BEHAVIORAL AND BRAIN SCIENCES (1993) 16, 217-265

Printed in the United States of America

"What" and "where" in spatial

language and spatial cognition
Barbara Landau
Department of Cognitive Science, University of California, Irvine, CA 92717
Electronic mall: blandau@orion.uci.edu

Ray Jackendoff
Linguistics Program, Brandeis University, Waltham, MA 02254
Electronic mail: jackendoff@brandeis.bitnet

Abstract: Fundamental to spatial knowledge in all species are the representations underlying object recognition, object search, and
navigation through space. But what sets humans apart from other species is our ability to express spatial experience through language.
This target article explores the language of objects and places, asking what geometric properties are preserved in the representations
underlying object nouns and spatial prepositions in English. Evidence from these two aspects of language suggests there are
significant differences in the geometric richness with which objects and places are encoded. When an object is named (i.e., with count
nouns), detailed geometric properties - principally the object's shape (axes, solid and hollow volumes, surfaces, and parts) - are
represented. In contrast, when an object plays the role of either "figure" (located object) or "ground" (reference object) in a locational
expression, only very coarse geometric object properties are represented, primarily the main axes. In addition, the spatial functions
encoded by spatial prepositions tend to be nonmetric and relatively coarse, for example, "containment," "contact," "relative
distance," and "relative direction." These properties are representative of other languages as well. The striking differences in the way
language encodes objects versus places lead us to suggest two explanations: First, there is a tendency for languages to level out
geometric detail from both object and place representations. Second, a nonlinguistic disparity between the representations of "what"
and "where" underlies how language represents objects and places. The language of objects and places converges with and enriches
our understanding of corresponding spatial representations.
Keywords: count nouns; identification; language; localization; object recognition; object structure; prepositions; spatial language;
spatial representation: what/where system

The representations underlying object recognition, ob- human language. Such analyses have been important not
ject search, and navigation through space are fundamen- only because of the richness and complexity of spatial
tal to spatial knowledge in all species. What sets humans language itself, but also because the organization of spatial
apart from other species is our ability to use these repre- language extends readily to many abstract domains such
sentations to express our spatial experience, talking about as time, status, possession, and social organization (Gru-
what things are and where they are located. Clearly, ber 1976; Jackendoff 1976; Lakoff & Turner 1980). The
language and spatial understanding map onto each other. burden of this target article is to show that the latter
In this target article we ask what that mapping might be - concerns can be brought to bear as a new source of
how language draws on our spatial representations such evidence on the nature of human spatial cognition. Spatial
that we can manage to talk about what we perceive. language, properly analyzed, can shed light on spatial
Our focus will be the language of objects and places in thinking.
English. Our specific goal will be to describe these We should begin by clarifying what we mean by spatial
linguistic domains in a way that is compatible with con- representation and what we mean by the language of
straints on nonlinguistic spatial cognition. At the same objects and places. By spatial representation, we intend a
time, however, we will use evidence from language to level of mental representation devoted to encoding the
provide boundary conditions on a satisfactory theory of geometric properties of objects in the world and the
spatial cognition: Our premise is that any aspect of space spatial relationships among them. Because spatial infor-
that can be expressed in language must also be present in mation can be derived from vision, audition, and the
nonlinguistic spatial representations. Simply put, what- haptic (touch) faculty, the format must be either hetero-
ever we can talk about we can also represent. l modal or amodal. That is, this representation is not
In the main tradition of research on spatial cognition, exclusively visual or haptic or aural, but spatial; under-
psychological and neuroscientific techniques have been standing spatial configurations in the world involves ei-
used to study our ability to visually perceive objects and ther translating modality-specific information into a com-
their locations and motions in space. At the same time, mon format or providing interfaces between modalities.
there has been a substantial tradition in linguistics and We also assume that spatial representations must be
psycholinguistics of studying the spatial expressions in translatable into a form of representation specific to the

© 1993 Cambridge University Press 0140-525X193 $5.00+.00 217

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language
Visual naming is particularly dramatic in the case of representa-
information
tional art. As one example, the 60-foot metal sculpture by
Claes Oldenburg that graces clowntown Philadelphia is
Linguistic Spatial Auditory
representations Representation information
universally recognized as and labeled "the clothespin,"
although it clearly violates most of the critical properties
Haptic
of true clothespins: Its shape is the dominating criterion
\
information in choosing its name.3
Aside from such informal observations, experimental
Motor
information
evidence shows that two-dimensional representations of
object shape can support object identification and nam-
Figure 1. Spatial representations take as input information ing. Adults can easily recognize and label familiar com-
from vision, audition, and the haptic system, and provide mon objects - both artifacts and natural kinds - on the
information to the motor system and language. basis of line drawings or silhouettes (Biederman 1987;
Roschetal. 1976). Such representations elicit recognition
immediately, automatically, and irresistibly.4 This recog-
motor system, used to initiate and guide behavior: We can nition ability is present in infancy (DeLoache et al. 1979)
touch what we see, look at what we hear, and avoid and may be innate: Even a child who has never learned
obstacles as we navigate through space2 (see Figure 1). name-picture correspondences can easily identify famil-
To account for language about space, there must be a iar common objects from simple black-and-white line
translation between spatial representations and language. drawings (Hochberg & Brooks 1962). (In contrast, imag-
For our purposes, the language of space will concern ine how difficult it would be to name most objects based
those words and simple phrases that encode objects and only on a patch of their color or texture.)
places. In English, objects are represented by count Object shape also seems to have a privileged status in
nouns and places are represented canonically by preposi- learning names for novel objects. A number of studies
tions or prepositional phrases (see below for some quali- have shown that young children rely specifically on shape
fications). Although people can obviously construct syn- when learning labels for novel objects (Au & Markman
tactically complex descriptions of objects (e.g., "the thing 1987; Bornstein 1985; Heibeck & Markman 1987; Landau
over there with the two pointed tops") or places (e.g., "not et al. 1988). For example, Landau et al. found that when
less than 3,000 miles away"), we will restrict our discus- young children or adults were shown a novel object
sion to count nouns and prepositions because these single labeled as a count noun (e.g., "This is a dax"), they tended
words are likely to correspond to the simplest well- to generalize that label to objects of the same shape as the
formed formulas in our conceptual system (Fodor 1983; original, even in the face of rather large differences in
Jackendoff 1983; 1990). texture or size. In addition, they found that this "shape
Our discussion will be in three parts. First, we will bias" becomes stronger over age, with 2-year-olds show-
discuss language pertaining to object representation - in ing a rather weak but stable preference, 3-year-olds show-
particular, that required for identification -and its bear- ing a stronger one, and adults showing the bias in an
ing on some current theories of the encoding of object extreme form, consistently rejecting even quite small
shape. Second, we will discuss language pertaining to deformations in object shape. Finally, they found that the
places, the locations of objects and the spatial relation- shape bias does not strictly mirror perceptual preferences
ships holding between different objects. We will highlight and can even occur in the context of highly salient com-
the differences between these two aspects of language, in peting properties (Jones et al. 1991; Landau et al. 1992;
particular, the fact that object naming draws on rich Smith et al. 1992). The shape bias appears most consis-
shape-based representations of objects, whereas place tently and strongly in the context of the word-learning
naming draws on quite sparse elements of object shape. task, suggesting a developmentally early link between
These differences will lead us to the third issue of why names for things and representations of object shape.
the identification and location of objects draws on such The importance of shape imposes a basic constraint on
disparate representations. We will consider two hypoth- the relation between spatial representations and lan-
eses, one focusing on the design of language and one guage: The spatial representations that are linked to
focusing on the design of spatial representations. We will object names must provide enough different shape de-
conclude with some predictions derived from these scriptions, configured in the proper way, to be able to
hypotheses. distinguish all the kinds of objects we categorize (or
partially categorize) linguistically on the basis of shape.
We propose that just as the number of possible shapes
1. Talking about objects must meet or exceed the number of object names in a
language (Biederman 1987), any spatial distinctions we
In the average adult vocabulary, there are roughly 10,000 can encode linguistically must be capable of correspond-
names for things - count nouns that label different kinds ing to spatial representations. Such a correspondence is
of objects. For a large proportion of object categories, necessary if talking about objects and places is linked to
shape is among the most important criteria for identifica- thinking about or acting on them. This section begins by
tion, and in particular for judgments of what a thing reviewing some current approaches to the spatial repre-
should be called: Categories of things with the same sentation of object shape. We then proceed to show that
shape, including natural kind objects and artifacts, often linguistic evidence motivates interesting augmentations
share the same name. The importance of shape in object of the theory of spatial descriptions.

