut short in this
way, allowing a person to act without becoming lost in thought and plan¬
ning. Damasio reiterates here a point made by Simon (1969) that, far
from being irrelevant by-products of decision making, emotions are cru¬
cial for resource management in a system with limited capacity, such as
human working memory.
There is another line of evidence from cognitive science that appears to
require an explicit account of the self. Many observations about language
 Beyond text 413

and language use have been amassed by such writers as Lakoff (1987) and
Glenberg (1997) to show that language expresses relationships with refer¬
ence to the human body and that we interpret language by means of our
body schema. Through the experience with our own bodies we acquire a
number of schemas, such as the container schema, the part-whole schema,
or the source-path-goal schema. These body schemas are transferred to
language and the world. We understand the world in terms of these
schemas, and language has evolved around the use of them. We know what
a container is because of our direct experience with our body. It has an
inside, an outside, and a boundary, and we can map our understanding of
these relationships, which is extralinguistic and experiential, on the world
by means of metaphorical language. We understand Mary felt trapped in
her marriage and wanted to get out by assigning to marriage the role of the
container, \nfohn and Mary are splitting up, we map marriage into the part-
whole schema instead (examples from Lakoff, 1987). We understand the
world with reference to our own body experience and action; when lan¬
guage evolved, it reflected this bodily basis of understanding.
A working memory in which somatic nodes are continuously active may
provide a mechanism for the explanation of the linguistic phenomena
described by Lakoff (1987). All spatial perception occurs in the context of
activated somatic nodes that represent the position and state of our body.
Thus, from the very beginning, spatial information about the world is
linked to our own body. For instance, the concept more is acquired in a spa¬
tial context, as an increase in quantity along the verticality dimension,
which is given directly in our bodily experience. Once acquired in this con¬
text, the use of the word more is extended metaphorically to nonspatial
contexts, resulting in the more is up, less is down metaphor. Language is
embodied because it is linked to the representations of our own body in
working memory — in its daily use, in its acquisition by the individual, and
in its cultural evolution.

/1.3.1 Evaluative and motivational functions of self-representation

If a model of working memory can be formulated that includes self nodes

in the manner envisaged here, it might be possible to broaden the scope
of comprehension theory to include a number of phenomena that we
know are important but that we heretofore have had no way of dealing
with systematically within a cognitive framework. Most relevant for dis-
414 Models of comprehension

course comprehension is the ubiquitous finding that self-relevant, emo¬

tional material has a special status in comprehension and memory. More
generally, however, we may have a way for dealing with attention, motiva¬
tion, and emotion.

Emotional factors in comprehension. Given that the self is an undifferenti¬

ated whole but a complex structure, one ought to distinguish, at a mini¬
mum, between nodes with positive and negative valences, GOOD and
BAD self-nodes. Propositions in a text (or events in the environment,
in general) are not only self-relevant to varying degrees but may be self¬
relevant in a positive or negative sense, giving rise to positive or negative
emotions. It has often been pointed out that when people read a text, their
emotional responses may be more important and salient than the propo¬
sitional content of the text (for example, Miall, 1989; Oatley, 1992).
I illustrate this point with an example from the literature on decision
biases, continuing the argument introduced in section 11.2.1 that some of
the well-documented decision biases can be understood as comprehen¬
sion problems. If readers construct a text representation at the narrative
level only and base their decisions on the activation values that various
decision-relevant propositions achieve in this representation, it is easy to
see why certain decision errors occur. If, on the other hand, readers con¬
struct an abstract representation, applying schemas appropriate to the
domain (e.g., probability theory in the example we have discussed), they
will arrive at the formally correct answer. This also holds in a context
where emotions play a central role. Consider the following decision prob¬
lem first used byTversky and Kahneman (1983).

(2) Imagine that the United States is preparing for the outbreak of
an unusual Asian disease that is expected to kill 600 people. Two
alternative programs to combat the disease have been proposed.
Assume that the exact scientific estimates of the consequences of
the programs are as follows:

If Program A is adopted, 200 people will be saved,

If Program B is adopted, there is a one-third probability that 600

people will be saved and a two-thirds probability that no people
will be saved.
Which of the two programs would you favor?
 Beyond text 415

People overwhelmingly prefer Program A (72%). Why? The classical

answer is that people’s subjective utility function is negatively acceler¬
ated, so when they compute the utility of the two choices, saving 600
lives is weighted less than three times as highly as saving 200 lives. The
Cl theory suggests a different answer. People may not compute utilities
at all. Instead, Program A will become more activated because it is linked
only to positive self-nodes, whereas Program B will be less activated
because it is risky and hence inhibited by negative self-nodes. Specifi¬
cally, if we let SI, S2, and S3 stand for the first three sentences of (2), we
obtain the network shown in Figure 11.20, in which GOOD and BAD
are self-nodes with their activation values clamped. GOOD is connected
with a strength of 1 to saving 600 people and a strength of .5 to saving
200 people (reflecting the negatively accelerated subjective utility func¬
tion), and BAD is connected with a value of-1 to saving none. All other
links have a value of 1, except for the mutually exclusive decision alter¬
natives, which have a negative link. Integrating this network yields an
activation value of .57 for Progr m A (save-200) and .19 for Program B
(save-600-or-none).
Tversky and Kahneman also showed that if the alternatives in (2) are
reworded as

(3) If Program A is adopted, 400 people will die.

If Program B is adopted, there is a one-third probability nobody
will die and a two-thirds probability that 600 people will die.

Program B is now preferred over Program A (78% vs. 22%). The Cl the¬
ory readily predicts this switch, as Figure 11.21 shows. The negative

Figure 11.20
416 Models of comprehension

Figure 11.21

wording produces inhibition from the BAD node (.5 for 400 dying, 1 for
600 dying), whereas the GOOD node is connected to only none-dying.
Integrating this new network yields much lower activation values for both
choices, but their relative strength is now reversed: Program A has zero
activation value, whereas Program B is favored with an activation value of
.05. I am not aware of any data showing that negatively worded choice-
pairs in planning decisions are less activated than positively worded
choices, but that is a prediction of this kind of model.
Note that there is nothing in the Cl model as such that yields these pre¬
dictions. Rather, they are a consequence of the assumption we made
about the nature of the mental representation that is being constructed.
If people employ a nonanalytical comprehension strategy, treating the
text just like a story, they will make these “errors” in judgment. However,
if they take a more analytic approach and construct an abstract represen¬
tation, employing their knowledge of probability theory, both decision
alternatives will be equally strong (exp-value abbreviates the proposition
expected value equals 200) (see Figure 11.22).
In this case, both save-200 and save-600-or-none receive activation val¬
ues of .41. Thus, it is not a question whether people have “decision
biases but, rather, how they approach the decision problem. If they treat
it as an everyday comprehension problem, their choices will be quite dif¬
ferent than if they employ an analytic strategy, representing the informa¬
tion provided by the text in terms of the appropriate probability schema.
Some decision makers cannot use this strategy because they do not have
the necessary knowledge of probability theory and hence are unable to
construct a representation at this level. Tversky and Kahneman’s point is
that many decision makers who have the right schema still make errors.
Thus, by formally incorporating into the Cl model the self-relevance
of textual information, in both a positive and negative way, some inter-
 Beyond text 417

Figure 11.22

esting and important empirical phenomena may be accounted for within

the framework of the theory.

Attention and motivation. If we admit self-nodes into working memory,

representations of the environment that are linked to these self-nodes will
receive an extra boost in activation. That is, they will more likely be
attended to than environmental stimuli without such links. A tight clus¬
ter of self-nodes in working memory to which some of the propositional
nodes are linked will modulate the pattern of activation in such a way that
self-linked nodes become more strongly activated. Hence, self-relevant
nodes will also obtain a greater long-term memory strength. The likeli¬
hood that they will attain enough strength to become conscious increases,
as does the likelihood that these nodes will be retrieved from memory at
a later time. Indeed, the relatively stable self-nodes could play a signifi¬
cant role as retrieval cues in the process of retrieving other episodic or
semantic memories.
The self-nodes, being permanently activated, thus can serve as ener¬
gizers in working memory, activating other nodes in working memoiy
that they are linked to. If the nodes being activated by the self are repie-
sentations of environmental stimuli, we talk about attention (leav ing aside
the question whether there are sources of attention other than self-rele¬
vance). If the nodes being activated by the self are plans and action
schemas, we talk about motivation (leaving aside the possibility that moti¬
vation need not always be tied to specific goals). Thus, although the view
of working memory explored here may not provide a complete account of
attention and motivation, it might be a useful way to model some aspects
of attention and motivation.
Consider three hypothetical plan nodes in working memory. Plan A is
closely linked to the self; it is associated with pleasant emotions and good
418 Models of comprehension

body states and is positively linked to higher-order plans. Plan B is only

distantly related to the self by some chain of other nodes in long-term
memory. Finally, let Plan C be associated with negative emotional states.
Plan A, because it is linked to highly and permanently activated self¬
nodes, will stay continuously activated, at least at a high enough level that
it could control action under appropriate conditions. Plan B will be much
harder to activate, even if appropriate conditions exist, and Plan C will be
inhibited. It might be possible in this way to build a system that is moti¬
vated, in that it will spontaneously tend to act in certain ways and avoid
acting in other ways.

Consciousness. Plan A does not need to be conscious, however, and neither

do the self-nodes themselves, in spite of their high levels of activation.
The equation of short-term memory/focus-of-attention with conscious¬
ness is no longer permissible in the expanded model of working memory.
Consciousness requires not only a high level of activation in working
memory but also an appropriate form of encoding. Just where the bound¬
aries must be placed is a question that cannot be discussed here, but in
terms of the hierarchy of mental representations described in chapter 2,
a sufficiently high level might have to be reached before the contents of
working memory could become conscious. Thus, even a highly activated
procedural representation may not be conscious, whereas representations
at the linguistic or abstract level would be more likely to be conscious.
Many plans are vague, represented at a level that is not readily accessible
to consciousness. They may be strong and very much part of our self, but
represented at a subsymbolic, sublinguistic, unconscious level. Whether
such an approach to consciousness would have sufficient explanatory
power to be worth pursuing remains to be explored, however.

Interest. Representing the self in working memory may also provide us

with a new way of looking at what kinds of things people find interesting.
I have speculated in Kintsch (1980) that there are two forms of interest.
Some things are inherently interesting, like sex and violence. Anything
that has to do with the self, in fact, is in that category. According to the
extended working memory model, these interests are represented by
nodes in the network that are directly and perhaps multiply linked to the
cluster of self-nodes. Hence, they will be more attended to, better remem¬
bered, and judged to be more interesting than nodes without such direct
 Beyond text 419

links. The other class of interesting things are those that are neither too
familiar nor too strange - objects of intermediate novelty (Berlyne, 1960).
Those are the things that arouse our curiosity, without necessarily being
linked to our own self. This form of interest seems rather different in its
psychological origin than the intrinsic interest based on links to the self.
“Interest,” like so many other psychological concepts, is a folk concept,
and perhaps scientific theory will have to separate what language has put
together. James (1969) thought so for emotions, which he claimed were a
heterogeneous class of phenomena lumped together by our language. In
the Cl theory, it may be possible to treat emotion as a unitary phenome¬
non after all - in much the same way as James wanted to, namely, as per¬
ceptions of bodily states.

11.3.2 Cognition and emotion

In the expanded model of working memory that I have proposed, I have

suggested that attention be treated as links to the self and motivation as
attention to plan nodes. Consider the network of self-nodes again. One can
think of it as a representation of our own state, body, and mind. It is a
dynamic representation, continually changing and continually being-
reconstructed. The representation of the body includes information about
the biochemical state of the brain, the state of the viscera, and current
information about the musculoskeletal frame. The body not only furnishes
a place for our thoughts but contributes content, too (Damasio, 1994).
Much of this representation is at a level not accessible to consciousness,
yielding phenomenologically a vague background feeling of the body.
When new representations about the world or the body are being con¬
structed in working memory as a consequence of the ongoing perception
and action of the organism, these may or may not match the existing self¬
representation. Such mismatches are perceived as emotions or feelings.
A crucial role in these mismatches is played by the plans and goals cur¬
rently active in working memory. An active goal is a node in the self-net,
and if representations are constructed in working memory that conflict
with this goal, an emotion is experienced. These mismatching represen¬
tations may be based on perceptions, but they also may be purely mental
in origin. Because mental representations, in contrast to body represen¬
tations, tend to involve higher levels (language, abstract thought), they
would be more likely to be conscious. Different types of emotions are
420 Models of comprehension

experienced because of the contents of the representations that produce

a mismatch.
The idea that emotions arise out of a mismatch between a representa¬
tion constructed from an experience and our self-representation is found
in Damasio (1994), Oatley and Johnson-Laird (1987), and Oatley (1992),
among others. It seems quite possible to incorporate it into a network
model of cognition and emotion such as the extended working-memory
model sketched here. Working out the details of such a model and testing
its empirical adequacy must remain as a challenge at this time. However,
there are a few topics in the field of cognition and emotion that have been
thoroughly investigated to which the ideas sketched here are clearly
relevant.
There is first the mood dependence of memory. In general, memories
acquired in one emotional state are best retrieved when one is again in the
same emotional state, though the evidence for this claim is neither strong
nor unequivocal (Bower, 1981). The expanded working-memory model
can in principle account for this result as follows. If a memory node is
linked to self-nodes of a particular emotional tone, the reinstatement of
that emotional tone can only facilitate the retrieval of that memory,
though it may not be sufficient by itself. It is interesting that the strongest
evidence for the mood dependence of memory comes from studies of
autobiographical memory (Eich, Macauley, & Ryan, 1994), where one
would expect links to self-nodes to be most relevant.
Although mood may bias memory retrieval, a retrieved memory may
affect mood by the same mechanism. A depressive person may encode an
experience by linking it with negative emotional nodes. If the experience
is reinstated, it will tend to activate these negative emotions. Thus,
depression may feed on itself in a vicious feedback cycle (Baddeley,
1989). The expanded working-memory model may eventually help us to
better understand this process.
Another phenomenon that has been extensively researched is the emo¬
tional distinctiveness of memories. Certain memories - flashbulb memo¬
ries (Neisser & Harsch, 1993) - are associated with very strong emotional
responses. Recall is particularly vivid for such memories. Actually flash¬
bulb memories are not notably accurate; it is just that people have great
confidence in them (Weaver, 1993). Consider two memories, both poorly
encoded and fragmentary, one linked to a very strong, emotional self¬
node, the othei only weakly linked to the self. Suppose that upon retrieval
 Beyond text 421

of these fragmentary nodes, both memories may be reconstructed more or

less faithfully. However, the first reconstruction will be highly activated
because of its strong emotional links, whereas the second will be only
weakly activated. Confidence in the first memory would be high and con¬
fidence in the second would be low, though both could be equally accurate
or inaccurate. An explanation of flashbulb memories along these lines, as
suggested by the expanded working-memory model, might be possible.
Cognition and emotion and cognition and motivation are research top¬
ics for the future. I have not much to say about these topics at present,
except that they are important and should no longer be neglected. I can
do no more here than sketch a possible path that might lead to a serious
investigation of these phenomena in the future, a path that is consistent
with the framework for cognition developed here.