218 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language

1.1. Whole objects figure and various animals, whereas Biederman describes
such objects as airplanes and cameras (see Figure 2a and
From the nonlinguistic side, a traditional approach to the 2b). In principle, these systems should be able to gener-
problem of shape description takes object shapes to be rep- ate enough descriptions to cover all named or unnamed
resented componentially by simple three-dimensional objects. With a suitable similarity metric, they could
components such as cylinders. This approach dates back account for differences among named categories such as
at least to Leonardo da Vinci, who described the human person versus gorilla, whose shapes are quite similar.6
figure as a combination of units fitting a particular set of Furthermore, using these componential systems, one can
proportions. Twentieth-century work in computational characterize transformations that preserve the geometric
vision and psychology has also drawn on componential structure of objects capable of internal movement, cap-
analyses to describe object shape (Biederman 1987; Bin- turing for example the possible limb movements of a
ford 1971; Lowe 1985; Marr 1982). That is, a limited human figure, as described by Marr and Vaina (1982) (see
number of shape components are taken to be the units Figure 2c). Finally, the componential systems are capable
used by perceivers in recognizing object shapes. of characterizing many kinds of object parts, an issue to
What corresponds most closely to our notion of spatial which we turn next.
representation is dubbed by Marr the "3-D model level,"
in which objects are encoded in an object-centered for-
1.2. Names for object parts
mat, independent of the viewer's perspective. Even
though this approach has been developed primarily for A significant part of the lexicon includes names for coher-
vision, it should be capable of accepting inputs from ent object parts: handles, noses, legs, stems, and so on.
haptic/kinesthetic sources as well, making it non- For some parts, the componential approach provides
modality-specific. Such an interpretation of the compo- straightforward ways of parsing the host object: Parts are
nential approach would be compatible with the extensive represented by the individual generalized cones that
evidence that object shape is encoded in some detail by combine to yield the whole object. Thus, some of the
the congenitally blind (see, e.g., Landau 1991; Lederman named parts of a camera, an airplane, or a person could be
& Klatzky 1987). Moreover, Jackendoff(1987a; 1987b) has described by particular cones. Using Marr and Vaina's
shown that 3-D model representations can be translated scheme, one can add that certain body parts - the head,
in part into representations suitable for linguistic expres- legs, or arms of a person - follow from parsing at critical
sion. Hence, the 3-D model level, insofar as it can be movement joints. The regions of the joints themselves
fleshed out, has the properties with which we have char- correspond in turn to names for parts such as elbows,
acterized spatial representation. knees, and wrists.
In most componential approaches, the primitives for For other parts, the componential approach seems less
3-D object description are related to Binford's (1971) appropriate: Many named parts, especially parts of natu-
notion of the generalized cylinder. In Marr's model, for ral objects, are not well described by cylinders. For
example, object representations are built up from (1) a set example, how does one tell where a forehead or a nose
of principles for describing "generalized cones" in terms begins and where it ends? In such cases, two different
of an axis and a varying cross-section, and (2) a principle approaches have been suggested. Hoffman and Richards
for elaborating a generalized cone by adding a subsidiary (1984) suggest that parts may be defined by certain
generalized cone whose axis is of a particular size and characteristics of the object's boundaries rather than by a
orientation relative to the main axis. Principle (2) applies set of prespecified primitive shapes. They propose that a
recursively, so that objects in the 3-D model representa- part is perceptually defined as the segment between two
tion are composed of parts, each of which may have a consecutive contours of negative minima within the
further decomposition. As a result, highly detailed shapes boundary's principal curvatures. The nose, for example,
can emerge, making the representations in principle rich would begin and end at points of negative extrema along
enough to support the extensive vocabulary of object the external boundary of the face (see Figure 3a).
names seen in language. A second approach, offered by Leyton (1989), suggests
As another example, Biederman (1987) proposes that that parts may be those units produced by the causal
the parts can be encoded in terms of a small specific set of processes underlying shape formation. For example, pro-
generalized cones, 36 in number, which he calls "geons.' cesses such as "squashing," "protrusion," or "indentation'
These geons are meant to have "nonaccidental" properties can operate on an initial shape to yield quite a different
(Lowe 1985) - properties in the image that are likely to shape (which can nevertheless be seen as related to the
represent true properties of the object rather than acci- original by the relevant causal processes). The nose, for
dents of viewpoint. For example, the geons are meant to example, could be the result of a growth process wherein
be contour-determined and invariant over size and view- a segment of a simple curved surface is pushed out from
point transformations.5 Given a small set of attachment the inside along a principal axis to form a new segment
relationships among the cones - such as "end-to-end" or (Figure 3b).
"top-to-side" - and only a few iterations, Biederman's Each approach is promising in that it affords us a way of
system can generate at least as many object shapes as linking descriptions of further object parts with their
there are object names (as would be expected from any labels. Of course, this linking is not one-to-one: We can
componential system of sufficient complexity).
often find parsable parts of objects that go unnamed.
The componential systems are all capable of generating Even with these enrichments, however, there are a
a wide range of particular object shapes from primitives. number of cases in which evidence from language sug-
For example, Marr describes schemata for the human gests the need for aspects of shape descriptions that are

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 219

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language
Human

MODEL AXIS AT THE BEGINNING

HUMAN MODEL AXIS AT THE END

OBJECT MODEL AXIS

IN HUMAN FRAME

Figure 2. Componential theories of object shape can provide detailed descriptions of complex shapes based on the combination of a
few primitives. A: Marr's (1982) schemata for the human figure. B: after Biederman's (1987) camera. C: Marr and Vaina's (1982)
schemata for the human figure under movement transformations. (Figure 2A reprinted from Marr [1982] and 2C from Marr & Vaina
[1982].)

not a prominent part of these proposals about object be crucial in nonlinguistic tasks such as mental rotation,
representation. The rest of this section makes some ele- where judgments of object identity may require discrimi-
mentary observations that point to a number of such gaps. nating right-left reversals of an object (Parsons 1987;
In each case we will suggest enrichments of spatial repre- Shepard & Cooper 1982; see Tarr & Pinker 1989 for
sentation that seem compatible with the approaches de- emphasis on orientation-dependence). A principal ori-
scribed so far. ented object axis appears to be critical for our representa-
tions of objects in the "implicit" memory system (Cooper
1992).
7.3. Names for spatial parts: Axes and axial parts
Both the linguistic and the nonlinguistic facts therefore
Many objects can be described as having a top and a suggest the need for explicit representation of oriented
bottom, a front and a back, and sides and/or ends. These and directed axes in object descriptions, that is, the axes
terms do not describe parts in the same sense as, say, required to distinguish top from bottom, back from front,
handle or wing: They are not subsidiary parts tacked onto and right from left. The componential approaches we
the object. Rather, they denote regions of the object have described provide for representing local axes (the
based on its inherent orientation (as opposed to contex- axis of each part) and overall object axes (such as the
tually imposed uses; see sect. 2.7.2). How can these terms principal axis of a canonically oriented object; see Marr &
be derived from a spatial representation? Nishihara 1978); but as far as we can tell, they make no
Leaving language aside for a moment, notice that the provision for representing the orientation of these axes.
orientation of an object is necessary for describing certain In other domains, however, such as morphological de-
relationships among parts: for example, the fact that one's velopment (Thompson 1961) or perceptual analysis of
nose, feet, and navel point in the same direction (hence changes in shape (Leyton 1992), oriented axes are critical
are on the front of the body) or that one's arms are to understanding the structure and transformations of
attached opposite one another and orthogonal to the front natural shapes.
(hence are on the sides). Relative orientation of parts can One way to express regularities of orientation is to

220 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language

Second, an axis can be marked optionally as intrin-

sically directed or symmetric (see Figure 4). A directed
axis indicates inherent regularities that distinguish one
end from the other: This can distinguish top from bottom
or front from back. For example, the directed front-to-
back axis of the human figure establishes the regularity of
alignment for the nose, feet, and navel. A symmetric axis
indicates equivalent elaborations of the object at both
ends of the axis. For example, the side-to-side axis of the
human figure establishes the symmetry of the limbs and

O
face parts.
The generating axis of a cone as well as its orienting axes
can be directed or symmetric. In the human figure, for
example, the main generating axis is a directed one that
distinguishes top from bottom. In Biederman's camera, if
the long side-to-side dimension is the generating axis for
Figure 3. Theories of object parsing can provide descriptions the principal geon, then that axis will also be symmetric
of coherent object parts named by nouns such as nose and (since right/left need not be distinguished in a camera,
handle. A: Hoffman and Richards (1984) define a part as the aside from minor details).
segment between two consecutive contours of negative minima Other combinations of generating and orienting axes
within the boundary's principal curvatures. B: Leyton's (1989)
theory suggests that parts may be defined as those units pro- are also possible. For example, an arrow has a directed
duced by causal processes underlying shape formation. Here, a generating axis but no significant orienting axes. The
process of protrusion creates a "nose' within a uniformly curved human hand has an oriented generating axis (wrist-to-
boundary. (Figure 3A reprinted from Hoffman & Richards fingers, following Marr), and two directed orienting axes
[1984].) (back-to-palm and pinky-to-thumb).
Returning to the linguistic description of objects, we
can use the system of directed and orienting axes to define
extend the theory of axes in two ways. First, we will call the terms brought up at the beginning of this section,
the axis that is expanded into a generalized cone the which we can call "axially determined parts." The top and
generating axis; this is the axis that is central to Marr's and bottom of an object are the regions (or parts of the surface)
Biederman's accounts of object shape. Let us impose on of the object at the ends of whichever axis is vertical in the
this cone up to two further axes, called orienting axes, object's normal orientation. If the object is relatively long
which are orthogonal to the generating axis and to each and narrow, that is, if it has a horizontal generating axis
other. These axes will serve to orient the principal cone
significantly longer than the other axes, it can be said to
radially (see Figure 4). In the case of the human body, the
have ends — the regions at the termination of this axis. If
principal generating axis is vertical; it defines the gener-
the object has a horizontal directed axis, with one that
alized cone of the torso. The two orienting axes determine
normally faces the observer or determines the normal
the front-to-back and side-to-side directions.
direction of motion, the region determined by that end of
the axis is the object's front; the opposite end of this axis
determines the back. Finally, the region determined by
the termination of any other horizontal axis can be called a
side. 7 Thus, linguistic and nonlinguistic facts about shape
converge in motivating use of the axial system as an
~ \ important part of object representation.

\ \ 1.4. Names for objects best described as surfaces

Consider the spatial representations of sheets of paper,
phonograph records, crackers, table tops, blackboards,
rugs, roads, and lakes. What these have in common is that
they are principally extended in two dimensions, with a
relatively negligible thickness (at least in the relevant
context); the linear boundary of this surface can then be
defined as its edge.
How are such objects to be encoded in spatial represen-
Generating axis + Orienting axes + Directed axes tation? It seems wrong to treat a phonograph record as a
very fat cylinder with a very short main axis passing
Figure 4. Three types of axes are required to account for through the hole. However, this is the only way to
linguistic terms describing aspects of an object's orientation. generate it formally in the componential framework of
The generating axis is the object's principal axis as described by
volumetric primitives. Furthermore, a lake hardly lends
Marr (1982). In the case of a human, this axis is vertical. The
orienting axes are secondary and orthogonal to the generating itself at all to such a description. For example, if its
axis and to each other (e.g., corresponding to the front/back and generating axis is taken as going from the surface to the
side/side axes). The directed axes differentiate between the two bottom, the description is entirely counterintuitive, es-
ends of each axis, marking top vs. bottom or front vs. back. pecially if the lake is irregular in shape. Alternatively, its

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 221

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language

generating axis might be taken as parallel to the surface only be enlarged in two dimensions: A big square is large
(say if the lake is relatively long and narrow). Our own in each of its two dimensions, a big cube is large in each of
intuition, however, is that it would be odd to have it fall in its three dimensions (see Figure 5). Surface-type objects,
the interior of the geon (as in Biederman's [1987] reper- like record or lake, behave like square in this analysis: A
toire): That would put the axis under water. big record is large in two dimensions, but if significantly
The problem in these cases is that models using vol- enlarged in its third dimension, it ceases to be seen as a
umetric primitives require one to generate a volume record (see Figure 5). Similarly, a big lake is one that is
directly from a linear axis. A more intuitively satisfying extended in length and width; its depth is irrelevant.
analysis of these objects is that they are schematized as Other dimensional adjectives have related application.
surfaces, possibly elaborated into a volume by adding a For example, the adjectives thick and thin can be seen to
thickness. In this analysis, the record is schematized as a place a metric on the elaboration of a surface into a
disk rather than a volume; the lake is schematized more as volume: A thick record is one that is relatively large in the
its surface, with depth as an elaboration. The edge of such third dimension.
objects is the linear boundary of the schematized surface.
In addition to object names for such "surface-like" 1.5. Names for "negative" object parts
objects and for the term edge, there are other words that
Some entities may be best conceptualized as "negative
benefit from having such analysis available in spatial
parts" of objects, as alluded to by Hoffman and Richards
representation. Two classes come to mind: (1) two-
(1984) and Herskovits (1986). Compare a ridge and a
dimensional shape terms like square, circle, oval, trape-
groove. A ridge is conceptualized as a protrusion from
zoid, and so on; (2) general terms for "thickened surfaces"
the surface of a host object. It has an extended linear
such as slab, sheet, layer, slice, lamina, and stratum.
To further motivate the distinction between "surface- generating axis parallel to the surface of the host object. In
type" and "volume-type" objects, consider what happens addition, it has a directed orienting axis that projects out
when these different types of objects are modified by a of the surface of the host object, giving the ridge a top and
dimensional term such as big. Like many adjectives, big a bottom, and a (roughly) symmetrical orienting axis that
selects different dimensions, depending on the nature of defines its sides. It can therefore be easily described as a
the object. If an object is inherently surface-like, it can part of the host object using the volumetric system as
elaborated so far.
Now consider a groove. It is conceptualized as a depres-
sion in the surface of a host object. It has an extended
linear generating axis parallel to the surface of the object
plus a directed orienting axis that projects into the surface
of the host object, giving the groove a top and a bottom,
plus a (roughly) symmetrical axis that defines its sides.
Although one could describe it within the volumetric
A square A big square - A big square
system as an indented volume, we suggest an alternative
representation.
A natural way to think of a groove is as a "negative part,"
a shaped volume scooped out of the object instead of
added to it. That is, it is a shape defined by "lack of
substance" rather than by presence of substance, as in the
case of normal parts. Other than that, a negative part
evidently has shape descriptors - and a linguistic descrip-
tion - essentially parallel to those of ordinary object parts.
A groove not only has a top, bottom, and sides; it can be
A cube A big cube •A big cube
described as long or short (along its principal axis), deep or
shallow (along its secondary axis), and broad or narrow
(along its tertiary axis). Notice further that the terms deep
and shallow play the same role for negative parts that high
and low do for ordinary parts: We speak of a deep groove
rather than a *high groove, for instance.
Other negative part names are hole, pit (a "negative"
bump), notch, slot, scratch, depression, cavity, and possi-
A record A big record "A big record
bly dent. Words that name "negative objects" are valley,
ditch, cave, and well, as well as door and window (in the
Figure 5. The distinction between "surface-type" and sense of "opening in a wall" rather than the object used to
"volume-type" objects can be seen when they are modified by a close off such an opening). Thus, a simple enrichment of
dimensional term such as big, which has the function of enlarg-
spatial representation again affords revealing analyses for
ing the object in each of its critical dimensions. A square is
inherently two-dimensional, so a "big square" can only be en- a wide variety of things we can name (though strictly
larged in those two dimensions, but not in the third dimension speaking they are not objects this time).
(in which case it becomes a cube). A cube is three-dimensional,
so it must be enlarged equally in its three dimensions. In this 7.6. Names for containers and related objects
analysis, a record behaves like the two-dimensional object: A
"big record" must be large in its two primary dimensions, but As Miller and Johnson-Laird (1976) point out, English has
not in the third. an extensive set of labels for containers: objects like cups,

222 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language

bowls, boxes, jars, tanks, and so forth. What is their where the object is or, if in motion, its path of movement:
spatial representation? One possibility is that a cup, for That is an essential part of spatial cognition - and an
instance, is a volume - a cylinder out of which a large essential part of spatial language.
coaxial negative cylinder has been scooped. An alterna- There is a large and diverse literature on spatial repre-
tive with a certain intuitive appeal is that a cup should be sentation in humans and other species. In addition to the
represented as a thickened surface that encloses a cylin- vast body of research on space perception, the variety of
drical space - that is, the sides and bottom of the cup are perspectives includes that of psychologists interested in
not the residue of extensive scooping, they are surfaces. how perceptual-motor coordination is achieved (Hein &
Although these two alternative descriptions require an Jeannerod 1983); how spatial knowledge is structured
empirical test, there is some evidence that language (Gallistel 1990), how it develops (Stiles-David et al. 1988)
distinguishes containers from solid objects; roughly, con- and how it is represented neurally (O'Keefe & Nadel
tainers are objects that can hold things inside them (see 1978). There is also a considerable tradition in city plan-
sect. 2.4.1). To encode a class of containers in spatial ning and environmental psychology that seeks to under-
representation, we tentatively adopt the second alterna- stand how we represent large spatial layouts such as the
tive and introduce the notion of "hollow" volumes, distin- cities in which we live (Coucelis et al. 1987; Downs & Stea
guished from the standard "solid" volumes (see Marr 1973; Hooper 1978; Kuipers 1978; Lynch 1960).
1982, for a similar suggestion). Solid geons would be The variety of perspectives makes it difficult to identify
encoded as uniformly substantial; hollow geons would be a single unified theory of spatial representation that we
shapes with a substance that is distributed only over their might use as a model for thinking about how language
surfaces, leaving a shaped empty space inside. encodes spatial relations. In taking these approaches as a
In addition to the containers mentioned above, hollow whole, however, one theme is pervasive: Understanding
volumes would permit the theory of spatial representa- our representations of space requires invoking mental
tion to neatly encode such objects as cars and other closed elements corresponding to places and paths, where
vehicles, houses and other buildings, stomachs, egg- places are generally understood as regions often occupied
shells, balloons, bubbles, violins, and drums. Again, a by landmarks or reference objects. Objects (including
simple parameter added to spatial representation affords oneself) are then located in these places. Paths are the
an intuitively natural treatment of a significant new class routes along which one travels to get from place to place.
of objects. These elements are likely to be critical in any complete
theory of spatial representation and will serve as a skeletal
organization for our discussion of the language of space.
1.7. Summary We draw especially on the work of Bennett (1975), Clark
To name objects and object parts, spatial representation of (1973), Fillmore (1975), Hawkins (1984), Herskovits
objects by shape must be a rich combinatorial system. Its (1986), Jackendoff (1983; 1990), Miller and Johnson-Laird
basic units include not only generalized cones, but also (1976), and Talmy (1978; 1983). As with our discussion of
surfaces. Cones may be marked as "solid,' "hollow," or objects, we will show that linguistic evidence both con-
"negative." Each unit has an axial structure: the generat- verges with and enriches findings from nonlinguistic
ing axis around which the cone or surface is elaborated, studies of spatial cognition.
plus up to two orthogonal orienting axes. In turn, each of
the three axes may be directed or symmetric. These basic 2.1. Basic elements: Figure, reference object, region
units are combined hierarchically to form complex object
descriptions. 8 The standard linguistic representation of an object's place
In addition to our proposed additions to the descriptive requires three elements: the object to be located (or
power of the componential framework, there remains the figure), the reference object (called ground by Talmy9),
task of showing that the whole system of decomposition and their relationship. In the canonical English expres-
into parts can be adapted to the tolerances necessary for sion of an object's location, the figure and reference
object category discrimination. For example, descrip- objects are encoded as noun phrases; the relationship is
tions must be potentially fine-grained enough that one encoded as a spatial preposition that, with the reference
can decide which objects are to be named horse and object, defines a region in which the figure object is
which donkey, or which dog and which wolf. On the other located. For example, in the sentence, "The cat is sitting
hand, they must be potentially indeterminate enough to on the mat," the figure (the cat) is located in the region
allow considerable variation in shape within these named described by the prepositional phrase on the mat. The
categories, for example, the differences between Dalma- region is in turn described by the reference object (the
tians, German Shepherds, and Pekingese dogs, and to mat) and the spatial relation expressed by the preposition
allow such variation as occurs in the number, placement, on, roughly, "contact with the surface of the reference
size, and shape of arms on a saguaro cactus. How these object." In addition to prepositions, there are many verbs
tolerances for discrimination are to be formalized is be- that incorporate spatial relations; these can (almost invari-
yond the scope of this target article (see Note 6). ably) be paraphrased by a simpler verb plus a preposition.
For example, enter can be paraphrased by go into, ap-
proach by go toward, and cross by go across. (See Jacken-
2. Talking about places doff 1983; 1990 for formalization of these relations.)
Thus, the key element in the English expression of
The componential framework and all the amplifications place is the preposition: We can develop a fairly compre-
we have discussed concern object recognition and catego- hensive idea of the spatial relations expressed in language
rization, that is, what an object is. None of this addresses by focusing on spatial prepositions. We present a fairly