11.4 Outlook

How good a paradigm is comprehension for the study of cognition? More

precisely, how good a paradigm is the model of comprehension that
was developed in these pages, the construction-integration archi¬
tecture? Whether or not the extension beyond cognition that was ten¬
tatively explored in the previous section can be realized is of sec¬
ondary importance. The primary question concerns the adequacy of the
construction-integration framework as a paradigm for cognition proper.
In my not entirely unbiased view the evidence appears rather positive
and the outlook promising. First, the model does well in its core area, text
comprehension. Second, extensions to related cognitive phenomena that
do not, or do not only, involve text comprehension have been successful
and more appear feasible. Third, we are beginning to see the limitations
of the comprehension paradigm; once we know what is lacking, we can try
to supply the missing components of the theory, or, better, attempt to
coordinate the Cl theory with other approaches that complement it in
just the right way.
The field of text comprehension includes most of the central topics of
cognition. Compare the topics discussed in chapters 5 through 10 of the
present volume with the outline of a typical textbook of cognitive psy
chology. Most of the traditional topics of cognitive psychology come up
for discussion in the context of research on text comprehension. Word
identification, lexical access, priming studies, inference, spatial cognition,
422 Models of comprehension

and imagery are represented. Psycholinguistic topics such as sentence

parsing, anaphora resolution, and metaphor have been discussed. Mem¬
ory issues are central to text research: working memory, the theory of
long-term working memory (which in fact has been developed in part
within the context of discourse processing), free recall, cued recall, and
recognition, as well as sentence verification. Problem solving appears in
the research on word arithmetic problems, as well as in the work on action
planning. Skill acquisition is treated in the context of reading skills.
Learning, a topic often neglected in the past by cognitive psychologists,
has become a major concern in text research. Thus, one can more or less
match the major topics of a cognitive psychology text with the chapters
of the present book. Some topics are underrepresented, especially the
general area of perception and attention. This reflects more an idiosyn¬
crasy of the present approach to text research rather than a necessary lim¬
itation - in fact, these topics are central in theories of reading, which
focus more on the decoding end of the process than on the comprehen¬
sion end, as was done here. A few topics considered here are not usually
found in cognitive psychology texts, such as the research on macrostruc¬
tures, which is text specific. But on the whole one can truthfully say that
text research faithfully mirrors the concerns of cognitive psychology in
general. Therefore, to say that the Cl model accounts well for the phe¬
nomena discussed here is a strong claim.
Although the scope of the theory presented here is broad, some limi¬
tations must be acknowledged. Some topics that one would expect in a
book on discourse comprehension are neglected here. The decoding
aspects of reading have already been mentioned. Another major omission
is the whole area of conversation with its fascinating interplay of prag¬
matic and cognitive factors. Of course, there is no reason in principle why
the present approach could not be extended in either direction.
The present chapter has been concerned with another issue: trying to
extend the Cl theory beyond the text or discourse domain. To some
extent, this appears feasible. The Cl model has brought useful insights in
the area of skill acquisition, elucidating the kinds of qualitative differ¬
ences that characterize expert versus novice performance. In the hands of
Kitajima and Poison (1995) (section 11.1.3), this kind of theorizing has
proved to be capable of dealing with complex problem-solving tasks and
is beginning to have practical implications for training. The Cl models
have enabled us to explore the boundaries between automatic compre-
 Beyond text 423

hension and strategic problem-solving processes, even accounting for

how biases and emotions may influence comprehension. There are many
situations in which we comprehend our environment in much the same
way as we comprehend a text. That is, we construct bottom-up, crude,
local interpretations that are then integrated via a constraint satisfaction
process. However, the effectiveness of such processes in complex situa¬
tions is limited; successful problem solving requires more than that on
occasion.
The qualifying phrase on occasion in the previous sentence is impor¬
tant. Although bottom-up construction rules in combination with an
integration process are not sufficient to account for all of cognition, they
account for a great deal. At some point we need something like impasse
recognition and impasse resolution procedures that are more powerful
than the mechanisms described by the Cl theory. What I find surprising
and significant is how far cognitive theories get without such analytic
mechanisms.
Traditionally, cognitive scientists, psychologists, but especially lin¬
guists and philosophers, have thought of cognition (thinking) in terms of
powerful, analytic procedures, modeled more or less closely after the
example of logic. Alternatively, we can look at cognition in a very differ¬
ent way, in terms of separate levels of representation and distinctive
processes operating on these levels. That is the approach of the Cl theory.
The basis of cognition may be an associative system, constructed from
a person’s experience with the world. Latent semantic analysis (LSA) can
be a model for such a system, with the crucial difference that LSA learns
from the dependencies among printed words only, whereas people learn
by interacting with the world and other people, including but not
restricted to reading. Dimension reduction may play a role in enabling
us to generalize from experience, just as for LSA. Many tasks are per¬
formed on the basis of this system alone, or partly on the basis of this sys¬
tem, even tasks that at first glance would seem to require more analytic
processing, as is suggested by our successful use of LSA in a variety of
situations.
The Cl theory adds to this associative basis a powerful computational
device, the constraint-satisfaction mechanism via spreading activation.
The cognitive operations postulated by the Cl theory dominate at the
level of linguistic representations, the narrative-linguistic level of chap¬
ter 2. This is the level at which the comprehension paradigm is most
424 Models of comprehension

appropriate. Initial processes deal with associative relations of different

strengths and responses based directly on these relations. The compre¬
hension mechanism is a way of forming contextually sensitive construc¬
tions from this associative basis.
Human cognition encapsulates different levels of representation, from
direct representations that are not dependent on language to the level of
abstract, analytic thought that are the traditional focus of problem-solv¬
ing theories in cognitive science. Thus, human cognition employs other,
more complex, more analytic computations than can be achieved by
spreading activation networks. It is not my goal here to specify what these
are, or how they interface with the mechanisms described by the Cl the¬
ory. However, a complete theory of cognition has to account for interac¬
tions among at least these three levels of processing and representation:
the associative, the linguistic, and the abstract symbolic.
 References

Albrecht, J. E., O’Brien, E. J., Mason, R. A., & Myers, J. L. (1995). The role of
perspective in the accessibility of goals during reading. Journal of Experi¬
mental Psychology: Learning, Memory, and Cognition, 21, 364—72.
Albrecht, J. E., & Myers, J. L. (1995). The role of context in accessing distant
information during reading. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 21, 1459-68.
Albrecht, J. E., & O’Brien, E. J. (1991). Effects of centrality on retrieval of text-
based concepts. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 17, 932-9.
(1993). Updating a mental model: Maintaining both global and local coher¬
ence. ‘Journal of Experimental Psychology: Learning, Memory, and Cognition,
19, 1061-70.
(1995). Goal processing and the maintenance of global coherence. In R. F.
Lorch & E. J. O’Brien (Eds.), Sources of coherence in reading (pp. 263-78).
Hillsdale, NJ: Erlbaum.
Altmann, G. T. M., & Steedman, M. (1988). Interaction with context during
human sentence processing. Cognition, 30, 191-238.
Anderson, A., Garrod, S., & Sanford, A. J. (1983). The accessibility of pronom¬
inal antecedents as a function of episode shift in narrative text. Quarterly
Journal of Experimental Psychology, 35, 427-40.
Anderson, J. R. (1978). Arguments concerning representations for mental
imagery. Psychological Review, 85, 249—77.
(1983). The architecture of cognition. Cambridge, MA: Harvard University
Press.
(1990). The adaptive character of thought. Hillsdale, NJ: Erlbaum.
(1993). Rules of the mind. Hillsdale, NJ: Erlbaum.
Anderson, J. R., & Bower, G. H. (1972). Human associative memory. Washington,
DC: Winston.
Anderson, R. C., & Ortony, A. (1975). On putting apples into bottles - a prob¬
lem of polysemy. Cognitive Psychology, 7, 167-80.
426 References

Baddeley, A. D. (1986). Working memory. New York: Oxford University Press.

(1989). The psychology of remembering and forgetting. In T. Butler (Ed.),
Memory: History, culture, and mind. London: Basil Blackwell.
Baddeley, A. D., Papagno, C., & Valla, G. (1988). When long-term learning
depends on short-term storage. Journal of Memory and Language, 27,
586-96.
Baddeley, A. D., & Wilson, B. (1988). Comprehension and working memory: A
single-case neuropsychological study. Journal of Memory and Language, 27,
479-98.
Barclay, J. R., Bransford, J. D., Franks, J. J., McCarrell, N. S., & Nitsch, K.
(1974). Comprehension and semantic flexibility. Journal of Verbal Learning
and Verbal Behavior, 13, 471—81.
Barr, A., & Feigenbaum, E. A. (1982). The handbook of artificial intelligence. Los
Altos, CA: William Kaufman.
Barsalou, L. W. (1993). Flexibility, structure, and linguistic vagary in concepts:
Manifestations of a compositional system of perceptual symbols. In A. C.
Collins, S. E. Gathercole, & M. A. Conway (Eds.), Theories of memories.
London: Erlbaum.
Barshi, I. (1997). Message length and misunderstandings in aviation communi¬
cation: Linguistic properties and cognitive constraints. Ph.D. dissertation.
Boulder: University of Colorado.
Bartlett, F. C. (1932). Remembering. Cambridge, UK: Cambridge University
Press.
Bateson, G. (1972). Steps to an ecology of mind. New York: Ballantine.
Battig, W. F. (1979). The flexibility of human memory. In L. S. Cermak &
F. I. M. Craik (Eds.), Levels of processing in human memory (pp. 23M-4).
Hillsdale, NJ: Erlbaum.
Beck, I. L., McKeown, M. G., Sinatra, G. M., & Loxterman, J. A. (1991). Revis¬
ing social studies texts from a text-processing perspective: Evidence of
improved comprehensibility. Reading Research Quarterly, 27, 251-76.
Beck, I. L., McKeown, M. G., Worthy, J., Sandora, C., & Kucan L. (1996).
Questioning the author: A yearlong classroom implementation to engage
students with text. Elementary School Journal, 96, 385-414.
Bereiter, C., & Scardamalia, M. (1989). Intentional learning as a goal of instruc¬
tion. In L. B. Resnick (Ed.), Knowing, learning, and instruction (pp. 361—92).
Hillsdale, NJ: Erlbaum.
Berlyne, D. E. (1960). Conflict, arousal, and curiosity. New York: McGraw-Hill.
Beyer, R. (1990) Psychologische Analyse kogmtiver Prozesse bet der Textverar-
beitung. Berlin: Habilitationsschrift, Humboldt Universitat.
Bierwisch, M. (1969). On certain problems of semantic representation. Founda¬
tions of Language, 5, 153-84.
 References 427

Bisanz, G. L., Das, J. P., Varnhagen, C. K., & Henderson, H. K. (1992). Struc¬
ture and components of reading times and recall for sentences in narra¬
tives: Exploring changes with age and reading ability. Journal of Educational
Psychology, 84, 102-14.
Bisanz, G. L., LaPorte, R. E.,Vesonder, G. T., &Voss, J. F. (1978). Contextual
prerequisites for understanding: Some investigations of comprehension
and recall .Journal of Verbal Learning and Verbal Behavior, 17, 337-57.
Bousfield, W. A., & Sedgwick, C. H. (1944). The analysis of sequences of
restricted associative responses. Journal of General Psychology, 30, 149-65.
Bovair, S., & Kieras, D. E. (1985). A guide to propositional analysis for research
on technical prose. In B. K. Britton & J. B. Black (Eds.), Understanding
expository text (pp. 315-62). Hillsdale, NJ: Erlbaum.
Bower, G. H. (1981). Mood and memory. American Psychologist, 36, 129-48.
Bower, G. H., Black, J. B., & Turner, T. J. (1979). Scripts in memory for text.
Cognitive Psychology, 11, 177-220.
Bransford, J. D., Barclay, J. R., & Franks, J. J. (1972). Sentence memory: A con¬
structive versus interpretive approach. Cognitive Psychology, 3, 193-209.
Bransford, J. D., & Franks, J. J. (1971). The abstraction of linguistic ideas. Cog¬
nitive Psychology, 2, 331-50.
Bransford, J. D., & Johnson, M. K. (1972). Contextual prerequisites for under¬
standing: Some investigations of comprehension and recall. Journal of Ver¬
bal Learning and Verbal Behavior, 11, 717-26.
Briars, D. J., & Larkin, J. H. (1984). An integrated model of skill in solving ele¬
mentary word problems. Cognition and Instruction, 1, 245-96.
Britton, B. K., & Eisenhart, F. J. (1993). Expertise, text coherence, and con¬
straint satisfaction: Effects on harmony and settling rate. In Proceedings of
the Cognitive Science Society (pp. 266-71). Hillsdale, NJ: Erlbaum.
Britton, B. K., & Gulgoz, S. (1991). Using Kintsch’s model to improve instruc¬
tional text: Effects of inference calls on recall and cognitive structures.
Journal of Educational Psychology, 83, 329-45.
Broadbent, D. E. (1975). The magic number seven after fifteen years. In
A. Kennedy & A. Wilkes (Eds.), Studies in long-term memory (pp. 3-18).
London: Wiley.
Brown, A. L., & Palincsar, A. S. (1989). Guided, cooperative learning and indi¬
vidual knowledge acquisition. In L. B. Resnick (Ed.), Knowing, learning,
and instruction. Hillsdale, NJ: Erlbaum.
Brown, J. S., &VanLehn, K. (1980). Repair theory: A generative theory of bugs
in procedural skills. Cognitive Science, 4, 379—426.
Bruner, J. S. (1986). Actual minds, possible worlds. Cambridge, MA: Harvard Uni¬
versity Press.
Cacciari, C., & Glucksberg, S. (1994). Understanding figurative language. In
428 References

M. A. Gernsbacher (Eds.), Handbook of psycholinguistics (pp. 447-78). San

Diego: Academic Press.
Cantor, J., & Engle, R. W. (1993). Working memory capacity as long-term mem¬
ory activation: An individual difference approach. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 19, 1101—14.
Cantor, J., Engle, R. W., & Hamilton, G. (1991). Short-term memory, working
memory, and verbal abilities: How do they relate? Intelligence, 15, 229—46.
Card, S. K., Moran,T. R, & Newell, A. (1983). The psychology of human-computer
interaction. Hillsdale, NJ: Erlbaum.
Carpenter, P. A., &Just, M. A. (1977). Reading comprehension as the eyes see it.
In M. A. Just & P. A. Carpenter (Eds.), Cognitive processes in comprehension.
Hillsdale, NJ: Erlbaum.
Carpenter, T. P, & Moser, J. M. (1983). The acquisition of addition and sub¬
traction concepts. In R. Lesh & M. Landau (Eds.), Acquisition of mathe¬
matical concepts and processes (pp. 7-44). Hillsdale, NJ: Erlbaum.
Chafe, W. L. (1974). Language and consciousness. Language, 58, 111-33.
Chandrasekaran, B., & Narayanan, N. H. (1990). Integrating imagery and visual
representations. In The Twelfth Annual Conference of the Cognitive Science
Society (pp. 670-7). Hillsdale, NJ: Erlbaum.
Charniak, E. (1993). Statistical language analysis. Cambridge, MA: MIT Press.
Chase, W. G., & Ericsson, K. A. (1982). Skill and working memory. In G. H.
Bower (Eds.), The psychology of learning and motivation. New York: Acade¬
mic Press.
Cheng, P. W., & Holyoak, K. J. (1985). Pragmatic reasoning schemas. Cognitive
Psychology, 17, 391-416.
Chi, M. T. H. (1976). Short-term memory limitations in children: Capacity of
processing deficits? Memory 15 Cognition, 4, 545-80.
Chi, M. T. H., Feltovich, P. J., & Glaser, R. (1981). Categorization and represen¬
tation of physics problems by experts and novices. Cognitive Science, 5,
121-52.
Chomsky, N. (1966). Cartesian linguistics. New York: Harper.
Clark, H. H., & Marshall, C. R. (1981). Definite reference and mutual knowl¬
edge. In A. K. Joshi, B. Webster, & I. A. Sagg (Eds.), Elements of discourse
understanding. Cambridge, UK: Cambridge University Press.
Clark, H. H., & Sengul, C. J. (1979). In search of referents for nouns and pro¬
nouns. Memory 15 Cognition, 7, 35—41.
Collins, A. M., & Quillian, M. R. (1969). Retrieval from semantic memory.
nal of Verbal Learning and Verbal Behavior, 8, 240-7.
Conrad, C. (1972). Cognitive economy in semantic memory. Journal of Experi¬
mental Psychology, 92, 149-54.
Cummins, D. D. (1991). Children’s interpretations of arithmetic word problems.
Cognition and Instruction, 8, 261-89.
 References 429

Cummins, D. D„ Kintsch, W., Reusser, K., & Weimer, R. (1988). The role of
understanding in solving word problems. Cognitive Psychology, 20, 405-38.
Damasio, A. R. (1994). Descartes' error: Emotion, reason, and the human brain.
New York: Putnams.
Daneman, M., & Carpenter, P. A. (1980). Individual differences in working
memory and reading. Journal of Verbal Learning and Verbal Behavior, 19,
450-66.
DeCorte, E., Verschaffel, L., & DeWinn, L. (1985). The influence of rewording
verbal problems on children’s problem representation and solutions. Jour¬
nal of Educational Psychology, 77, 460-70.
Deerwester, S., Dumais, S. T., Furnas, G. W., Landauer, T. K., & Harshman, R.
(1990). Indexing by Latent Semantic Analysis. Journal of the American Soci¬
ety for Information Science, 41, 391-407.
Dell, G. S., McKoon, G., & Ratcliff, R. (1983). The activation of antecedent
information during the processing of anaphoric reference in reading. Jour¬
nal of Verbal Learning and Verbal Behavior, 22, 121-32.
Dellarosa, D. (1986). A computer simulation of children’s arithmetic word prob¬
lem solving. Behavior Research Methods, Instruments, (V Computers, 18,
147-54.
Dewey, J. (1897). The University Elementary School: History and character.
University [of Chicago] Record, 2, 72-75.
Doane, S. M., Kintsch, W., & Poison, P. G. (1989). Action planning: Producing
UNIX commands. In The Eleventh Annual Conference of the Cognitive Sci¬
ence Society (pp. 458-65). Hillsdale, NJ: Erlbaum.
Doane, S. M., McNamara, D. S., Kintsch, W., Poison, P. G., & Clawson, D.
(1992). Prompt comprehension in UNIX command production. Memory,
& Cognition, 20, 327-43.
Doane, S. M., Pellegrino, J. W., & Klatzky, R. L. (1990). Expertise in a computer
operating system: Conceptualization and performance. Human-Computer
Interaction, 5, 267—304.
Donald, M. (1991). Origins of the modern mind. Cambridge, MA: Harvard Uni¬
versity Press.
Ehrlich, K. (1980). Comprehension of pronouns. Quarterly Journal of Experi¬
mental Psychology, 32, 247-55.
Ehrlich, K., & Rayner, K. (1983). Pronoun assignment and semantic integration
during reading: Eye-movements and immediacy of processing. Journal of
Verbal Learning and Verbal Behavior, 22, 75—87.
Eich, E., Macauley, D., & Ryan, L. (1994). Mood dependent memory for events
of the personal past. Journal of Experimental Psychology: General, 123,
201-15.
Ericsson, K. A. (1985). Memory skill. Canadian Journal of Psychology, 39, 188—
231.
430 References

Ericsson, K. A., & Charness, N. (1994). Expert performance: Its structure and
acquisition. American Psychologist, 49, 727-47.
Ericsson, K. A., & Chase, W. G. (1982). Exceptional memory. American Scientist,
70, 607-15.
Ericsson, K. A., & Kintsch, W. (1995). Long-term working memory. Psycholog¬
ical Review, 102, 211—45.
Ericsson, K. A., & Simon, H. A. (1993). Protocol analysis: Verbal reports as data
(Rev. ed.). Cambridge, MA: MIT Press.
Estes, W. K. (1986). Array models for category learning. Cognitive Psychology,
18, 500-49.
(1995). A general theory of classification and memory applied to discourse
processing. In W. A. Weaver, S. Mannes, & C. R. Fletcher (Eds.), Discourse
comprehension: Essays in honor of Walter Kintsch (pp. 35-48). Hillsdale, NJ:
Erlbaum.
Farnham-Diggory, S., & J3regg, L. (1975). Short-term memory function in
young readers. Journal of Experimental Child Psychology, 19, 279-98.
Ferreira, F., & Clifton, C. (1986). The independence of syntactic processing.
Journal of Memory and Language, 25, 348-68.
Ferstl, E. C. (1994a). The Construction-Integration Model: A framework for
studying context effects in sentence processing. In The Sixteenth Annual
Conference of the Cognitive Science Society (pp. 289-94). Hillsdale, NJ: Erl¬
baum.
(1994b) Context effects in syntactic ambiguity resolution: The location of preposi¬
tional phrase attachment. Ph.D. dissertation, University of Colorado.
Ferstl, E. C., & Kintsch, W. (in press). Learning from text: Structural knowledge
assessment in the study of discourse comprehension. In S. R. Goldman &
H. van Oostendorp (Eds.), The construction of mental representations during
reading.
Fincher-Kiefer, R. H. (1993). The role of predictive inferences in situation
model construction. Discourse Processes, 16, 99-124.
Fischer, B., & Glanzer, M. (1986). Short-term storage and the processing of
cohesion during reading. Quarterly Journal of Experimental Psychology, 38,
431-60.
Fischer, G., Henninger, S., & Redmiles, D. (1991). Cognitive tools for locating
and comprehending software objects for reuse. In The Thirteenth Interna¬
tional Conference on Software Engineering (pp. 318-28). Austin, TX: ACM,
IEEE.
Fletcher, C. R. (1981). Short-term memory processes in text comprehension.
Journal of Verbal Learning and Verbal Behavior, 20, 264—74.
(1984). Markedness and topic continuity in discourse processing. Journal of
Verbal Learning and Verbal Behavior, 23, 487—93.
 References 431

(1985). Understanding and solving arithmetic word problems: A computer

simulation. Behavior Research Methods, Instruments, and Computers, 17,
565-71.
Fletcher, C. R., & Bloom, C. R (1988). Causal reasoning in the comprehension
of simple narrative texts .Journal of Memory and Language, 27, 235—44.
Fodor, J. A. (1983). The modularity of mind. Cambridge, MA: MIT Press.
Foertsch, J., & Gernsbacher, M. A. (1994). In search of complete comprehen¬
sion: Getting “Minimalists” to work. Discourse Processes, 18, 271-96.
Foltz, P. W. (1996). Latent Semantic Analysis for text-based research. Behavior
Research Methods, Instruments, & Computers, 28, 197-202.
Forster, K. I. (1970). Visual perception on rapidly presented word sequences of
varying complexity. Perception and Psychophysics, 8, 215-21.
Fox, B. A. (1987). Discourse structure and anaphora. New York: Cambridge Uni¬
versity Press.
(1984). Anaphora in popular written English narratives. In R. S. Tomlin
(Eds.), Coherence and grounding in discourse. Amsterdam: Benjamins.
Franzke, M. (1994) Exploration, acquisition, and retention of skill with display-
based systems. Ph.D. Dissertation, University of Colorado.
Frederiksen, J. R. (1981). Sources of process interaction in reading. In A. M.
Lesgold & C. A. Perfetti (Eds.), Interactive processes in reading. Hillsdale,
NJ: Erlbaum.
Furnas, G. W. (1990). Formal models for imaginal deduction. In The Twelfth
Annual Conference of the Cognitive Science Society (pp. 662—9). Hillsdale,
NJ: Erlbaum.
Gallistel, C. R. (1990). The organization of learning. Cambridge, MA: MIT Press.
Garrod, S., Freudenthal, D., & Boyle, E. (1994). The role of different types of
anaphor in the on-line resolution of sentences in a discourse. Journal of
Memory and Language, 33, 38-68.
Garrod, S., O’Brien, E. J., Morris, R. K., & Rayner, K. (1990). Elaborative infer-
encing as an active or passive process. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 16, 250—7.
Garnham, A., Oakhill, J., & Johnson-Laird, P. N. (1982). Referential continuity
and the coherence of discourse. Cognition, 11, 29-46.
Gernsbacher, M. A. (1989). Mechanisms that improve referential access. Cogni¬
tion, 32, 99-156.
(1990). Language comprehension as structure building. Hillsdale, NJ: Erlbaum.
(1993). Less skilled readers have less efficient suppresion mechanisms. Psy¬
chological Science, 4, 294-8.
Gernsbacher, M. A., & Faust, M. E. (1991). The mechanism of suppresssion: A
component of general comprehension. Journal oj Experimental Psychology:
Learning, Memory, and Cognition., 17, 245—62.
432 References

Gernsbacher, M. A., Varner, K. R., & Faust, M. E. (1990). Individual differences

in general comprehension skills. Journal of Experimental Psychology: Learn¬
ing, Memory, and Cognition, 16, 430—45.
Gibbs, R. W. (1994). Figurative thought and figurative language. In M. A. Gerns¬
bacher (Ed.), Handbook of psycholinguistics (pp. 411-46). San Diego: Acade¬
mic Press.
Gibson, J. J. (1977). The theory of affordances. In R. Shaw & J. Bransford (Eds.),
Perceiving, acting, and knowing: Toward an ecological psychology (pp. 67-82).
Hillsdale, NJ: Erlbaum.
Gillund, G., & Shiffrin, R. M. (1984). A retrieval model for both recognition and
recall. Psychological Review, 91, 1—67.
Givon, T. (1983). Topic continuity in discourse. Amsterdam: Benjamins.
(1995). Functionalism and grammar. Amsterdam: Benjamins.
Glanzer, M., Dorfman, D., & Kaplan, B. (1981). Short-term storage in the
processing of text. Journal of Verbal Learning and Verbal Behavior, 20,
656-70.
Glanzer, M., Fischer, B., & Dorfman, D. (1984). Short-term storage in reading.
Journal of Verbal Learning and Verbal Behavior, 23, 467-86.
Glenberg, A. M. (1997). What memory is for. Behavioral and Brain Sciences, 20,
1-55.
Glenberg, A. M., Kruley, P., & Langston, W. E. (1994). Analogical processes in
comprehension. In M. A. Gernsbacher (Ed)., Handbook of Psycholinguistics
(pp. 609-40). San Diego, CA: Academic Press.
Glenberg, A. M., & Langston, W. E. (1992). Comprehension of illustrated text:
Pictures help to build mental models. Journal of Memory and Language, 31,
129-51.
Glucksberg, S., Gildea, R, & Bookian, H. D. (1982). On understanding literal
speech: Can people ignore metaphors? Journal of Verbal Learning and Ver¬
bal Behavior, 21, 85-98.
Goetz, E. T., Anderson, R. C., & Schallert, D. L. (1981). The representation of
sentences in memory. Journal of Verbal Learning and Verbal Behavior, 20,
369-85.
Goldman, S. R., & Varma, S. (1995). CAPping the construction-integration
model of discourse comprehension. In W. A. Weaver, S. Mannes, & C. R.
Fletcher (Eds.), Discourse understanding: Essays in honor of Walter Kintsch.
Hillsdale, NJ: Erlbaum.
Golden, R. M. (1997). Causal network analysis validation using synthetic recall
protocols. Behavior Research Methods, Instruments, IV Computers, 29, 15-24.
Gordon, P. C., & Scearce, K. A. (1995). Pronominalization and discourse coher¬
ence, discourse structure and pronoun interpretation. Memory 15 Cogni¬
tion, 23, 313-23.
Graesser, A. C. (1981). Prose comprehension beyond the word. New York: Springer.
 References 433