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 223

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language

Table 1. Prepositions of English many are not object names, so this estimate is only rough.
But even supposing the estimate is drastically biased,
about between outside there is a difference of approximately two orders of magni-
above betwixt over tude: For every spatial relation expressible in English,
across beyond past there are perhaps a hundred object names. This qualita-
after by through tive difference is reproduced in every language we know
against down throughout of. (If there were a language with even a thousand prepo-
along from to sitions, someone would certainly have raised a big hue
alongside in toward and cry about it.)
amid(st) inside under Given the small number of prepositions, the word class
among(st) into underneath most clearly devoted to expressing spatial relations, one is
around near up led to ask what constrains the range of possibilities so
at nearby upon severely. Our hypothesis is that there are so few preposi-
atop off via tions because the class of spatial relations available to be
behind on with expressed in language - the notions prepositions can
below onto within mean - is extremely limited. This section presents what
beneath opposite without we believe is a rather comprehensive enumeration of the
beside out factors involved in defining the spatial relations expressed
Compounds in English: the totality of meanings of spatial prepositions.
far from on top of These factors divide into four classes. The first, the
in back of to the left of asymmetry between figure and reference objects, sets
in between to the right of the basic parameters for spatial relations (sect. 2.2). The
in front of to the side of remaining three concern geometric possibilities for the
in line with three key elements of the spatial relation: the reference
object (sect. 2.4), the figure object (sect. 2.5), and the
Intransitive prepositions region based on the reference object (sect. 2.7).
afterward(s) forward right
apart here sideways
away inward south 2.2. Asymmetry between figure and reference object
back left there It is logically possible that spatial relations could be
backward N-ward (e.g., together mentally encoded as binary relations between objects,
downstairs homeward) upstairs that is, as propositional functions of the form R(a,b),
downward north upward where a and b are the objects to be related. In human
east outward west languages, however, a predominant way to express spatial
Nonspatial prepositions relations is asymmetrical. In the canonical form, the
ago for figure is encoded as grammatical subject, and the refer-
as like ence object is encoded as the object of the spatial preposi-
because of of tion or of the verb itself.
before since To illustrate, we have annotated the figure and refer-
despite until ence objects in the sentences in (1).
during (1) a. The book (figure) is lying on the table (reference object).
b. The train (figure) reached the station (reference object).
c. The star (figure) is inside the circle (reference object).
d. The circle (figure) lies around (surrounds) the star (refer-
complete list of the prepositional repertoire of English in ence object).
Table I. 1 0 Additional prepositions exist, but these are
Note that (lc) and (Id) can describe the very same
typically either archaic or reserved for technical usage
physical stimulus (see Figure 6A). They organize it differ-
(e.g., betwixt, athwart, abaft, etc.) and, in any case, these
ently, however, exchanging figure and reference object.
prepositions do not violate the principles set forth below.
These different organizations appear to reflect differences
A salient fact about prepositions is that there seems to
in the encoding of the stimulus in spatial representation,
be surprisingly few of them in comparison to the number
with primary attention switching from one object to the
of names for different kinds of objects. (In fact, there are
other.
few enough prepositions that they are usually considered
There appears to be a canonical way of expressing
part of the "closed-class" vocabulary, along with auxilia-
linguistically the assessment of object to the roles of figure
ries, determiners, and inflections.) We can get an idea of
and reference object. Whereas the exchanges listed above
the order of magnitude of different spatial relations ex-
are not unusual, they are not always possible. As noted by
pressed in English by counting the prepositions (see
Miller and Johnson-Laird (1976) and Talmy (1983), if the
Table 1). There is something on the order of 80 to 100,
objects are unequal in size or mobility, the larger and
depending on how one counts. Many of these are poly-
more stable is invariably encoded as the reference object.
semous (e.g., the different senses of over), and quite a few
For example, in (2), an exchange comparable to (lc,d)
are nonspatial (during, for instance, is purely temporal),
produces the odd-sounding result (2b).
so this estimate gives us only a ballpark figure. But
compare it to the number of count nouns in English - tens (2) a. The book is on the table,
of thousands. Again, many of these are polysemous, and b. ?The table is under the book.

224 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language

cases, the object that surrounds the other or tends to

dominate it by greater size or intensity will tend to
become the reference object (Oppenheimer 1934).
A similar asymmetry appears to guide our assignment
of objects to the role of figure versus ground when making
explicit judgments of the relative distance between ob-
(a) The star is inside the circle.
jects in spatial layouts. Sadalla et al. (1980) asked people to
(b) The circle lies around (surrounds) the star. judge distances between different campus landmarks that
were thought to be either good or poor reference objects.
Good reference objects had been shown by independent
ratings to be relatively large, stable, familiar, or culturally
significant. People tended to judge the distance of a poor
reference object (e.g., the architecture building) from a
good reference object (e.g., the student union) as shorter
than the reverse, suggesting that reference objects are
critical anchors in structuring cognitive maps (see also
Coucelis et al. 1987). Thus, converging with the linguistic
evidence presented above, a logically symmetrical rela-
(a) The bike is near the garage.
tionship - distance - is treated as a psychologically
asymmetrical relationship when people make certain
(b) ? The garage is near the bike. kinds of distance judgments.
Evidence from language tasks also confirms the linguis-
Figure 6. A: The same scene can be described in two different tic reflection of figure-reference object asymmetry, even
ways, by reversing the figure and ground objects. B: Even an
apparently symmetrical spatial relation - "nearness" - is subject in young children. A classic study by Huttenlocher and
to asymmetry offigureand ground, as is shown by the relative Strauss (1968) showed that children and adults respond
naturalness of the two sentences. more quickly to sentences in which the mobile object is
named as grammatical subject of the sentence and the
stable object is named as grammatical object of a preposi-
Even what would seem to be a symmetrical spatial tion than they do to sentences in which these roles are
relation - adjacency - is subject to the asymmetry of reversed. For example, subjects were quicker and more
figure-reference object dyads, as shown in (3) and Figure accurate when asked to "Make it so the (mobile) block is
6b. on top of the (fixed) block" than when asked to "Make it so
the (fixed) block is on top of the (mobile) block" (i.e., by
(3) a. The bicycle is next to the house, placing the mobile block under the fixed block).
b. ?The house is next to the bicycle. This effect even appears for the symmetrical predicate
This is not to say that spatial expressions must be near. Landau et al. (forthcoming) showed children and
asymmetrical (consider, e.g., "the bicycle and the house adults pairs of objects and asked them to "Make it so the
are adjacent to each other"; see Landau & Gleitman 1985, (mobile object) is near the (fixed object)" or the reverse.
for discussion), but only that this is the norm. Reaction times to the canonical form were much quicker,
What causes these asymmetries? It does not seem to and some adults even claimed they could not carry out the
follow from any fact specifically pertaining to language request in its noncanonical form (e.g., "Make it so the
that, in these contexts, the table and the house are more house is near the bicycle"), because the fixed object could
plausible reference objects and the book and bicycle are not be moved.
more plausible figures. (In particular, if "the house" in Finally, this sensitivity to asymmetry appears in ex-
(3b) happens to refer to a toy house, which is smaller and plicit judgment tasks. Landau etal. (forthcoming) showed
more mobile, the sentence becomes much more accept- children (2-, 4-, and 6-year-olds) and adults drawings of
able.) Rather, we believe that this linguistic asymmetry pairs of objects, and asked them to make judgments of
follows from principles of spatial organization, which asymmetry. In each pair of objects was one that could
require that an object be anchored (or located) relative to naturally serve as a reference object for the other because
some other object. Reference objects should have proper- it was larger or more stable - for example, a picture of a
ties that facilitate search: In many contexts, they should house (reference object) adjacent to a bicycle (figure).
be large, stable, and distinctive (and in environmental Subjects were asked, "Which one is near which one?" or,
contexts they are often landmarks; Lynch 1960). That is, in some cases, "Is the house near the bicycle, or is the
in this case, the organization of language parallels the bicycle near the house?" Both of these requests pushed
organization of spatial cognition.11 subjects to make an asymmetry judgment for a predicate
A variety of experimental evidence supports the notion (near) that is logically symmetrical. Even the youngest
that these asymmetries are fundamental to our spatial children tended to place the smaller and more mobile
representations - in particular, that people expect figure object in subject position and the larger and more stable
and reference objects to differ in size and stability. For object in prepositional object position, saying, for exam-
example, in perception, some cases of "induced move- ple, "The bicycle is near the house" rather than "The
ment" show the importance of differentially marking house is near the bicycle." A similar finding was obtained
figure and reference frame: If a stationary dot is placed by Sadalla et al. (1980): When people were asked to assign
within a moving rectangular frame, observers see the dot two well-known campus buildings to the sentence "
as moving rather than the frame (Duncker 1929). In such is close to ," they tended to assign the clear refer-

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 225

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language

ence objects to the prepositional object position, again For a third case, it is easy to think of many dozens of
consistent with linguistic observations on asymmetry. container terms of the sort discussed in section 1.6, for
In sum, asymmetry between figure and reference ob- example, cup, bowl, vase, box, carton, crate, coffin, bag,
ject emerges in a variety of tasks, ranging from strictly pouch, sack, sheath, case, tank, barrel, and so on. Yet
perceptual judgments to cognitive judgments of distance there are essentially only two spatial relations that pertain
and linguistic judgments of semantic naturalness. 12 This to containers: Something can be either in a container or
suggests that the asymmetry arises as a consequence of out of it (but see discussion of verbs in sect. 2.6). For
our nonlinguistic representations of space and that lan- example, there is no preposition plin that describes con-
guage draws on them as a means of expressing a critical tact with the inner surface of a container, so that one can
aspect of these representations. say:
Bill crawled plin the water tank, in the sense
2.3. The sparse constraints on the geometry of figure "Bill crawled along the inner surface of the tank"; or
and reference object Bill spread paint plin the carton, in the sense
"Bill spread paint on the inner surface of the carton."
Next we explore in more detail the components of spatial
relations, beginning with how the system of spatial rela- The inner surface must be explicitly mentioned. Like
tions expressed in language makes use of the shapes of the sprough and betwaft, this hypothetical term represents a
objects being related. Having just discussed the intricate perfectly plausible spatial relation, but a perfectly horri-
shape descriptions required for object naming, it is worth ble preposition.
asking to what extent the same descriptors are used when About the most complicated cases we have found in
describing an object's place. What proves surprising is English - in which some elements of object shape are
how sparsely both the figure and reference objects appear relevant to the preposition's meaning - are the terms
to be represented. along and across. Along requires its reference object to
Taking a simple case, there seem to be no prepositions have a principal axis of significant elongation, so one can
with a figure or reference object that must be analyzed in travel along a road or along a beach, but not along a chair
terms of a particular geon. A hypothetical example would or along a round table. One can travel along the edge of a
be the preposition sprough, "reaching from end to end of round table, but then the linear edge, not the table as a
a cigar-shaped object," appearing in sentences like (4a) whole, is serving as reference object. In addition, this
but not (4b) (see Figure 7). principal linear axis must be (more or less) horizontal: A
bug can be said to crawl along a flagpole only if the
(4) a. The rug extended sprough the airplane.
flagpole is lying down. (We treat across and some further
The weevil bored sprough the cigar.
wrinkles in along shortly.)
The major axis of an ellipse goes right sprough it.
And that is more or less it with regard to specific shape
b. *The rug extended sprough my dining room.
requirements. The detailed descriptions of shape rele-
*The weevil bored sprough the chair.
vant to the naming of objects appear to be irrelevant to the
*The major axis of a cup handle goes right sprough it.
descriptions of the same objects in their role as figures or
Similarly, there are no prepositions that insist on anal- reference objects. Only very sparse schematization of the
ysis of the figure or reference object into its constituent objects is relevant.
parts. An example might be the hypothetical preposition Landau and Stecker (1990) produced experimental
betwaft in (5), which requires the reference object io have evidence for this bifurcation between the description of
a protruding part. objects as objects and the description of objects in their
(5) a. The bug crawled betwaft my face. role in constructing spatial relations. They showed 3-year-
"The bug crawled down the junction between my nose olds, 5-year-olds, and adults a single scene in which a
and the main body of my face." novel object was being placed on the top front corner of a
box. Half of the subjects were told, "This is a corp," as if
b. The water ran betwaft the airplane.
the novel word named the object itself. The other half
'The water ran down the junction between the wing and
were told, "This is acorp my box," with the same novel
the fuselage."
phonetic sequence now playing the role of a preposition,
c. A stripe extended betwaft the cup.
as if it described the object's location. All subjects were
"A stripe extended along/down the junction between the
then shown a series of novel objects being placed one at a
body of the cup and the handle."
time in novel locations on a second identical box. Each
time, they were asked either, "Is this a corp?" or, "Is this
acorp my box?" with the choice of question matching the
(b) syntactic context of the introductory sentence they had
heard.
Subjects who had heard the sentence with a novel
count noun generalized it to objects of the same shape as
(a) The weevil bored "sprough" the oigar. the original, completely ignoring the object's location on
the box. In contrast, subjects who had heard the sentence
(b) *The weevil bored "sprough" the pillow.
with a novel preposition generalized it to objects of any
shape, as long as they were in roughly the same location as
Figure 7. There is no preposition in English whose reference
object must be analyzed as a particular geon, such as one that the original. This suggests that both children and adults
would require a cigar-shaped reference object, making only (a) attended to details of an object's shape when it was named
grammatical. as an object, but completely ignored the same objects

226 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language

shape when what was at issue was its role as figure in a conceptualized as enclosures or containers (see sect. 2.9
locational expression. for other usage distinctions based on convention and
In a further experiment, Landau and Stecker asked idiom).
whether subjects would ever preserve any components of
object shape in the preposition condition. This time,
2.4.2. Axial structure. A sizable number of prepositions,
subjects were shown a straight rod lying across the top of a
such as on top of, under, in front of, in back of, and
box. The test objects included a straight rod, a squiggly
beside, make reference to an object's axial structure and
rod of the same extent, and a cube, each placed in
its axially determined parts, such as top, bottom, and sides
different positions on and off the box. As in the first
(sect. 1.3). In the case of an object that lacks inherent axes,
experiment, children and adults preserved precise shape
such as a sphere, axes are contextually imposed (see sect.
in the count noun condition (accepting only the straight
2.7.2). But these prepositions can also make use of the
rod) and location in the preposition condition. But in the
reference object's inherent axes. On top of and under
preposition condition, they also accepted both rods of
project a region from whichever directed axis is vertical in
equal extent, even though they were quite different in
the object's normal orientation. In front of, in back of,
precise shape. Thus, a novel preposition led subjects to
and behind make use of the directed horizontal front-to-
preserve both location and a very general component of
back axis; beside and alongside make use of a horizontal
the figure object's shape - linearity - while still ignoring
axis perpendicular to the front-to-back axis. For the pur-
the detailed shape.
poses of these prepositions, it does not matter whether
These studies tell us that even young children pay the axes in question are generating axes or orienting axes
attention to only very sparse properties of objects when (see sect. 1.3).
what is at issue is their location: Either they ignore shape
As mentioned above, along requires its reference ob-
entirely or they preserve only very crudely schematized
ject to be basically linear and horizontal. Its partner
components of shape. We assume that children come
across appears to require its reference object to be or to
prepared or learn very rapidly to draw on qualitatively
have a surface with sides, so that one can go across, "from
different kinds of representations when learning the
one side to the other." The best case appears to involve
names of objects (count nouns) versus the names of spatial
opposite sides, although adults accept almost all cases
locations. We return to this issue in section 3.2.3.
where one object intersects two segments of another
We will now enumerate the linguistic evidence for object (e.g., two arcs of a circle or two adjacent sides of a
overall constraints on the range of geometric properties rectangle, see Williams 1991). Just in case the reference
relevant to the description of figure and reference ob- object has a significant linear elongation, the sides are
jects. Reference objects are the more complex of the two distinguished from the ends: A square table has four sides,
and we describe their constraints first. (A summary of but a long rectangular table has two sides and two ends. In
constraints appears in Table 2 at the end of sect. 2.) such a case, across appears to pertain specifically to the
sides and not to the ends, so that across the rectangular
2.4. Reference objects table describes a region that traverses the table's shorter
dimension.13 Nonrectilinear objects in this framework
The few restrictions on reference objects concern their
tend to be schematized as though they were rectilinear, so
treatment as certain geometric types (volumes, surfaces, that, with respect to across, a round table behaves like a
points, and lines), their axial structure, and their quantity. square table and an oval table like a rectangular one.
We take these up in turn. A further restriction on across is that, like along, it
describes a horizontally oriented region. For example,
2.4.1. Volumes, surfaces, points and lines. The terms in,
one draws a line across a blackboard in the horizontal
on, near, and at require very little in the way of detailed
direction, not the vertical. This follows from the stipula-
geometry (see Herskovits 1986, for details on normal use
tion that across pertains to the sides of the object, which
types). There is not even any requirement for particular are normally the boundaries of a horizontal axis.
axes. For something to be in X, X must have an interior,
but nothing more is necessary. In other words, the refer-
ence object for in needs a form descriptor less specific 2.4.3. Quantity. A different sort of restriction on the
than any particular generalized cone, something like a reference objects appears in the distinction among the
"lump" or "blob" that would indicate its capacity to sur- prepositions between, among, and amidst. For between,
round or contain. Similarly, near and at require only that the reference object is not a single object, but rather a
the reference object be bounded in extent; they place no pair. In the case of among and amidst, the reference
requirements at all on its shape. On is slightly more object is an aggregate (or collection of objects), as in
complex: It requires that its reference object possess a among the people or amidst the waves.
surface, whether it be a line (on the border), a surface (on To sum up, the restrictions placed on the form of the
the square), or an object with a boundary that is a line or a reference object by expressions for spatial relations are
surface (a house on the lake or on the hill, respectively). not at all severe, compared to the potential complexity of
A minor distinction appears in the contrast between objects themselves. At most, these restrictions appeal to
inside and in. Inside is the more specific, and seems to the very gross geometry of the coarsest level of represen-
require that its reference object be or contain a bounded tation of the object - whether it is a container or a surface.
enclosure (a negative part or the interior of a hollow In addition, the object's axial structure plays a crucial
volume). Thus, as pointed out by Talmy (1983), one can be role, that is, whether the object is relatively elongated,
either in or inside a cave or a bottle; but one can be only in, whether the elongation is horizontal, and whether it has
not inside, a swimming pool or lake, because these are not sides.14

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 227

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language

2.5. Figure objects over, it can be a pair of objects or an aggregate that forms a
linear virtual object, so that one can say "The house is in
There are even fewer constraints on figure objects than on
line with the trees." If the figure object is not linear (e.g.,
reference objects. For the rriost part, the figure object
the house in the previous sentence), it is simply under-
needs no geometric specification at all; Talmy (1983)
stood as lying on the axis of the reference object or the
suggests that the figure is usually conceived as "point-
extension of this axis at any distance. If the figure object
like." We have found only two specifications of figural
has a linear axis, however, as in "The road is in line with
geometry for the English prepositions: axial structure and
the trees," this axis must be aligned with that of the
quantity.
reference object.
2.5.1. Axial structure. Four prepositions, along, in line Across, as mentioned above, involves a linear region
with, across, and around, express spatial relations be- that goes from one side to the other of the reference
tween the reference object and the linear axis of the figure object. Various senses of across locate the figure object
object. For example, consider "The road is along the differently with respect to this region. The two most
river." This specifies that the main axis of the figure object relevant ones place the figure object within the region,
(the road) is parallel to (as well as horizontally proximate where it is either (1) linear and coaxial with this region
to) the main axis of the river. If the figure object is an "the stick lay across the road"), or (2) distributed along the
aggregate, as in "The trees are along the river," this axis of the region ("The trees extend across the field"). As
aggregate is preferentially understood as forming a virtual with along, other senses do not require a linear figure.
object with an axis that is parallel to the main axis of the For example, one sense (3) places the figure on the other
river (see Figure 8). The adjectival form parallel to places side of the. region in relation to the observer or a second-
constraints similar to those ofalong. (Note, however, that ary reference object ("Bill is across the road (from here)").
not all uses of along impose this linearity condition, in Another (4) expresses a figure moving along the axis of the
particular, if the figure has no main horizontal axis, as in region ("Bill ran across the road").15
"The tree is along the river." Similarly, if the figure object Around also has a number of variants. Other than the
is in motion, as in "the dog loped along the river," it is the one that means roughly near ("There are lots of trees
trajectory of the figure rather than the figure itself that around here"), they all specify a hollow region with an
must be conceptualized as linear and parallel to the main interior that contains the reference object. The constraint
axis of the river.) on the figure is that it occupies this region, either as a
The compound preposition in line with, like along, linear object surrounding a two-dimensional reference
requires a linear reference object. Unlike along, how- object ("The road goes around the city") or as a shell or
ever, the reference object need not be horizontal; more- thickened surface surrounding a three-dimensional refer-
ence object ("There is chocolate around the core of the
candy"). A distributed figure object is again acceptable
("There are trees around the house"). As with along and
across, a moving figure need not be a line or a surface; it
may either circumnavigate the reference object (go all the
way around) or detour around it.