Graesser, A. C., Hoffmann, N. L., & Clark, L. F. (1980). Structural components

of reading times. Journal of Verbal Learning and Verbal Behavior, 7 9, 131-51.
Graesser, A. C., & Kreuz, R. J. (1993). A theory of inference generation during
text comprehension. Discourse Processes, 16, 145-60.
Graesser, A. C., Singer, M., &Trabasso, T. (1994). Constructing inferences dur¬
ing narrative text comprehension. Psychological Review, 101, 375-95.
Graesser, A. C., & Zwaan, R. A. (1995). Inference generation and the construc¬
tion of situation models. In C. A. Weaver, S. Mannes, & C. R. Fletcher
(Eds.), Discourse comprehension: Essays in honor of Walter Kmtsch (pp. 117-
39). Hillsdale, NJ: Erlbaum.
Graf, P., & Schacter, D. L. (1985). Implicit and explicit memory for new associ¬
ations in normal and amnesic subjects. Journal of Experimental Psychology:
Learning, Memory and Cognition, 11, 501-18.
Greene, S. B., Gerrig, R. J., McKoon, G., & Ratcliff, R. (1994). Unheralded pro¬
nouns and management by common ground. Journal of Memory and Lan¬
guage, 33, 511-26.
Greene, S. B., McKoon, G., & Ratcliff, R. (1992). Pronoun resolution and dis¬
course models. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 18, 266—83.
Greeno, J. G. (1989). Situations, mental models, and generative knowledge. In D.
Klahr & K. Kotovsky (Eds.), Complex information processing (pp. 285-318).
Hillsdale, NJ: Erlbaum.
(1994). Gibson’s affordances. Psychological Review, 101, 336—42.
(1995). Understanding concepts in activity. In C. A. Weaver, S. Mannes,
& C. R. Fletcher (Eds.), Discourse comprehension: Essays in honor of Walter
Kintsch (pp. 65-96). Hillsdale, NJ: Erlbaum.
Groen, G. J., & Patel, V. L. (1988). The relationship between comprehension and
reasoning in medical expertise. In M. T. H. Chi, R. Glaser, & M. J. Farr
(Eds.), The nature of expertise (pp. 287—310). Hillsdale, NJ: Erlbaum.
Grosz, B. J., & Sidner, C. (1986). Attention, intention, and the structure of dis¬
course. Computational Linguistics, 12, 175-204.
Guindon, R., & Kintsch, W. (1984). Priming macropropositions: Evidence for
the primacy of macropropositions in memory. Journal of Verbal Learning
and Verbal Behavior, 23, 508-18.
Guthke, T. (1991). Psychologische Untersuchungen zu Inferenzen beim Satz und
Textverstehen. Berlin: Ph.D. dissertation, Humboldt Universitat.
Guthrie, J. T., & Greaney,V. (1991). Literacy acts. In R. Barr, M. L. Kamil, P B.
Mosenthal, & P. D. Pearson (Eds.), Handbook of reading research (pp. 68-
96). New York: Longman.
Haberlandt, K., & Bingham, G. (1984). The effect of input direction on the pro¬
cessing of script statements. Journal of Verbal Learning and Verbal Behavior,
23, 162-77.
434 References

Haberlandt, K., & Graesser, A. C. (1989). Processing of new arguments at clause

boundaries. Memory 15 Cognition. 17, 86-193.
Haberlandt, K., Graesser, A. C., Schneider, N. J., & Kiely, J. (1986) Effects of
task and new arguments on word reading times. Journal of Memory and
Language, 25, 314-22.
Haenggi, D., Kintsch, W., & Gernsbacher, M. A. (1995). Spatial situation mod¬
els and text comprehension. Discourse Processes, 19, 173—200.
Hall, R., Kibler, D., Wenger, E., & Truxaw, C. (1989). Exploring the episodic
structure of algebra story problem solving. Cognition and Instruction, 6,
223-83.
Hammond, K. R. (1996). Human judgment and social policy: Irreducible uncer¬
tainty, inevitable errors, unavoidable injustice. New York: Oxford University
Press.
Hanson, S. J., Kraut, R. E., & Farber, J. M. (1984). Interface design and multi¬
variate analysis of UNIX command use. ACM Transactions on Office Infor¬
mation Systems, 2, 42-57.
Henley, N. M. (1969). A psychological study of the semantics of animal terms.
Journal of Verbal Learning and Verbal Behavior, 8, 176—84.
Hintzman, D. U. (1988). Judgments of frequency and recognition memory in a
multiple-trace memory model. Psychological Review, 95, 528-51.
Hudson, T. (1983). Correspondences and numerical differences between disjoint
sets. Child Development, 54, 84-90.
Hull, C. L. (1943). Principles of behavior. New York: Appleton-Century.
Hutchins, E. (1995). Cognition in the wild. Cambridge, MA: MIT Press.
Huttenlocher, J., & Burke, D. (1976). Why does memory span increase with age?
Journal of Verbal Learning and Verbal Behavior, 8, 1-31.
James, W. (1969). What is emotion? In William James: Collected essays and
reviews. New York: Russell and Russell. (Originally published in 1884.)
Jenkins, J. J. (1974). Remember that old theory of memory? Well, forget it! Amer¬
ican Psychologist, 29, 785-95.
Johnson, C. S. (1967). Hierarchical clustering schemes. Psychometnka, 32, 241—
54.
Johnson-Laird, P. N. (1983). Mental models. Cambridge, MA: Harvard Univer¬
sity Press.
Johnson-Laird, P. N., Byrne, R. M. J., & Schaecken, W. (1992). Propositional
reasoning by model. Psychological Review, 99, 418-39.
Jurafsky, D. (1996). A probabilistic model of lexical semantic access and disam¬
biguation. Cognitive Science, 20, 177-94.
Just, M. A., & Carpenter, P. A. (1980). A theory of reading: From eye fixations to
comprehension. Psychological Review, 87, 329-54.
 References 435

(1987). The psychology of reading and language to comprehension. Boston: Allyn

& Bacon.
(1992). A capacity theory of comprehension: Individual differences in work¬
ing memory. Psychological Review, 99, 122-49.
Karmiloff-Smith, A. (1992). Beyond modularity. Cambridge, MA: MIT Press.
Katz, J. J., & Fodor, J. A. (1963). The structure of semantic theory. Language, 39,
170-210.
Kelly, M. C., & Rubin, D. C. (1988). Natural rhythmic patterns in English: Evi¬
dence from child counting-out rhymes. Journal of Memory and Language,
27, 718—40.
Kieras, D. E. (1980). Initial mention as a signal to thematic content in technical
passages. Memory & Cognition, 8, 345—53.
Kimble, G. A. (1996). Psychology: The hope of a science. Cambridge, MA: MIT
Press.
Kintsch, E. (1990). Macroprocesses and microprocesses in the development of
summarization skill. Cognition and Instruction, 7, 161-95.
Kintsch, E., & Kintsch, W. (1995). Strategies to promote active learning from
texts: Individual differences in background knowledge. Swiss Journal of
Psychology, 54, 141-51.
Kintsch, W. (1966). Recognition learning as a function of the length of the reten¬
tion interval and changes in the retention interval. Journal of Mathematical
Psychology, 3, 412-33.
(1972). Notes on the structure of semantic memory. In E. Tulving & W. Don¬
aldson (Eds.), Organization and memory (pp. 247-308). New York: Acade¬
mic Press.
(1974). The representation of meaning in memory. Hillsdale, NJ: Erlbaum.
(1976). Memory for prose. In C. N. Cofer (Ed.), The structure of human mem-
ory (pp. 90-113). San Francisco: Freeman.
(1977a). Memory and cognition. New York: Wiley.
(1977b). On comprehending stories. In M. A. Just & P. Carpenter (Eds.), Cog¬
nitive processes in comprehension (pp. 33—61). Hillsdale, NJ: Erlbaum.
(1980). Learning from text, levels of comprehension, or: Why anyone would
read a story anyway. Poetics, 9, 87-98.
(1985). Text processing: A psychological model. InT. A. van Dijk (Ed.), Hand¬
book of discourse analysis (pp. 231—43). London: Academic Press.
(1987). Generating scripts from memory. In E. van der Meer & J. Hoffmann
(Eds.), Knowledge aided information processing (pp. 61—80). Amsterdam:
Elsevier.
(1988). The use of knowledge in discourse processing: A construction-
integration model. Psychological Review, 95, 163—82.
436 References

(1989). The representation of knowledge and the use of knowledge in dis¬

course comprehension. In R. Dietrich & C. F. Graumann (Eds.), Language
processing in social context (pp. 185-209). Amsterdam: North Holland.
(1991). A theory of discourse comprehension: Implications for a tutor for
word algebra problems. In M. Pope, R. J. Simmons, J. I. Pozo, & M. Car-
retero (Eds.), Proceedings of the 1989 EARLI conference (pp. 235-53). Lon¬
don: Pergamon Press.
(1992a). A cognitive architecture for comprehension. In H. L. Pick, P. van-
denBroek, & D. C. Knill (Eds.), The study of cognition: Conceptual and
methodological issues (pp. 143-64). Washington, DC: American Psychologi¬
cal Association.
(1992b). How readers construct situation models for stories: The role of syn¬
tactic cues and causal inferences. In A. F. Healy, S. M. Kosslyn, & R. M.
Shiffrin (Eds.), From learning processes to cognitive processes: Essays in honor
of William K. Estes (pp. 261-78). Hillsdale, NJ: Erlbaum.
(1993). Information accretion and reduction in text processing: Inferences.
Discourse Processes, 16, 193—202.
(1994a). Discourse processes. In G. d’Ydewalle, P. Eelen, & P. Bertelson
(Eds.), International perspectives on psychological science, vol. 2: The state of
the art (pp. 135-55). Hillsdale, NJ: Erlbaum.
(1994b). Kognitionspsychologische Modelle desTextverstehens: Literarische
Texte. In K. Reusser & M. Reusser (Eds.), Verstehen (pp. 39-54). Bern:
Hans Huber.
(1994c). Learning from text. American Psychologist, 49, 294-303.
(1996). Lernen ausTexten. In J. Hoffmann & W. Kintsch (Eds.), Lernen (pp.
503-28). Gottingen: Hogrefe.
Kintsch, W., & Greene, E. (1978). The role of culture-specific schemata in the
comprehension and recall of stories. Discourse Processes, 1, 1-13.
Kintsch, W., & Greeno, J. G. (1985). Understanding and solving word arithmetic
problems. Psychological Review, 92, 109-29.
Kintsch, W., & Keenan, J. M. (1973). Reading rate and retention as a function of
the number of propositions in the base structure of sentences. Cognitive
Psychology, 5, 257-79.
Kintsch, W., & Kozminsky, E. (1977). Summarizing stories after reading and lis¬
tening. Journal of Educational Psychology, 69, 491—9.
Kintsch, W., Kozminsky, E., Streby, W. J., McKoon, G., & Keenan, J. M. (1975).
Comprehension and recall as a function of content variables. Journal of Ver¬
bal Learning and Verbal Behavior, 14, 196-214.
Kintsch, W., & Lewis, A. B. (1993). The time course of hypothesis formation in
solving arithmetic word problems. In M. Denis & G. Sabah (Eds.), Mode-
 References 437

les et concepts pour la science cognitive: Homrnage a Jean-Franfois Le Ny (pp.