2.5.2. Quantity. A different class of prepositions requires

the figure object to be distributed in space, either as a
fa) The road is along the river. substance or as an aggregate. Consider: "There was water
all over the floor." All over specifies a figure object (water)
distributed over and in contact with the entire extent of
the surface of the reference object. In "There were raisins
throughout the pudding," throughout specifies an aggre-
gate figure object (raisins) more or less evenly distributed
in the volume of the reference object. Thus, all over and
throughout are "distributive" forms of the spatial relations
normally expressed as on and in, respectively. All along,
(b) The trees are along the river. all around, and all across are similar distributive forms
corresponding to the prepositions along, around, and
across.

2.6. Summary
In comparison to the rich description of objects as cate-
gory members (named by count nouns), most of the
geometrical distinctions among objects are disregarded
when specifying their role as either figure or reference
(c) The dog loped along the river.
object in a spatial relation. A reference object can be
Figure 8. The preposition along requires a linear figure and schematized as a point, a container, or a surface, as a unit
reference object. In (a), both figure and ground are ribbonlike; with axial structure, or as a single versus aggregate entity.
in (b), the aggregate of trees forms the linear figure; and in (c), No more detail is necessary. Similarly, a figure object is
the dog's path is the figure that is linear. schematized at most as either a simple lump or blob (with

228 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language

no geometric structure whatsoever), a unit with axial region of the reference object. For example, the cat can
structure along at most one of its dimensions, or a single be said to be on the mat just in case it is located in contact
versus distributed entity. Further internal complexity of with the uppermost surface of the mat; that is, on maps
the object, describable in terms of componential configu- the reference object (mat) into the region consisting of its
ration or object parts, is simply irrelevant. (upper) surface. The bicycle is near the house just in case
Before turning to a discussion of regions, we should it is located within a region surrounding the house and
note some examples from other languages that might extending up to some critical distance; near thus maps the
appear to conflict with the description set forth above. reference object into such a region.
These concern spatial verbs with an application that is The notion of region appears necessary to describe the
restricted to figure or reference objects belonging to a meanings of spatial prepositions (see e.g., Miller &
certain object category or to those possessing specific Johnson-Laird 1976). It is also necessary to describe
configurational properties. For example, Talmy (1983) people's nonlinguistic representation of space. Kevin
describes Atsugewi, a Native American language of Cali- Lynch (1960) was one of the first to show the importance of
fornia, in which there are roughly a dozen verb suffixes components such as "nodes" and "districts," which are not
that mark distinctions finer than the English preposition evidently circumscribed in physical space but serve as
into. The distinctions concern the geometry of the refer- foundational elements in people's conceptualization of
ence object, with separate suffixes for "into an aggregate," the cities in which they live.
"into a gravitic container" (e.g., basket, pocket), "into an Several studies by cognitive psychologists have pointed
areal enclosure" (e.g., a field or corral), "over the rim into to the importance of regions in organizing people's knowl-
a volume" (e.g., gopher hole or mouth), and so on. edge of space. Stephens and Coup (1978) showed that
English prepositions distinguish a few such cases; for people's distance and direction judgments were not iso-
example, inside can be used only for enclosures. morphic to physical space but rather were biased by the
There are also spatial verbs that differentiate according organization of individual cities into larger regions de-
to figure object type. Talmy (1985) describes a set of fined by the states in which they were located. Hirtle and
spatial verbs in Atsugewi that have restrictions on the Jonides (1985) showed that people tend to judge distances
figure object, for example, "for a small shiny spherical within a particular region as smaller than they do when
object to move/be-located," "for a slimy lumpish object to the same distances cross regions.
move/be-located," "for runny icky material to move/be- Given that regions are part of our spatial representa-
located," and so on. Analogous English verbs include "to tions, we can ask how many ways spatial prepositions can
spit" and "to rain." Similar distinctions are reported by encode regions. What kinds of distinctions among regions
Bowerman (1991) in her description of Tzeltal. There, are encoded by the set of spatial prepositions? In princi-
spatial relationships are apparently expressed with ple, they could represent regions in acute detail, for
"closed-class positional" verbs that classify on the basis of example, in terms of precise distance and direction from
the figure. Thus, there are separate terms for the follow- the reference object, using coordinate systems and the
ing types "to be located": (for) a bowl-shaped figure (to be specialized domain of measurement terms. This degree of
located), a narrow-mouthed container in upright position, precision is surely represented in the spatial cognitive
an inverted object with flat surface down, a small sphere, system and is used in some cases, although not in all cases
a large sphere, and things in a bulging bag. Analogous as evidence from nonlinguistic studies such as Hirtle and
distinctions in English might be verbs like smear or pin, Jonides suggests (see sect. 3 for further discussion).
which describe kinds of attachment restricted to different Within language, the very restricted number of spatial
classes of figural objects: The kinds of things you can prepositions suggests that precise location is not encoded
smear on the wall are quite different from those you can in any individual term. And, as it turns out, the spatial
pin on the wall. relations expressed in English factor into just a few inde-
These cases appear to encode more detailed geometry pendent features that combine to produce some of the
in the figure and reference object than do English prepo- complexity of the system. The principal features include
sitions, although they still do not encode as much detail as several degrees of distance and several kinds of direction.
is found in object names. From our perspective, these
cases are somewhat different because they involve verbs 2.7.1. Relative distance. This factor concerns the distance
that are open class and clearly encode more than just between figure and reference object. The most salient
location. It is well known that verbs impose all kinds of fact is that distance is digitized into several discrete
restrictions: Drink applies only to ingestion of liquids, categories. There appear to be four levels described by
disperse applies only to aggregates, diagonalize applies English prepositions: (1) location in the region interior to
only to matrices, and the German verb fressen can be the reference object (in, inside); (2) location in the region
predicated (nonironically) only of animals. However, the exterior to the reference object but in contact with it (on,
cross-linguistic evidence does offer interesting chal- against); (3) location in the region proximate to the refer-
lenges, to which we return after presenting the remaining ence object (near); and (4) location distant from the
evidence and our hypotheses. reference object (far and perhaps beyond).
Some languages provide additional values for the dis-
tance feature, for example, "not near but within reach" or
2.7. Constraints on regions
"not near but visible." One such language is Korean (Soo-
Recall that the third element of a linguistically expressed Won Kim, personal communication), in which the expres-
spatial relation is the region (of the reference object), in sions yup and kiyut both translate as "near," but yup is
which the figure is located. A spatial preposition can be confined to more immediately proximate cases. (Other
thought of as a function that determines the relevant examples are cited by Anderson and Keenan [1985] in

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 229

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language

connection with systems of spatial deixis corresponding to sit on top of it (referring to the contextually determined
English here and there.) On the other hand, some lan- top, in this case a long horizontal surface). Levelt (1983)
guages have prepositions that collapse two adjacent values and Olson and Bialystok (1983) describe the many com-
of distance in English. For example, English in and on can plexities that arise from assigning different frames of
both be translated by Spanish en. Thus, English repre- reference.
sents an intermediate degree of complexity in the dis- Not all spatial expressions involving axes leave this
tance parameter. To our knowledge, no language encodes choice of reference system open. On the top of, by
more than five or six levels of distance, although we are contrast with on top of, refers only to the inherent top of
able to represent distance at much finer levels for other the reference object (presumably because it contains the
cognitive purposes. full noun phrase the top). Beyond, by contrast with be-
At least two of the degrees of relative distance found in hind, refers only to the region projected to the contex-
English have a corresponding "negative," which actually tually determined rear of the reference object. Which-
means "farther away from the reference object than." ever type of reference system is used, however, its
These are "farther away than the interior" (out of, outside), structure still depends on analyzing the reference object
and "farther away than in contact" (off of). The pair near into its three principal axes.
and far from might also form such a contrast.
Several prepositions involve the distance feature in a 2.7.3. Combinations of distance and direction. Further
less obvious way. As mentioned above, the reference distinctions can be derived among prepositions by various
object for among is an aggregation of objects, which combinations of distance and direction. For example,
together define a group or virtual object that contains compare over, above, and on top of. Over is indifferent to
them all. The figure object is then specified as interior to contact versus noncontact: A cloth may be put over a table
this virtual object. The case of between is similar, except (contact), and clouds may fly over a city (noncontact).
that the virtual object is the minimal space bounded by Above, however, specifies noncontact: Though clouds
the pair of reference objects. may fly above a city, one can only put a cloth above a table
by putting it on a higher shelf. Finally, on top o/strongly
2.7.2. Direction. Direction of the figure object from the favors a contact reading.
reference object provides the second key parameter in In back of and behind, which share directionality, also
specifying spatial relations. Crucially, the entire set of differ in distance. A tree may be right behind (proximal),
directions derives from the axial structure of the refer- way behind (distal), or right in back of a house, but "The
ence object (or its axially- determined parts): The three tree is way in back of the house" sounds odd or colloquial
principal axes can be viewed as extending from the center (to us, anyway). The standard use of in back of seems to be
of the reference object to provide six possible directions. restricted to proximal distance (and possibly contact),
Centered around each half-axis is a region that defines the whereas behind and the colloquial in back of are unre-
acceptable space for different prepositions. stricted. In any case, distance can be combined with
Regions determined by the vertical axis are given (in direction to yield finer distinctions.
the canonical case) by gravitation, defining over, above, For a somewhat different case, to move up or down
under, below, and beneath. Orthogonal to gravitation is a mountain, tree, or wall is to move in an upward/
the horizontal plane, which helps define beside, by, downward direction while maintaining contact with (or,
alongside, and next to. To see that the horizontal is crucial marginally, proximity to) the surface of the reference
to these terms, notice that if a bird is beside, by, object.
alongside, or next to a house, it must not be on the roof or Sergey Avrutin (personal communication) has pointed
flying overhead: It must be in proximity to the house and out that the conceptual features of distance and direction
no higher than the house. Thus, these prepositions desig- can be used to predict the case-marking pattern of Rus-
nate the relation "proximate to the reference object in the sian place prepositions. According to his analysis, the two
horizontal direction." If, in addition, the reference object closest grades of distance - interior and contact - assign
has inherent axes that distinguish front and back from the prepositional case to their objects; prepositions that
sides, the terms beside, by, alongside, and next to tend to involve an axis-based directional feature (over, under,
mean "horizontally proximate to the sides of the reference behind, etc.) assign the instrumental case; the remaining
object." For example, Bill is not beside me if I am facing prepositions (at, near, close to, etc.) assign the genitive
him. case. 16 Thus, these features are grammatically as well as
Similarly, if the reference object has an axis that distin- conceptually significant.
guishes an inherent front and back, in front of can mean
"horizontally proximate to the inherent front of the refer- 2.7.4. Visibility and occlusion. A further distinction ap-
ence object," and in back of can mean "horizontally pears to be subsidiary to (but to some degree independent
proximate to the inherent back of the reference object." of) the distance and direction distinctions. A case in
However, an alternative interpretation of these preposi- English where this distinction is evident is in speaking of
tions results from contextually assigning a front-to-back paint on a wall being on top of or underneath the wall-
axis to the reference object: The front is the surface facing paper. Here on top of evidently means "in contact with
the speaker (or addressee) and the back is the surface visible surface." whatever its orientation, and underneath
opposite. In this case in front of and in back of mean means "in contact with the surface opposite the visible
"horizontally proximate to the contextual front/back of surface." (Notice, by the way, that one cannot speak of the
the reference object." A parallel ambiguity occurs with on bottom of the wallpaper in this context; not all the words
top of: If a flagpole is lying on its side, one can paint the of vertical orientation generalize to this use.) Vandeloise
ball on top of it (referring to the inherent top), or one can (1986) argues that occlusion of the reference object is the