11—23). Grenoble: Press Universitaires.
Kintsch, W., & Mannes, S. M. (1987). Generating scripts from memory. In
E. van der Meer & J. Hoffmann (Eds.), Knowledge-aided information pro¬
cessing (pp. 61-80). Amsterdam: Elsevier Science Publishers.
Kintsch, W., & Mross, E. F. (1985). Context effects in word identification. Jour¬
nal of Memory and Language, 24, 336—49.
Kintsch, W., & van Dijk, T. A. (1978). Towards a model of text comprehension
and production. Psychological Review, 85, 363-94.
Kintsch, W., & Welsch, D. M. (1991). The construction-integration model: A
framework for studying memory for text. In W. E. Hockley & S. Lewan-
dowsky (Eds.), Relating theory and data: Essays on human memory in honor of
Bennett B. Murdock (pp. 367-85). Hillsdale, NJ.: Erlbaum.
Kintsch, W., Welsch, D. M., Schmalhofer, F., & Zimny, S. (1990). Sentence mem¬
ory: A theoretical analysis. Journal of Memory and Language, 29, 133-59.
Kintsch, W., & Witte, R. S. (1962). Concurrent conditioning of bar press and
salivary responding. Journal of Comparative and Physiological Psychology,
55, 963-8.
Kintsch, W., & Yarborough, J. J. (1982). Role of rhetorical structure in text com¬
prehension. Journal of Educational Psychology. 74, 828-34.
Kitajima, M., & Poison, P. G. (1995). A comprehension-based model of correct
performance and errors in skilled, display-based human-computer interac¬
tion. International Journal of Human-Computer Studies, 43, 65-99.
Kohler, I. (1962). Experiments with goggles. Scientific American (May issue).
Kolodner, J. L. (1993). Case-based reasoning. Hillsdale, NJ: Erlbaum.
Kosslyn, S. M. (1980). Image and mind. Cambridge, MA: Harvard University
Press.
(1994). Image and brain. Cambridge, MA: MIT Press.
Kotovsky, K., Hayes, J. R., & Simon, H. A. (1985). Why are some problems hard?
Evidence from Tower of Hanoi. Cognitive Psychology, 17, 248-94.
Kunda, Z., &Thagard, P (1996). Forming impressions from stereotypes, traits,
and behaviors: A parallel constraint satisfaction theory. Psychological
Review, 103, 284—308.
Lakoff, G. (1987). Women, fire, and dangerous things. Chicago: University of
Chicago Press.
Landauer, T. K., & Dumais, S. T. (1997). A solution to Plato’s problem: The
Latent Semantic Analysis theory of acquisition, induction and representa¬
tion of knowledge. Psychological Review, 104, 211—40.
Landauer, T. K., Foltz, P. W., & Laham, D. (in press). An introduction to Latent
Semantic Analysis. Discourse Processes.
438 References

Larkin, J. H. (1989). Display-based problem solving. In D. Klahr & K. Kotovsky

(Eds.), Complex information processing (pp. 319-432). Hillsdale, NJ: Erl-
baum.
(1983). Expert and novice differences in solving physics word problems. In D.
Gentner & A. L. Stevens (Eds.), Mental models. Hillsdale, NJ: Erlbaum.
Lave, J. (1988). Cognition in practice: Mind, mathematics, culture in everyday life.
New York: Cambridge University Press.
Lehnert, W. G. (1981). Plot units and narrative summarization. Cognitive Sci¬
ence, 4, 293—332.
Leibowitz, H. W., & Post, R. B. (1982). The two modes of processing concept
and some implications. In J. Beck (Ed.), Organization and representation in
perception (pp. 343-64). Hillsdale, NJ: Erlbaum.
Lesgold, A. M., Roth, S. F., & Curtis, M. E. (1979). Foregrounding effects in
discourse comprehension. Journal of Verbal Learning and Verbal Behavior,
18, 291-308. ^ '

Linde, C. (1979). Focus of attention and the choice of pronouns. In T. Givon

(Ed.), Syntax and semantics. New York: Academic Press.
Long, D. L., & Golding, J. M. (1993). Superordinate goal inferences: Are they
automatically generated during reading? Discourse Processes. 16, 55-73.
Long, D. L., Golding, J. M., & Graesser, A. C. (1992). A test of the on-line sta¬
tus of goal-related inferences. Journal of Memory and Language, 31, 634—47.
Long, D. L., Oppy, B. J., & Seely, M. R. (1994). Individual differences in the time
course of differential processing. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 20, 1456-70.
Lorch, R. F., Lorch, E. P, & Inman, W. E. (1993). Effects of signalling topic
structure on text recall. Journal of Educational Psychology, 85, 281-90.
Lorch, R. F., Lorch, E. R, & Mathews, P. D. (1985). On-line processing of the
topic structure of a text .Journal of Memory and Language, 24, 350-62.
Lord, A. B. (1960). The singer of tales. Cambridge, MA: Harvard University
Press.
Magliano, J. P, Baggett, W. B., Johnson, B. K., & Graesser, A. C. (1993). The
time course of generating causal antecedent and causal consequence infer¬
ences. Discourse Processes, 16, 35-53.
Mandler, J. M., & Johnson, N. S. (1977). Remembrance of things parsed: Story-
structure and recall. Cognitive Psychology, 9, 111—51.
Mani, K., & Johnson-Laird, P. N. (1982). The mental representation of spatial
descriptions. Memory & Cognition, 10, 181-7.
Mannes, S. M. (1994). Strategic processing of text. Journal of Educational Psy¬
chology, 86, 577-88.
Mannes, S. M., & Kintsch, W. (1987). Knowledge organization and text organi¬
zation. Cognition and Instruction, 4, 91-115.
 References 439

(1991). Planning routine computing tasks: Understanding what to do. Cogni¬

tive Science, 15, 305-42.
Marr, D. (1982). Vision. San Francisco: Freeman.
Masson, M. E. J., & Miller, J. A. (1983). Working memory and individual differ¬
ences in comprehension and memory of text. Journal of Educational Psy¬
chology, 75, 314-18.
Mayer, R. E. (1981). Frequency norms and structural analysis of algebra story
problems into families, categories, and templates. Instructional Science, 10,
135-75.
McClelland, J. L., & Rumelhart, D. E. (1986). Parallel distributed processing.
Cambridge, MA: MIT Press.
McConkie, G. W., & Zola, D. (1981). Language constraints and the functional
stimulus in reading. In A. M. Lesgold & C. A. Perfetti (Eds.), Interactive
processes m reading. Hillsdale, NJ: Erlbaum.
McDaniel, M. A., Blischak, D., & Einstein, G. I. (1995). Understanding the spe¬
cial mnemonic characteristics of fairy tales. In I. C. A. Weaver, S. Mannes,
& C. A. Fletcher (Eds.), Discourse comprehension: Essays in honor of Walter
Kintsch (pp. 157-76). Hillsdale, NJ: Erlbaum.
McDonald, J. L., & MacWhinney, B. (1995). The time course of anaphora reso¬
lution: Effects of implicit verb causality and gender .Journal of Memory and
Language, 34, 543-66.
McKeown, M. G., Beck, I. L., Sinatra, G. M., & Loxterman, J. A. (1992). The
contribution of prior knowledge and coherent text to comprehension.
Reading Research Quarterly, 27, 78-93.
McKoon, G, & Ratcliff, R. (1988). Contextually relevant aspects of meaning.
Journal of Experimental Psychology: Learning, Memory, and Cognition, 14,
331-43.
(1992). Inference during reading. Psychological Review, 99, 440-66.
(1995). The minimalist hypothesis: Directions for research. In I. C. A. Weaver,
S. Mannes, & C. R. Fletcher (Eds.), Discourse comprehension: Essays in honor
of Walter Kintsch (pp. 97-116). Hillsdale, NJ: Erlbaum.
McNamara, D. S., & Kintsch, W. (1996). Learning from text: Effect of prior
knowledge and text coherence. Discourse Processes, 22, 247-88.
McNamara, D. S., & Healy, A. F. (1995). A generation advantage for multiplica¬
tion skill and nonword vocabulary acquisition. In A. F. Healy & L. E.
Bourne (Eds.), Learning and memory of knowledge and skills (pp. 132-69).
Newbury Park, CA: Sage.
McNamara, D. S., Kintsch, E., Songer, N. B., & Kintsch, W. (1996). Are good
texts always better? Text coherence, background knowledge, and levels of
understanding in learning from text. Cognition and Instruction, 14, 1-43.
McNamara, D. S., & Kintsch, W. (1996). Working memory in text comprehen-
440 References

sion: Interrupting difficult text. In The Seventeenth Proceedings of the Cog¬

nitive Science Society. Hillsdale, NJ: Erlbaum.
Meyer, B. J. F. (1975). The organization of prose and its effect on memory. Amster¬
dam: North Holland.
Meyer, D. E., & Schvaneveldt, R. W. (1971). Facilitation in recognizing pairs of
words: Evidence of a dependence between retrieval operations, fournal of
Experimental Psychology, 90, 227—34.
Miall, S. D. (1989). Beyond the schema given: Affective comprehension of liter¬
ary narratives. Cognition and Emotion, 3, 55-78.
Miller, G. A. (1956). The magical number seven, plus or minus two: Some limits
of our capacity for processing information. Psychological Review, 63, 81-97.
Miller, J. R., & Kintsch, W. (1980). Readability and recall for short passages: A
theoretical analysis. Journal of Experimental Psychology: Human Learning
and Memory, 6, 335-54.
Minsky, M. (1975). A framework for representing knowledge. In R H. Winston
(Ed.), The psychology of computer vision. New York: McGraw-Hill.
Mishkin, M., & Petri, H. L. (1984). Memories and habits: Some implications for
the analysis of learning and retention. In L. R. Squire & N. Butters (Eds.),
Neuropsychology of memory (pp. 287-96). New York: Guilford Press.
Miyake, A., Just, M. A., & Carpenter, P. A. (1994). Working memory constraints
on the resolution of lexical ambiguity: Maintaining multiple interpretations
in neutral contexts. Journal of Memory and Language, 33, 175—202.
Moravcsik, J. E., & Kintsch, W. (1993). Writing quality, reading skills, and
domain knoweldge as factors in text comprehension. Canadian Journal of
Experimental Psychology, 47, 360-74.
Morrow, D. G., Greenspan, S. L., & Bower, G. H. (1987). Accessibility and sit¬
uation models in narrative comprehension. Journal of Memory and Lan¬
guage, 26, 165-87.
Mross, E. F. (1989) Macroprocessing in expository text comprehension. Ph.D. dis¬
sertation, University of Colorado, Boulder.
Mross, E. F., & Roberts, J. (1992). The construction-integration model: A program
manual. No. 92-14. University of Colorado, Institute of Cognitive Science.
Murdock, B. B., Jr. (1982). A theory for the storage and retrieval of items and
associative information. Psychological Review, 89, 609-26.
(1993). TODAM2: A model for the storage and retrieval of item, associative,
and serial order information. Psychological Review, 100, 187-203.
Murray, L., & Trevarthen, C. (1985). Emotional regulation of interactions
between two-month-olds and their mothers. In T. M. Field & N. A. Fox
(Eds.), Social perception in infants (pp. 177-97). Norwood, NJ: Ablex.
Myers, J. L., O Brien, E. J., Albrecht, J. E., & Mason, R. A. (1994). Maintaining
 References 441

global coherence during reading. Journal of Experimental Psychology:

Learning, Memory, and Cognition, 20, 876-86.
Myers, J. L., Shinjo, M., & Duffy, S. A. (1987). Degree of causal relatedness and
memory. Journal of Memory and Language, 26, 453-65.
Nathan, M. J. (1991) A tutor for word algebra problems. Ph.D. dissertation, Uni¬
versity of Colorado.
Nathan, M. J., Kintsch, W., & Young, E. (1992). A theory of word algebra prob¬
lem comprehension and its implications for the design of learning environ¬
ments. Cognition and Instruction, 9, 329-89.
Neisser, U. (1994). Multiple systems: A new approach to cognitive theory. The
European Journal of Cognitive Psychology, 6, 225—42.
Neisser, U., & Harsch, N. (1993). Phantom flashbulbs: False recollections of
hearing the news about Challenger. In E. Winograd & U. Neisser (Eds.),
Affect and accuracy in recall: Studies of flashbulb memory. New York: Cam¬
bridge University Press.
Nelson, K. (1996). Language in cognitive development: The emergence of the medi¬
ated mind. New York: Cambridge University Press.
Newell, A. (1990). Unified theories of cognition. Cambridge, MA: Harvard Uni¬
versity Press.
Newell, A., & Simon, H. A. (1972). Human problem solving. Englewood Cliffs,
NJ: Prentice-Hall.
(1976). Computer science as empirical inquiry: Symbols and search. Commu¬
nications of the ACM, 19, 113-26.
Noordman, L. G. M., &Vonk, W. (1992). Readers’ knowledge and the control of
inferences in reading. Language and Cognitive Processes, 7, 373-91.
Norman, D. A. (1981). The trouble with UNIX. Datamation, 139-50.
Oatley, K. (1992). Best laid schemes: The psychology of emotion. New York: Cam¬
bridge University Press.
Oatley, K., & Johnson-Laird, P. N. (1987). Toward a cognitive theory of emotion.
Cognition and Emotion, 1, 29-50.
O’Brien, E. J. (1995). Automatic components of discourse comprehension. In
R. F. Lorch & E. J. O’Brien (Eds.), Sources of coherence in reading (pp.
159-76). Hillsdale, NJ: Erlbaum.
O’Brien, E. J., & Albrecht, J. E. (1991). The role of context in accessing
antecedents in text. Journal of Experimental Psychology: Learning, Memory,
and Cognition, 17, 94-102.
O’Brien, E. J., Shank, D. M., Myers, J. L., & Rayner, K. (1988). Elaborative
inferences during reading: Do they occur on-line? Journal of Experimental
Psychology: Learning, Memory, and Cognition, 14, 410-20.
Olson, J. R., & Biolsi, K. J. (1991). Techniques for representing expert knowl-
442 References

edge. In K. A. Ericsson & J. Smith (Eds.), Toward a general theory of exper¬

tise (pp. 240-85). Cambridge, UK: Cambridge University Press.
Omanson, R. C. (1982). The relation between centrality and story category vari¬
ation. Journal of Verbal Learning and Verbal Behavior, 21, 326—37.
Otero, J., & Kintsch, W. (1992). Failures to detect contradictions in text: What
readers believe vs. what they read. Psychological Science, 3, 229-34.
Paige, J. M., & Simon, H. A. (1966). Cognitive processes in solving algebra word
problems. In B. Kleinmuntz (Ed.), Problem solving. New York: Wiley.
Patel, V. L., Arocha, J. F., & Kaufman, D. R. (1994). Diagnostic reasoning and
medical expertise. In D. L. Medin (Ed.), The psychology of learning and
motivation (pp. 187-252). San Diego: Academic Press.
Pearlmutter, N.J., & MacDonald, M. C. (1995). Individual differences and prob¬
abilistic constraints in syntactic ambiguity resolution. Journal of Memory
and Language, 34, 521-42.
Pennington, N., & Hastie, R. (1993). Explanation-based decision making:
Effects of memory structure on judgment. Journal of Experimental Psychol¬
ogy: Learning, Memory, and Cognition., 14, 521-33.
Perfetti, C. A. (1985). Reading Ability. New York: Oxford University Press.
(1993). Why inferences might be restricted. Discourse Processes. 16, 181-92.
Perfetti, C. A., & Britt, M. A. (1995). Where do propositions come from? In
C. A. Weaver, S. Mannes, & C. R. Fletcher (Eds.), Discourse comprehension:
Essays in honor of Walter Kintsch pp. (11-34). Hillsdale, NJ: Erlbaum.
Perrig, W., & Kintsch, W. (1985). Propositional and situational representations
of text .Journal of Memory and Language, 24, 503-18.
Posner, M. I., & McCandliss, B. D. (1993). Converging methods for investigat¬
ing lexical access. Psychological Science, 4, 305-09.
Potter, M. C. (1983). Representational buffers: The eye-mind hypothesis in pic¬
ture perception, reading, and visual search. In K. Rayner (Ed.), Eye move¬
ments in reading: Perceptual and language processes (pp. 413-37). New York:
Academic Press.
Poulsen, D., Kintsch, E., Kintsch, W., & Premack, D. (1979). Children’s com¬
prehension and memory for stories. Journal of Experimental Child Psychol¬
ogy, 28, 379-403.
Pylyshyn, Z. W. (1981). The imagery debate: Analogue media versus tacit knowl¬
edge. Psychological Review, 88, 16-45.
(1984). Computation and cognition. Cambridge, MA: MIT Press.
Raajmakers, J. G., & Shiffrin, R. M. (1981). Search of associative memory. Psy¬
chological Review, 88, 93-34.
Ratcliff, R., & McKoon, G. (1978). Priming in item recognition: Evidence for
the propositional structure of sentences. Journal of Verbal Learning and Ver¬
bal Behavior, 17, 403-18.
 References 443