230 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language

main relation expressed by French devant ("in front of"). path that if extended backward would begin at the
We would not go quite so far, but we believe this criterion explosion."
does play a secondary role, possibly forming a preference
rule system (Jackendoff 1983) with the directional
2.9. Other factors
criteria.
Visibility and occlusion may well play a role in chil- The discussion above has presented more than just a
dren's learning of spatial prepositions. Johnston (1984) sample of the spatial relations expressed by English
found that 3-year-olds understood behind to mean "oc- prepositions. Rather, we believe we have extracted from
cluded from sight." For example, they agreed that X was the complicated grammatical and pragmatic facts about
behind Y only when Y was an object large enough to prepositions all the purely spatial information they are
occlude X. In contrast, adults agree that X is behind Y capable of encoding. This section will list briefly some of
whatever their relative sizes, as long as the figure object is the complications that remain, none of which involves
aligned properly with the reference object's front-back geometric properties per se.
axis. Combined with the evidence from French, this First, there are uses of prepositions that involve special
suggests that visibility versus occlusion may indeed be an functional situations. Herskovits (1986) points out that to
additional distinction in spatial relationships. be at a desk or at a sink usually implies more than being
located close to it; one is probably performing characteris-
tic actions, such as writing at the desk or washing at the
2.8. Spatial relations defining paths sink. For another case involving at, to throw a ball at X
In addition to regions, language expresses another spatial involves more than throwing it toward X, namely, an
category of paths or trajectories to specify a figure's intention to hit X. This difference accounts for the con-
motion ("The bird flew to the house") or orientation ("The trasts in (6):
sign points to New York"). There are only a few simple (6) a. Bill threw the ball toward/?at Harry without meaning to
ways of constructing trajectories, none of which draw any hit him.
further on the geometry of the figure or reference objects b. Bill shot at/?toward Harry.
than we already have. Once again, the main geometric
property involved is axial structure. Other special situations involve conventionalized con-
One class of paths specifies the figure's motion in terms ceptualizations of the reference object. For example,
of its own inherent horizontal axes: forward, backward, when traveling, one is in a bus or on a bus but only in, not
and sideways. Another specifies change of the figure on, a car. It seems that in English, large vehicles (buses,
object's orientation, again in terms of its own axes: turn yachts, trains, large airplanes) are conceptualized either
around, turn over, turn left, and turn right. Another as containers that one is in or sorts of platforms that one is
class draws on the axial structure of the earth: These terms on, but small vehicles (cars, rowboats, small airplanes) are
are the environmentally oriented directions up, down, only conceptualized as containers. The choices here show
north, south, east, and west. that alternative conceptualizations are available; presum-
The largest class of paths, however, is constructed from ably, the particular choice has less to do with principles of
the class of regions by attaching one of a set of five spatial representation than with historical and pragmatic
operators: via, to, toward, from, and away from (Jacken- issues. For a somewhat different case, a container can be
doff 1983). The operator "via" creates a path that passes conceptualized either in terms of the volume it surrounds
through a region that is in turn defined by the parameters or in terms of the body of its substance, so we can speak of
described in section 2.7. For example, to go through a either the water in the cup or the crack in the cup.
room involves a path that at some point involves being in Some uses of prepositions appear to involve forces
the room; to run by the house is to traverse a path that at exerted between the figure and the reference object. For
some point involves being near the house; to walk under a example, the preposition on is frequently said to involve
bridge is to follow a route that at some point involves support by the reference object (Cienki 1988; Herskovits
being under the bridge. If the region in question is linear 1986). This is not always the case, because we can speak of
for example, as in along, across, and around, the "via" the fly on the ceiling; but it may be a default interpreta-
path is coaxial with the region, so that going along X tion. According to Bowerman (1989), the Dutch preposi-
involves moving parallel to the axis of X. tion aan also involves support or attachment, specifically a
The operator "to" creates a path that terminates at the figure object hanging or projecting from a reference
region in question. For example, to X expresses a path object that is something other than a horizontal surface
that terminates at X. Into X and onto X express paths that (for example leaves aan a twig, a coathook aan a wall,
terminate in X and on X, respectively. Similarly, the clothes aan a line). The English preposition against, as in
operator "toward" constructs a path that would terminate "Bill leaned against the wall" or "The tree fell against the
at the region if extended, but does not in fact reach the house," describes contact with exertion of force, usually
region. To go toward X is therefore to undergo a motion in a horizontal or oblique direction. Among expressions of
that if extended would terminate at X. path, there is a reading of into found in "The car ran into
the pole," which means not traversal to the interior of the
The operator "from" is just the reverse of "to," perhaps
reference object, but rather coming into contact with the
forming a pair similar to in/out and on/oj[jf (Gruber 1976):
reference object with considerable force.
It constructs a path that begins at the region in question.
Examples are: "The bird emerged from under the table" Bowerman (1989) describes other uses of force-
and "The train came from inside the terminal." The dynamic properties encoded in Korean spatial verbs. It
operator "away from" is the reverse of "toward": "Bill ran appears that the verb kki-ta, roughly "put in, put to-
away from the explosion" describes him as traversing a gether," applies to situations in which the figure object

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 231

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language

fits fairly tightly into or around the reference object, such Table 2. Features of spatial relations
as a ring on a finger, a hand in a glove, a lid on ajar, and a (with sample prepositions)
button in a buttonhole. (The English verb insert appears
to cover part but not all of the same semantic territory; the Reference object geometry
verb fit also appears close in meaning.) The verb ppay-ta Volumes, surfaces, and lines: in, on, near, at, inside
describes the removal of the figure from a reference One or two axes
object with which it has been configured in this fashion. Vertical: on top of
This application of force-dynamic properties also occurs Horizontal: in front of, in back of, beside, along, across
in English verbs such as clasp, snap, and impress (see Quantity: between, among, amidst
Pinker 1989 for discussion). The Korean verbs again make Figure object geometry
little reference to the detailed geometry of the figure and One axis: along, across, around, in line with
the reference object, however, other than the fact that Distributed figure (substance or aggregate): all over, through-
there is a match between a positive part of one and a out, all along, all around, all across
negative part of the other.
Relation of region to reference object
Beyond this sort of complication, most of the complex-
ity of English prepositions appears to involve (1) how Relative distance
spatial configurations that are nonstereotypical or ambig- Interior: in, inside, throughout
uous are forced into the expressions available in the Contact: on, all over
language, (2) how particular prepositions are extended Proximal: near, all around
from core place meanings to different sorts of related Distal: far
paths and places (for example, the variants of across "Negatives"
mentioned above), (3) how preposition meanings are ex- Beyond interior: out of
tended to nonspatial domains such as time and posses- Beyond contact: off of
sion, and (4) how prepositions are used as purely gram- Beyond proximal: far
matical markers ("Bill believes in capitalism," "The letter Direction
was received by Bill," and "a picture of Bill"). Extended Vertical: over, above, under, below, beneath
discussions appear in Brugman (1981), Cienki (1988), Horizontal
Herskovits (1986), Jackendoff (1983; 1990), Lakoff(1987, Side-to-side: beside, by, alongside, next to
Chapter II.2), Miller and Johnson-Laird (1976), and Van- Front-to-back: in front of, ahead of, in back of, behind,
deloise (1986).lS Our goal Tiere has been to use linguistic beyond
research to see through all this complication in order to Choice of axis system:
find the specific characteristics of spatial representation Inherent: on the top of, in front of, ahead of, behind
necessary for the range of preposition meanings. These Contextual: on top of, in front of, behind, beyond
characteristics have proven to be quite limited. Visibility and occlusion: on top of, underneath
Paths (trajectories)
Earth-oriented: up, down, east, west, north, south
2.70. Summary Figure object axis-oriented: forward, ahead, backward, side-
The description of figures, reference objects, and regions ways, left, right
in English (both places and trajectories) recruit just a few Operators on regions
geometric properties and distinctions. The geometry of Via: through (= via inside), along (= via along)
figure objects specifies at most a single axis, whereas that To: to, into (= to in), onto (= to on)
of reference objects specifies at most three principal axes Toward: toward
of the object. The regions relevant to describing places From: from, from under, from inside
make further use of the axes, adding qualitative distinc- Away from: away, away from
tions pertaining to the distance and direction of the figure
from the reference object. The regions relevant to de-
scribing trajectories or object motions then draw on these
place descriptions, adding operators that specify the loca- tioned in sect. 1 with essentially the entire repertoire of
tion of the path relative to a given place, and where that English spatial prepositions as described in sect. 2).
path begins and ends. Table 2 provides a summary of the It is more significant that our investigation shows se-
factors discussed here. vere constraints on the ways in which the meanings of
spatial prepositions can invoke object geometries. It is as
if the spatial relations expressed by prepositions filter
3. Why do spatial prepositions make so little use object descriptions, removing much of the detail of object
of object shape? shape and preserving only certain key properties, primar-
ily the boundedness, surface, or volumetric nature of an
The picture that emerges from this brief overview (if it is object and its axial structure. Because of this extreme
anywhere near complete, as we believe it is) is that the limitation, more complex meanings such as those as-
geometric descriptions relevant to words describing what cribed to the hypothetical words sprough, betwaft, and
an object is are very different from those describing where plin are simply not available as possible meanings in the
an object is. This is initially suggested by the fact that language's basic stock of prepositions.
spatial prepositions in English are quite few in number This is not to say that one cannot express such mean-
relative to the class of object names (even taking polysemy ings; we clearly can, using combinatorial expressions.
into account, just compare the number of nouns men- Our claim is that such meanings will arise either through