Rayner, K., Pacht, J. M., & Duffy, S. A. (1994). Effects of prior encounter and
global discourse bias on the processing of lexically ambiguous words: Evi¬
dence from eye fixations .Journal of Memory and Language, 33, 527—44.
Recht, D. R., & Leslie, L. (1988). Effect of prior knowledge on good and poor
readers. Journal of Educational Psychology, 80, 16-20.
Reder, L. M. (1982). Plausibility judgments versus fact retrieval: Alternative
strategies for sentence verification. Psychological Review, 89, 250—80.
(1987). Strategy selection in question answering. Cognitive Psychology, 19,
90-38.
Rehder, R., Schreiner, M. E., Wolfe, M. B. W., Laham, D., Landauer, T. K., &
Kintsch, W. (in press). Using Latent Semantic Analysis to assess knowl¬
edge: Some technical considerations. Discourse Processes.
Reusser, K. (1989). From text to situation to equation: Cognitive simulation of
understanding and solving mathematical word problems. In H. Mandl,
E. DeCorte, N. Bennett, & H. F. Friedrich (Eds.), Learning and instruction
(pp. 477-98). Oxford: Pergamon Press.
(1993). Tutoring systems and pedagogical theory. In S. Lajoie & S. Derry
(Eds.), Computers as cognitive tools (pp. 143—77). Hillsdale, NJ: Erlbaum.
Richman, H. B., Staszewski, J. J., & Simon, H. A. (1995). Simulation of expert
memory using EPAM IV. Psychological Review. 102, 305-30.
Riley, M. S., Greeno, J. G., & Heller, J. H. (1983). Development of children’s
problem solving ability in arithmetic. In H. P. Ginsburg (Ed.), The develop¬
ment of mathematical thinking (pp. 153-96). New York: Academic Press.
Rips, L. J. (1994). The psychology of proof: Deductive reasoning in human thinking.
Cambridge, MA: MIT Press.
Rizzo, N. D. (1939). Studies in visual and auditory memory span with specific
reference to reading disability. Journal of Experimental Education, 8, 208-
44.
Rodenhausen, H. (1992). Mathematical aspects of Kintsch’s model of discourse
comprehension. Psychological Review, 99, 547-49.
Rosch, E., & Mervis, C. B. (1975). Family resemblance studies in the internal
structure of of categories. Cognitive Psychology, 7, 573-605.
Rubin, D. C. (1988). Learning poetic language. In F. S. Kessel (Ed.), The devel¬
opment of language and language researchers: Essays in honor of Roger Brown
(pp. 339-51). Hillsdale, NJ: Erlbaum.
Rumelhart, D. E., & McClelland, J. L. (Ed.). (1986). Parallel distributed process¬
ing. Cambridge, MA: MIT Press.
Sanford, A. J., & Garrod, S. C. (1981). Understanding written language. London:
Wiley.
Sanford, A. J., Moar, K., & Garrod, S. (1988). Proper names as controllers of dis¬
course focus. Language and Speech, 31, 43-56.
444 References

Santa, J. L. (1977). Spatial transformations of words and pictures. Journal of

Experimental Psychology: Human Learning and Memory, 3, 418-27.
Schank, R. C. (1982). Dynamic memory. Cambridge, UK: Cambridge University
Press.
Schank, R. C., & Abelson, R. P. (1977). Scripts, plans, goals, and understanding.
Hillsdale, NJ: Erlbaum.
Schmalhofer, F. (1986). Verlaufscharakteristiken des Informationsabruf beim
Wiedererkennen und Verifizieren von Saitzen. Zeitschrift fir Expenmentelle
und Angewandle Psychologic, 33, 133—49.
Schmidt, H. G., & Boshuizen, H. P. A. (1993). On the origin of intermediate
effects in clinical case recall. Memory £> Cognition, 21, 338-51.
Schmidt, R. A., & Bjork, R. A. (1992). New conceptualizations of practice: Com¬
mon principles in three paradigms suggest new concepts for training. Psy¬
chological Science, 3, 207-17.
Schneider, W., Korkel, J., & Weinert, F. E. (1989). Domain-specific knowledge
and memory performance: A comparison of high- and low-aptitude chil¬
dren. Journal of Educational Psychology, 81, 306—12.
Schonpflug, W., & Esser, K. B. (1995). Memory and its Graeculi: Metamemory
and control in extended memory systems. In C. A. Weaver, S. Mannes, &
C. R. Fletcher (Eds.), Discourse comprehension: Essays in honor of Walter
Kintsch (pp. 245-56). Hillsdale, NJ: Erlbaum.
Schvaneveldt, R. W. (Ed.). (1990). Pathfinder associative networks: Studies in
knowledge organization. Norwood, NJ: Ablex.
Schwanenflugel, P. J., & C. R. White (1991). The influence of paragraph infor¬
mation on the processing of upcoming words. Reading Research Quarterly,
26, 160-77.
Selz, O. (1922). Zur Psychologic des Denkens und Irrtums. Bonn: Cohen.
Shepard, R. N. (1967). Recognition memory for words, sentences, and pictures.
Journal of Experimental Psychology, 62, 302-9.
Shepard, R. N., & Metzler, J. (1971). Mental rotation of three-dimensional ob¬
jects. Science, 171, 701-3.
Shepard, R. N., & Teghtsoonian, M. (1961). Retention of information under
conditions approaching a steady state. Journal of Experimental Psychology,
62, 302-9.
Simon, H. A. (1969) The sciences of the artificial. Cambridge, MA: MIT Press.
(1995). The information-processing theory of mind. American Psychologist,
50, 507-8.
Singer, M. (1982). Comparing memory for natural and laboratory reading. Jour¬
nal of Experimental Psychology: General, 111, 331-47.
Singer, M., Graesser, A. C., & Trabasso, T. (1994). Minimal or global inference
during reading. Journal of Memory and Language, 33, 421-41.
Singer, M., & Kintsch, W. (in preparation). Recognition memory for sentences.
 References 445

Singer, M., & Ritchot, K. F. M. (1996). The role of working memory capacity
and knowledge access in text inference processing. Memory & Cognition,
24, 733-43.
Skinner, B. F. (1938). The behavior of organisms. New York: Appleton-Century-
Crofts.
Smith, E. E., & Medin, D. L. (1981). Categories and concepts. Cambridge, MA:
Harvard University Press.
Smith, E. E., Shoben, E. J., & Rips, L. J. (1974). Structure and process in seman¬
tic memory: A feature model for semantic decision. Psychological Review,
81, 214-41.
Smolensky, P. (1986). Information processing in dynamical systems: Founda¬
tions of harmony theory. In D. F. Rumelhart & J. L. McClelland (Eds.),
Parallel distributed processes. Cambridge, MA: MIT Press.
Spence, K. W. (1956). Behavior theory and conditioning. New Haven, CT: Yale
University Press.
Squire, L. R. (1992). Memory and the hippocampus: A synthesis of findings
with rats, monkeys, and humans. Psychological Review, 99, 195-231.
Squire, L. R., & Knowlton, B. (1995). Learning about categories in the absence
of memory. Proceedings of the National Academy of Sciences, 92, 12470.
Squire, L. R., Knowlton, B., & Musen, G. (1993). The structure and organiza¬
tion of memory. Annual Review of Psychology, 44, 453-95.
Stanners, R. F., Price, J. M., & Painton, S. (1982). Interrelationships among text
elements in fictional prose. Applied Psycholinguistics. 3, 95-107.
Stanovich, K. E. (1980). Toward an interactive-compensatory model of indi¬
vidual differences in reading fluency. Reading Research Quarterly, 16,
32-1.
Staub, F. C., & Reusser, K. (1995). The role of presentational structures in
understanding and solving mathematical word problems. In C. A. Weaver,
S. Mannes, & C. R. Fletcher (Eds.), Discourse comprehension: Essays in honor
of Walter Kintsch (pp. 285-306). Hillsdale, NJ: Erlbaum.
Steinhart, D. J. (1996) Resolving lexical ambiguity: Does context play a role? MA
thesis, University of Colorado.
Stern, E. (1993a). What makes certain arithmetic word problems involving com¬
parison of sets so difficult for children? Journal of Educational Psychology,
85, 7-23.
(1993b). Die Entwicklung des mathematischen Verstjindnisses im Kindes-
alter. Habilitationsschrift, Universitiit Munchen.
Stern, E., & Lehrndorfer, A. (1992). The role of situational context in solving
word problems. Cognitive Development, 7, 259-68.
Sternberg, S. (1969). Memory scanning: Mental processes revealed by reaction
time experiments. American Scientist, 57, 421—57.
Stigler, J. W., Fuson, K. C., Ham, M., & Kim, M. S. (1986). An analysis of addi-
446 References

tion and subtraction word problems in American and Soviet elementary

mathematics textbooks. Cognition and Instruction, 3, 153-71.
Suchman, L. A. (1987). Plans and situated actions: The problem of human-machine
communication. Norwood, NJ: Ablex.
Swinney, D. A. (1979). Lexical access during sentence comprehension: (Recon¬
sideration of context effects. Journal of Verbal Learning and Verbal Behav¬
ior, 18, 645-59.
Tapiero, I., & Denhiere, G. (1995). Simulating recall and recognition using
Kintsch’s Construction-Integration model. In C. A. Weaver, S. Mannes, &
C. R. Fletcher (Eds.), Discourse comprehension: Essays in honor of Walter
Kintsch (pp. 211-32). Hillsdale, NJ: Erlbaum.
Taraban, R., & McClelland, J. L. (1988). Constituent attachment and thematic
role assignment in sentence processing: Influence of content-based expec¬
tations. Journal of Memory and Language, 27, 597-632.
Taylor, H. A., & Tversky, B. (1992). Spatial models derived from survey and
route descriptions. Journal of Memory and Learning, 31, 261-92.
Thomson, D. M. (1972). Context effects in recognition memory. Journal of Ver¬
bal Learning and Verbal Behavior, 11, 497-511.
Till, R. E., Mross, E. F., & Kintsch, W. (1988). Time course of priming for asso¬
ciate and inference words in a discourse context. Memory & Cognition, 16,
283-98.
Tolman, E. C. (1932). Purposive behavior in animals and men. New York: Century.
Trabasso, T, & van den Broek, R (1985). Causal thinking and the representation
of narrative events .Journal of Memory and Language, 24, 612-30.
Trabasso, T., & Suh, S. (1993). Understanding text: Achieving explanatory
coherence through on-line inferences and mental operations in working
memory. Discourse Processes, 16, 3-34.
lulving, E., & Thomson, D. M. (1973). Encoding specificity and retrieval
processes in episodic memory. Psychological Review, 80, 352-73.
Tversky, A., & Kahneman, D. (1983). Extensional and intuitive reasoning: The
conjunction fallacy in probability judgment. Psychological Review, 90,
293-315.
Tyler, L. K., & Marslen-Wilson, W. D. (1982). Processing utterances in dis¬
course contexts: On-line resolution of anaphors. Journal of Semantics 1
297-315.
Underwood, B. J., & Freund, J. S. (1968). Errors in recognition learning and
retention. Journal of Experimental Psychology, 78, 55-63.
van Dijk, T. A. (1972). Some aspects of text grammars. The Hague: Mouton.
(1980). Macrostructures. The Hague: Mouton.
van Dijk, T. A., & Kintsch, W. (1983). Strategies of discourse comprehension. New
York: Academic Press.
 References 447

van den Broek, P. (1990). Causal inferences in the comprehension of narrative

texts. In A. C. Graesser, & G. H. Bower (Eds.), Psychology of learning and
motivation: Inferences and text comprehension (pp. 175-94). San Diego: Aca¬
demic Press.
van den Broek, P, Risden, K., Fletcher, C. R., & Thurlow, R. (1996). A “Land¬
scape” view of reading: Fluctuating patterns of activation and the con¬
struction of a stable meaning. In B. K. Britton & A. C. Graesser (Eds.),
Models of understanding text (pp. 165-88). Hillsdale, NJ: Erlbaum.
Varma, S., & Goldman, S. R. (1996). Resonance-based reinstatement of ana¬
phoric antecedents. Paper presented at the Seventh Annual Winter Confer¬
ence, Jackson Hole, WY.
Vonk, W., Hustinx, L. G. M. M., & Simons, W. H. G. (1992). The use of refer¬
ential expressions in structuring discourse. Language and Cognitive
Processes. 7, 301-33.
Voss, J. F., & Silfies, L. N. (1996). Learning from history texts: The interaction
of knowledge and comprehension skill with text structure. Cognition and
Instruction, 14, 45—68.
Vygotsky, L. S. (1986). Thought and language. Cambridge, MA: MIT Press.
Walker, C. H. (1987). Relative importance of domain knowledge and overall apti¬
tude on acquisition of domain-related knowledge. Cognition and Instruction,
4, 25-42.
Walker, C. H., & Yekovich, F. R. (1984). Script-based inferences: Effects of text
and knowledge variables on recognition memory. Journal of Verbal Learn¬
ing and Verbal Behavior, 23, 357—70.
Walker, W. H., & Kintsch, W. (1985). Automatic and strategic aspects of knowl¬
edge retrieval. Cognitive Science, 9, 261-83.
Wallace, W. T., & Rubin, D. C. (1988). “The wreck of the old 97”: A real event
remembered in song. In U. Neisser & E. Winograd (Eds.), Remembering
reconsidered: Ecological and traditional approaches to the study of memory (pp.
283-310). New York: Cambridge University Press.
Wanner, E. (1975). On remembering, forgetting and understanding sentences. The
Hague: Mouton.
Weaver, C. A. (1993). Do you need a “flash” to form a flashbulb memory? Jour¬
nal of Experimental Psychology: General, 122, 39-46.
Weinert, F. E., & Schneider, W. (Eds.), (in press). Individual development from 3
to 12: Findings from the Munich Longitudinal Study. Cambridge: Cambridge
University Press.
Weisskranz, L., Warrington, E. K., Sanders, M. D., & Marshall, J. (1974). Visual
capacity in the hemianopic field following restricted occipital ablation.
Brain, 97, 709-28.
Wertheimer, M. (1945). Productive thinking. New York: Harper.
448 References

Whitney, P., Budd, D., Bramuci, R. S., & Crane, R. C. (1995). On babies, bath¬
water, and schemata: A reconsideration of top-down processes in compre¬
hension. Discourse Processes, 20, 135—66.
Whitney, P., Ritchie, B. G.,& Clark, M. B. (1991). Working memory capacity and
the use of elaborative inferences in text comprehension. Discourse Processes,
14, 133-46.
Wisniewski, E. J. (1995). Prior knowledge and functionally relevant features in
concept learning. Journal of Experimental Psychology: Learning. Memory,
and Cognition, 21, 449-468.
Wolfe, M. B. W., & Kintsch, W. (in preparation). Story recall: Joining compre¬
hension theory and memory theory.
Wolfe, M. B., Schreiner, M. E., Rehder, R., Laham, D., Foltz, P. W., Landauer,
T. K., & Kintsch, W. (in press). Learning from text: Matching reader and
text by Latent Semantic Analysis. Discourse Processes.
Yekovich, F. R., Walker, C. H., Ogle, L. T., & Thompson, M. A. (1990). The
influence of domain knowledge on inferencing in low-aptitude individuals.
In A. C. Graesser & G. H. Bower (Eds.), The psychology of learning and moti¬
vation (pp. 175-96). New York: Academic Press.
Zwaan, R. A., Langston, M. C., & Graesser, A. C. (1995). The construction of
situation models in narrative comprehension: An event-indexing model.
Psychological Science, 6, 292-97.
Zwaan, R. A., Magliano, J. P., & Graesser, A. C. (1995). Dimensions of situation
model construction in narrative comprehension. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 21, 386-97.
 Name Index

Abelson, R. P., 36, 37, 94 Bingham, G., 82

Aesop, 268 Biolsi, K. J., 295
Albrecht, J. E., 175n, 181, 190, Bisanz, G. L., 282-3, 285, 286,
194,235 297
Altman, G. T. M., 172, 173 Bjork, R. A., 320
Anderson, A., 148 Black, J. B., 37, 82
Anderson, J. R., 2, 6, 7, 42, 44, Blischak, D., 319
45, 248, 373n Bloom, C. P., 234
Anderson, R. C., 71, 135 Bookian, H. D., 82
Aristotle, 35-6 Boshuizen, H. R A., 233
Bousfield, W. A., 83
Baddeley, A. D., 125, 126, 127, Bovair, S., 55
217, 239, 420 Bower, G. H., 37, 82, 201,202,
Baggett, W. B., 194, 198 248, 420
Barclay, J. R., 78, 104f, 105, 135, Boyle, E., 148, 149, 152, 153, 154
191,192 Bramuci, R. S., 37
Barr, A., 36 Bransford, J. D., 78, 104f, 105,
Barsalou, L. W., 30, 44, 79, 80 135, 191, 192, 244, 287, 291
Barshi, I., 70 Briars, D. J., 334
Bartlett, F. C., 29, 68, 94, 265 Britt, M. A., 60n
Bateson, G., 31 Britton, B. K., 169, 309, 310,
Battig, W. F., 320 311,313,317
Beck, I. L., 312, 329 Broadbent, D. E., 217
Bereiter, C., 329 Brown, A. L., 329
Berlyne, D. E., 419 Brown, J. S., 386
Beyer, R., 308, 309f, 311,313 Bruner, J. S., 27
Bierwisch, M., 37n Budd, D., 37
450 Name Index

Burke, D., 238 Deerwester, S., 87

Byrne, R. M. J., 192 Dell, G. S., 145, 146
Dellarosa, D., 338
Cacciari, C., 158 Denhiere, G., 266
Cantor, J., 239, 240 Descartes, R., 13
Card, S. K., 387 Dewey, J., 329
Carpenter, P. A., 101, 124, 148, DeWinn, L., 341
223,238-9, 240, 241,242, Doane, S. M., xv, 97, 381, 382,
243, 283, 284 383,384,386,388,391
Carpenter, T. R, 338 Donald, M., 27, 28,31
Chafe, W. L., 148, 226 Dorfman, D., 222, 223t, 230, 234
Chandrasekaren, B., 47 Duffy, S. A., 129, 145,320
Charness, N., 381 Dumais, S.T.,87, 88, 89, 134
Charniak, E., 174n
Chase, W. G., 212, 219 Eco, U., 157
Chatfield, N., 278 Ehrlich, K., 148, 149
Cheng, P. W., 402 Eich, E., 420
Chi, M. T. H., 238, 381 Einstein, G. E, 319
Chomsky, N., 32n, 344 Eisenhart, F. J., 169
Clark, H. H., 148, 235 Engle, R. W., 239, 240
Clark, L. E, 282, 285 Ericsson, K. A., xvi, 3, 76, 77,
Clark, M. B., 244 118, 190,212,213,216,
Clawson, D., 97, 381, 384, 385f, 217, 218f, 219, 220f, 222,
386,388, 391 223,231,238, 239, 241,
Clifton, C., 173 244, 245,381,396, 409
Collins, A. ML, 36, 75 Esser, K. B., 29
Conrad, C., 36 Estes, W. K., 35, 90
Crane, R. C., 37
Cummins, D. D., xv, 54, 338, Farber,J. M., 388
339, 340, 344, 345, 346, Farnham-Diggory, S., 238
347,354, 359 Faust, M. E., 284
Curtis, M. E., 148 Feigenbaum, E. A., 36
Feltovich, P. J., 381
Damasio, A. R., 409, 410, 412, Ferreira, F., 173
419,420 Ferstl, E. C., xvi, 101, 168, 170,
Daneman, M., 238-9, 240 171, 172, 173, 174, 297,
Das, J. P, 282-3, 285, 286 299f, 300, 301,302
DeCorte, E., 341 Fincher-Kiefer, R. H., 199
 Name Index 451

Fischer, B., 222 Glenberg, A. M., 44, 198, 413

Fischer, G., xv, 107, 371 Glucksberg, S., 82, 158
Fletcher, C. R., xv, 148, 156, Goetz, E.T., 71
179, 180, 198, 199, 234, 338 Golden, R. M., 268
Fodor, J. A., 34, 128, 344 Golding, J. M., 194
Foertsch, J., 193 Goldman, S. R., 99, 235
Foltz, P. W., 87, 88, 89, 90, 324, Gordon, P. C., 148
326f, 327 Graesser, A. C., 193, 194, 198,
Forster, K. I., 72 282,285
Fox, B.A., 156, 179 Graf, P, 25
Franks, J.J., 78, 104f, 105, 135, Greaney,V., 332
191, 192, 291 Greene, E., 68
Franzke, M., xvi, 390 Greene, S. B., 148, 156, 226
Frederiksen, J. R., 283 Greeno, J. G., xv, 23, 24, 29, 334,
Freudenthal, D., 148, 149, 152, 335—6t, 335, 337, 338, 339,
153,154 346, 347, 354, 359, 365
Freund, J. S., 248 Greenspan, S. L., 201, 202
Furnas, G. W., 45, 87 Gregg, L., 238
Fuson, K. C., 333 Groen, G. L., 233
Grosz, B. J., 156
Gallistel, C. R.,20,21 Guindon, R., 174, 175, 176n
Garnham, A., 148 Gulgoz, S., 309, 310, 311, 313,
Garrod, S., 148, 149, 152, 153, 317
154, 156, 194 Guthke, T., 189
Gernsbacher, M. A., 144, 147, Guthrie, J. T., 332
169n, 193, 201, 202f, 203f,
204,223,284 Flaberlandt, K., 82, 285
Gerrig, R. J., 148, 156, 226 Haenggi, D., 201, 202f, 203f, 204
Gibbs, R. W., 157, 161 Hall, R., 358, 363
Gibson, J. J., 21, 23, 29 Ham, M., 333
Gildea, R, 82 Hamilton, G., 239
Gillund, G., 35, 235, 249, 253, Hammond, K. R., 402
254, 256,257, 261,267, Hanson, S. J., 388
268, 273, 274, 275, 276 Harsch, N., 420
Givon,T., 68, 156, 226 Harshman, R., 87
Glanzer, M., 222, 223t, 230, 231, Hastie, R., 399, 402
234 Hayes, B., 247
Glaser, R., 381 Hayes, J. R., 397, 398
452 Name Index

Healy, A. F., 319 Kelly, M. C., 207, 208

Heller, J. H., 334, 335-6t, 335, Kibler, D., 358, 363
339,359 Kiely, J., 285
Henderson, H. K., 282-3, 285, Kieras, D. E., 55, 174
286 Kim, M. S., 333
Henley, N. M., 297 Kimble, G. A., 20
Henninger, S., 107, 371 Kintsch, E., xv, xvi, 50, 68, 207,
Hintzman, D. L., 35, 249, 253 309f, 311, 313, 315f, 316f,
Hoffman, N. L., 282, 285 318, 320, 322, 323
Holyoak, K. J., 402 Kintsch, W., 20, 34n, 37, 38,
Homer, 212 39, 46, 49, 54, 60n, 61, 67,
Hudson, T., 334, 346 68, 69-70, 72, 76, 77, 78,
Hull, C. L., 20 81, 83, 84f, 85, 95, 96, 97,
Hustinx, L. G. M. M., 180 101, 102, 104f, 105, 106,
Hutchins, E., 31 107, 108, 118, 130, 131,
Huttenlocher, J., 238 132, 133, 134, 136, 142,
158, 168, 174, 175, 176n,
Inman, W. E., 179 177, 181, 183,189, 190,
197, 199, 200, 201, 202f,
James, W., 419 203f, 204, 213, 216, 217,
Jenkins, J. J., 216 218f, 219, 220f, 221, 222,
Johnson, B. K., 194, 198 223,228, 229, 230, 231,
Johnson, C. S., 298 234, 238, 239, 241,244,
Johnson, M. K., 244, 287 245, 248, 251, 252f, 253f,
Johnson, N. S., 211 264, 265, 266, 274, 275,
Johnson-Laird, P. N., 118, 148, 276, 278, 285, 286, 287,
192,198,420 288f, 289f, 292, 293, 294f,
Jurafsky, D., 174n 297, 299,300, 301,302,
Just, M.A., 101, 124, 148, 223, 303f,306f, 310, 311,313,
239, 240, 241,242, 243, 315f, 316f, 317, 318, 319f,
283,284 320, 32If, 323, 324, 326f,
327, 335, 337, 338, 339,
Kahneman, D., 399, 400, 414, 340, 344, 345, 346, 347,
415,416 348, 354, 355, 358, 359,
Kaplan, B., 223, 230, 234 360, 361, 364, 365, 366,
Karmiloff-Smith, A., 27, 30 367f, 372, 376, 379, 380,
Katz, J. J., 34 381,382,383,384, 385f,
Keenan, J. M., 72, 197, 221, 229, 386,388, 391,396, 409,
285 418
 Name Index 453

Kitajima, M., 97, 388, 389, 390, Long, D. L., 132, 194, 198, 232,
391, 392f, 422 284, 285, 286
Klatzky, R. L., 382, 383 Lorch, E. P., 176, 179
Knowlton, B., 25, 26 Lorch, R. F., 176, 179
Kohler, I., 22 Lord, A. B., 212
Kolodner, J. L., 211 Loxterman, J. A., 312
Korkel, J., 287
Kosslyn, S. M., 44, 45 McCandliss, B. D., 126, 127
Kotovsky, K., 397, 398 McCarrell, N. S., 78, 135
Kozminsky, E., 174, 221, 285 Macauley, D., 420
Kraut, R. E., 388 McClelland, J. L., 35, 170, 171
Kreuz, R. J., 193 McConkie, G. W., 283
Kruley, P., 198 McDaniel, M. A., 319
Kucan, L., 329 McDonald, J. L., 148
Kunda, Z., 394, 402, 403, 404, MacDonald, M. C., 242
406, 408 McKeown, M. G., 312, 329
McKoon, G., 73,78, 145, 146,
Laham, D., 87, 88, 90, 324, 326f, 148, 156, 174, 193,221,
327 226,285
Lakoff, G., 43, 157,413 McNamara, D. S., xv, 97, 230,
Landauer, T. K., xv, xvi, 87, 309f, 311,313, 315f, 316f,
88, 89, 90, 134, 324, 326f, 317,318,319, 320, 323,
327 381,384, 385f, 386, 388,
Langston, M. C., 198 391
Langston, W. E., 198 MacWhinney, B., 148
LaPorte, R. E., 297 Magliano, J. P., 194, 198
Larkin, J. H., 31, 334, 339, 365, Mandler, J. M.,211
381 Mani, K., 198
Lave, J., 23 Mannes, S. M., xv, 37, 83, 84f,
Lehnert, W. G., 211 85,97,320, 32If, 322, 372,
Lehrndorfer, A., 346 376,379, 380,388,391
Leibniz, G. W. F., 32 Marr, D., 5
Leibowitz, H. W., 24 Marshall, C. R., 235
Lesgold, A. M., 148 Marshall, J., 24
Leslie, L., 287 Marslen-Wilson, W. D., 148
Lewis, A. B., 348, 349, 352, 353, Mason, R. A., 181, 190, 194, 235
354 Masson, M. E. J., 239
Lewis, C., xv Mathews, P. D., 176, 179
Linde, C., 179 Mayer, R. E., 359
454 Name Index

Medin, D. L., 34, 35 Noordman, L. G. M., 195-6, 197

Mervis, C. B., 35 Norman, D. A., 381
Metzler, J., 44
Meyer, B. J. F., 222 Oakill, J., 148
Meyer, D. E., 36 Oatley, K., 414, 420
Miall, S.D.,414 O’Brien, E.J., 175n, 181, 190,
Miller, G. A., 217 194,235
Miller, J. A., 239 Ogle, L.T., 289
Miller, J. R., 264, 266,310 Olson, J. R., 295
Minsky, M., 36 Omanson, R. C., 286
Mishkin, M., 25 Oppy, B. J., 132, 198, 232, 284,
Miyake, A., xvi, 243 285,286
Moar, K., 148 Or tony, A., 135
Moran, T. P., 387 Otero, J., xv, 181, 183
Moravcsik, J. E., xv, 104f, 105,
287, 288f, 289f, 291 Pacht,J. M., 129, 145
Morris, R. K., 194 Paige, J. M., 366
Morrow, D. G., 201, 202 Painton, S., 297
Moser, J. M., 338 Palinscar, A. S., 329
Mross, E. F., xv, 100, 101, 130, Papagno, C., 127
131, 132, 133, 134, 139n, Patel, V. L., 233
177-8, 179, 198, 228, 254, Pearlmutter, N. J., 242
286 Pellegrino, J. W., 382, 383
Murdock, B. B., Jr., 35, 249, 253 Pennington, N., 399, 402
Murray, L., 22 Perfetti, C.A.,60n, 199, 282,
Musen, G., 25 283
Musil, R., 205 Perrig, W., xv, 199, 200, 201
Myers, J. L., 181, 190, 194, 235, Petri, H. L., 25
238,320 Poison, P G., xv, xvi, 97, 381,
382, 383, 384, 385f, 386,
Narayanan, N. H., 47 388, 389, 390, 391, 392f, 422
Nathan, M. J., xv, 107, 360, 361, Posner, M. I., 126, 127
364, 366, 367f, 368 Post, R. B., 24
Neisser, U., 21, 22, 29, 420 Potter, M. C., 124, 125, 126
Nelson, K., 27 Poulsen, D., 68
Newell, A., 1,2, 6, 7, 29,42, 123, Premack, D., 68
216, 387, 392, 394, 396 Price, J. M., 297
Nitsch, K., 78, 135 Pylyshyn, Z. W., 6, 45
 Name Index 455

Quillian, M. R., 36, 75 Schallert, D. L., 71

Schank, R. C., 36, 37, 82, 94
Raajmakers, J. G., 85, 97 Schmalhofer, F., xv, 251, 252,
Ratcliff, R., 73, 78, 145, 146, 253f
148, 156, 174, 193,226 Schmidt, H. G., 233
Rayner, K., 129, 145, 148, 194 Schmidt, R. A., 320
Recht, D. R., 287 Schneider, N. J., 285
Reder, L. M., 252 Schneider, W., 287, 355
Redmiles, D., 107, 371 Schonpflug, W., 29
Rehder, R., 324, 326f, 327 Schreiner, M. E., xvi, 324, 326f,
Reusser, K., xv, 54, 338, 339, 327
340, 344, 345, 346, 347, Schvaneveldt, R. W., 36, 298
354, 359, 360 Schwanenflugel, RJ., 136, 137t,
Richman, H. B., 235 139, 142, 144
Riley, M. S., 334, 335, 335-6t, Sedgwick, C. H., 83
335,339,359 Seely, M. R., 132, 198, 232, 284,
Rips, L.J., 34, 192 285, 286
Risden, K., 198, 199 Selz, O., 94, 372
Ritchie, B. G., 244 Sengul, C. J., 148
Ritchot, K. F. M., 244 Shakespeare, W., 206
Rizzo, N. D., 238 Shank, D. M., 194,211
Roberts, J., 100, 139n, 254 Shepard, R. N., 44, 247
Rodenhausen, H., 99 Shiffrin, R. M., 35, 85, 97, 235,
Rosch, E., 35 249, 253, 254, 256, 257,
Roth, S. F., 148 261,267, 268, 273,274,
Rubin, D. C., 207, 208,212 275,276
Rumelhart, D. E., 35 Shinjo, M., 320
Ryan, L., 420 Shoben, E. J., 34
Sidner, C., 156
Sandburg, C., 159 Silfies, L. N., 288, 290, 323
Sanders, M. D., 24 Simon, H. A., 2, 3, 29, 31, 235,
Sandora, C., 329 366, 392, 394, 396, 397,
Sanford, A. J., 148, 156 398,412
Santa, J. L., 44 Simons, W. H. G., 180
Scardamalia, M., 329 Sinatra, G. M., 312
Scearce, K. A., 148 Singer, M., 193,216, 244, 251,
Schacter, D. L., 25 264
Schaecken, W., 192 Skinner, B. F., 20
456 Name Index

Smith, E. E., 34, 35 Truxaw, C., 358, 363

Smolensky, P., 169 Tulving, E., 250
Songer, N. B., 309f, 311, 313, Turner, T. J., 37, 82
315f, 316f, 318, 320, 323 Tversky, A., 399, 400, 414, 415,
Spence, K. W., 20 416
Squire, L. R., 25, 26 Tversky, B., 201
Stanners, R. F., 297 Tyler, L. K., 148
Stanovich, K. E., 283
Staszewski, J. J., 235 Underwood, B. J., 248
Staub, F. C., 347
Steedman, M., 172, 173 Valla, G., 127
Steinhart, D. J., xvi, 133, 134, van den Broek, P., xvi, 198, 199,
135 270, 285, 286, 320
Stern, E., 345, 346, 354, 355, van Dijk, T. A., 37n, 38, 39, 49,
356,357 50, 66, 67, 68, 69-70, 95,
Sternberg, S., 219 102, 118, 174, 184,216,
Stigler, J. W., 333 223, 234, 264, 265, 278,334
Streby, W. J., 221, 285 VanLehn, K., 386
Suchman, L. A., 23 Varma, S., xvi, 99, 235
Suh, S., 194, 198 Varner, K. R., 284
Swinney, D. A., 127, 128, 129, Varnhagen, C. K., 282-3, 285,
131 286
Verschaffel, L., 341
Tapiero, I., xv, 266 Vesonder, G. T., 297
Taraban, R., 170, 171 von Flumboldt, W., 32
Taylor, H. A., 201 Vonk, W., 180, 195-6, 197
Tegthsoonian, M., 248 Voss, J. F., 288, 290, 297, 323
Thagard, P., 394, 402, 403, 404, Vygotsky, L. S., 323
406, 408
Thompson, M. A., 289 Walker, C. EL, 251, 254, 287, 289
Thomson, D. M., 248, 250 Walker, W. H., 83, 274, 299
Thurlow, R., 198, 199 Wallace, W. T., 212
Till, R. E., 101, 131, 132, 133, Wanner, E., 70
134, 198, 228, 286 Warrington, E. K., 24
Tolman, E. C., 20 Weaver, C. A., xv, 420
Trabasso, T., 193, 194, 198, 270, Weimer, R., 54, 338, 339, 340,
285,286 344, 345, 346, 347, 354, 359
Trevarthen, C., 22 Weinert, F. E., 287, 355
 Name Index 457

Weisskranz, L., 24 Wolfe, M. B., xvi, 275, 276, 324,

Welsch, D. M., 78, 95, 251, 252f, 326f, 327
253f Worthy, J., 329
Wenger, E., 358, 363
Wertheimer, M., 290 Yarborough, J. J., 68
White, C. R., 136, 137t, 139, Yekovich, F. R., 251, 254, 289
142,144 Young, E., 107, 360, 361, 364,
Whitney, R, 37, 244 366, 367f
Wilhelm, G., 32
Wilson, B., 127 Zimny, S., xvi, 251, 252f, 253f
Wisniewski, E. J., 35 Zola, D., 283
Witte, R. S., 20 Zwann, R. A., 198
 Subject Index

abstract representation, 18 role in learning from text,

action planning 307-13
graphics application, 387-92 concepts, 74
routine computing, 372-81 consciousness, 418
UNIX, 381-7 construction-integration model
affordance, 23 story comprehension, 268-73
algebra word problems, 358-64 word arithmetic problems,
ANIMATE tutor, 364-9 347-54
anaphora, 144-57 contradictions in text, 93-103,
architecture of cognition, 5-7 181-4
arithmetic word problems cycles of processing in compre¬
errors, 339-45 hension, 101
schema, 337-8, 354-8
situation models, 347 decision making, 399-402,
types, 335-6 414-17
associative nets, 35 declarative memory, 25
direct and indirect representa¬
behaviorism, 20 tion, 20
buffer, memory discourse focus, 154
intermediate stores, 124-8
short-term memory in read¬ emergent structures, 82-6,
ing, 34-5 354-8
emotion, 22, 419-21
cognitive development, 27-9 encapsulated meaning, 42-4
coherence episodic memory
levels, 39 in text comprehension, 224-6
matrix, 98, 103 episodic representation, 17
460 Subject Index

errors macrostructures, 184-8

in arithmetic word problems, matching students and texts,
339-45 323-7
in skilled performance, 387—92 metaphor comprehension,
158-65
feature, 34, 41 retrieval structures, 235—8
frame, 36, 41 summaries, 277-80
free recall, 117 word meanings, 130-44
simulation of story recall, learning from text
268-77 active processing, 313-23,
of text, 264—80 329
advance organizer, 321-23
habit system, 24, 27 domain knowledge, 313-23
homographs, 128-36 matching students and texts,
323-7
imagery, 18, 30, 44-7, 108-11 measurement, 295-302
implicit memory, 25 vs. memory for text, 290—3
impression formation, 402-9 simulation, 293-5, 302—7
inferences, 97 levels of representation, 49
classification, 188-92 in sentence recognition,
in discourse, 193-97 251-3,258-64
in learning from text, 307, literary texts, 205-6
313-23 comprehension strategies,
time course, 198, 229 211-13
information processing theory,
6-7 macrostructure, 50, 64-9, 181-4
interest, 418-9 formation, 174-7
interrupting reading, 222-3, as LSA vectors, 184-8
230-1 signaling, 177-80
topic shift, 179-80
knowledge activation, 97 meaning
in text comprehension, of paragraphs, 185-7
227-34 of words, 77, 123-4, 130-6
knowledge net, 74 mental representation
embedding of levels, 30
latent semantic analysis, 86-91 redescription, 30
decision making, 399-402, role of culture, 31
414-17 role of environment, 31
essay grading, 325 types, 15, 16
 Subject Index 461

metaphor, 157-65 sentence recognition, 251-3

microstructure, 50, 54, 292 for text, 247-64
motivation, 413-18 reconstructive recall, 265,
288-90
narrative representation, 18 retrieval structure, 76, 224-34
novice and expert planning, modeled by LSA, 345-8
381-7 rhetorical schemas, 68

parsing of sentences in the schema, 36, 41, 82-6, 94

CJ-model, 168-74 in word problems, 337-8,
perceptual representation, 21 354-8, 360-4
predicate-argument schema, 37 script, 36, 41, 82-6, 111-18
priming, 73, 130-44 self
problem model, 107, 337-8, representation of, 409-13
339-45 semantic memory, 30
problem solving, 39/l—8 semantic nets, 36, 41, 75
procedural representation, 17 situated cognition, 23, 378
production rule, 42, 373 situation model, 49, 103-7,
proposition 108-18
atomic, 37, 273 in arithmetic word problems,
complex, 38, 273 345-7
psychological reality, 69-73 in literary texts, 209-11
propositional networks, 37 spatial situation models,
propositionalizing text, 54-60 199-204
propositional representations skilled memory, 217-21
of knowledge, 73-82 summaries, 118, 277-80
of text, 32, 33, 49-54 system model, 107, 371

reading skills textbase, 49, 103-7, 232, 290

decoding, 282-3
domain knowledge, 287-90, versification, 206-9
328
language skills, 283-7 working memory, 126, 215-21
reading span, 239 long-term working memory,
recognition memory 219-21
importance effects, 253-7 representation of self, 409-13
levels of representation,
258-64 zone of learnability, 323—7
- t
 MARYGROUE COLLEGE

3 ns? □□0LUM5B a

DATE DUE

DEMCO 38-296