232 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language
derivation (e.g., aboard), as technical terms (e.g., abaft), beginning of the reach the hand's exact angle of orienta-
or by composition with open class words (e.g., through a tion relative to the slot. Or one might initially make a
cigar, between the nose and mouth, 3 meters from the rough ballistic reach toward the slot and then modify it
door, etc.). using visual feedback about error. In either case, precise
One could stop here and accept this result as an inter- spatial representations of distance, angle, and joint posi-
esting fact about language. But to us it cries out for tion must be available to be converted into equally precise
explanation. We see two possible approaches. The first, muscular forces. A similar situation exists for the task of
which we will call the Design of Language Hypothesis, throwing a ball to someone, catching a ball, and, in
claims that the limitation on spatial relations expressible general, any act of navigation based on either visual or
in language is indeed just a fact about language, and that haptic-kinesthetic perception (see Gallistel 1990, for a
spatial cognition is just generally much richer in the discussion of the geometries underlying navigation).
spatial relations it can encode. According to the other Such metric representations of space emerge sponta-
approach, which we will call the Design of Spatial Repre- neously early in life; they are not the product of lengthy
sentation Hypothesis, this limitation in language reflects formal tutoring and are therefore probably part of our
a deeper constraint on how spatial cognition encodes the biological endowment. For example, evidence from von
relations among objects. We will argue that both fac- Hofsten (1980) suggests that even 4-month-old infants can
tors contribute to the relative scarcity of preposition "catch" both stable and moving objects successfully under
meanings. some circumstances. It is significant that the infant's
initial angle of trajectory is mapped more closely to the
object's (predicted) final position than to its actual posi-
3.1. Design of Language Hypothesis
tion at the initiation of the reach. Other evidence shows
According to the language design hypothesis, spatial that by around 2 or 3 years of age, children can use
representation can itself encode a rich range of spatial knowledge of metric properties of spatial layouts to search
relations, making use of detailed properties of object for objects, guide navigation, and use maps (DeLoache
shapes. Most of these are "invisible" to the language 1987; Landau 1986; Landau et al. 1984; Rieser & Heiman
faculty, however, and therefore neutralized or leveled out 1982).
in the translation into linguistic format. This hypothesis Yet, despite its importance to motor control and naviga-
implies that filtering should apply equally to the two tion and its naturalness in human perceptual and cogni-
domains of object and place and that it might be a design tive development, precise metric information is simply
feature required by any system that must collapse numer- not encoded in the language's stock of spatial terms, a
ous complex distinctions into a finite set of elements. 19 point emphasized by Talmy (1983). It is possible to be
There is abundant evidence that language does indeed precise in expressing distances and orientations, but to do
filter representations of spatial properties and relation- so, one must invoke a culturally stipulated system of
ships. For instance, language does not typically represent measurement that operates by counting units such as
the dimensions of objects in analog fashion, but rather meters or degrees (go 30 meters, turn 30 degrees).
digitizes them. Thus, dimensional adjectives such as Could this filtering account for the severe limitation in
biglsmall, thick/thin, and tall/short refer to continuous the ways spatial prepositions can take object shape into
dimensions of size, but the linguistic terms bifurcate account? It cannot be the whole story, for filtering also
these dimensions into pairs of relative contrasting terms. takes place in translating object shape descriptions into
Such binary (or ternary) relative contrasts are characteris- language. For one thing, not everything called dog is
tic of most adjective domains; they occur across most precisely the same shape; that is why a similarity metric is
languages and may be more natural than absolute con- needed for categorizing objects. And on the nonlinguistic
trasts in first language learning (Landau & Gleitman side there is abundant evidence that humans can encode
1985). detailed aspects of shape that do not appear in language.
Especially compelling evidence for this binary/relative For instance, one can recognize with great accuracy
constraint is provided by Newport (1988) and Supalla complicated contours and surface patterns (and this im-
(1990) from the study of American Sign Language. As proves with perceptual learning (Gibson 1969). To suc-
Newport points out, the manual mode affords a straight- ceed in tasks such as mental rotation and composition, we
forward means of expressing continuous quantities in an apparently represent in detail the metric composition and
analog fashion. For example, the sign for slow is repre- relationships among object parts (Cooper 1989), yet these
sented by moving the finger slowly along the opposite detailed descriptions are very hard to describe to some-
forearm. It would be feasible to "mimic" different rates in one else in words. Imagine, for example, trying to de-
a much more precise fashion, perhaps even matching the scribe the Shepard-Metzler objects or the Attneave fig-
object's actual rate of movement in one's sign. ASL, ures used in visual image rotation experiments (Shepard
however, like spoken languages, expresses different & Cooper 1982), the pattern of stripes on a particular
speeds using only two or three distinctions, qualitatively zebra, the shape of a violin, or one's mother's chin. What
categorizing the physically continuous dimension despite actually happens when we try to describe complex figures
the fact that the modality affords a more detailed repre- is that we describe them in terms of their parts, or with
sentation of speed. allusions to familiar objects, such as "spider" for a thing
with "legs" projecting from a round center (Fussell &
As another argument for filtering, notice that the spa-
Krauss 1989).
tial representations recruited by the motor system must
be much richer than those appearing in language. Con- These difficulties in describing precise shape are exac-
sider the task of inserting one's hand through a narrow erbated by the absence of linguistic terms for describing
slot. One might accomplish this by predicting from the exact sizes of objects (again apart from a culturally stipu-

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 233

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language
lated system of measurement). That is, thefilteringout of location. Such a task clearly requires the ability to select
metric information that occurs in the expression of spatial on the basis of pattern (or, in some cases, shape); it is
relations also occurs in the expression of objects. selectively impaired by damage to the inferior temporal
In short, language does not convey all the representa- cortex.
tional richness we have for encoding either locations or On another kind of task, an animal might be required to
objects. (That is why a picture is worth a thousand words.) find food when it is hidden in one of a pair of identical
However, thisfilteringout of metric information does not objects that is nearer a particular reference object. This
by itself explain why the encodings of object shape are requires the ability to select on the basis of position and it
limited in the particular way they are when spatial rela- is selectively impaired by damage to the posterior parietal
tions are at issue. If the problem were merely a design cortex. More generally, damage to the posterior parietal
need to filter, prepositional meanings could filter objects cortex impairs route following, reaching for objects, and
coarsely on many different properties, by representing using landmarks to locate objects.
just two values of brightness or color, size, texture, Ungerleider and Mishkin also review evidence sug-
animacy, and so on. But prepositions filter objects in gesting that the neurons in the two cortical regions in
particular ways, preserving just global and axis-based question have distinctly different receptive field proper-
structure. We are therefore still left with the question of
why objects that are being named are differentiated in ties. Most neurons in the inferior temporal cortex (the
relatively complex geometric terms, whereas objects that "what" system) are driven by complex sets of features and
are being located and the regions in which they are their receptive field is large: More than half have a
located are treated in terms of relatively simple schematic bilateral receptive field and all include the fovea. This
geometric descriptions. What accounts for this differ- would mean that information about the position of any
particular object shape being processed by these cells
ence r would be lost - that is, the position of the pattern would
simply not be represented. In contrast, neurons in the
3.2. Design of Spatial Representation Hypothesis: The parietal lobe (the "where" system) are not sensitive to
"what" and "where" systems stimulus features and most do not include the fovea in
their receptive field.
One possibility is that the disparity may be inherent in the Although Ungerleider and Mishkin's evidence is de-
spatial representations underlying language. According rived primarily from lesion studies on animals, there is
to this hypothesis, spatial representation is relatively rich converging evidence from human psychophysical studies
in its possibilities for describing object shape, but rela- that there may be two distinct streams of visual process-
tively limited in the way it can use object shape to encode ing: the "parvo cellular" system, specialized for detecting
spatial relations. If this is the case, the disparity observed color and detailed object shape, and the "magno cellular"
in language is a consequence of the disparity in the spatial system, which is color-blind but specialized for detecting
representations that language encodes. motion, depth, and location. These streams are segre-
We conjecture that this is indeed the case, and that the gated at relatively low levels of the visual system and the
disparity in spatial representation is partly reflected in segregation appears to become more pronounced at
some basic organizational facts about the human brain, in higher levels, providing converging evidence for the what
particular, that it arises in part from a functional bifurca- and where distinction (Livingstone & Hubel 1989; but see
tion of the system of spatial representation (perhaps into Van Essen et al. 1992 for a suggestion that the systems
"submodules" in the sense of Fodor 1983 [see also BBS overlap considerably at early stages of processing).
multiple book review of Fodor's The modularity of mind, Human clinical evidence appears to support the
BBS 8(1) 1985], as refined by Jackendoff 1987b, Chapter what/where distinction as well. Farah et al. (1988) docu-
12). One part of the system is devoted primarily to objects ment a case in which bilateral damage to inferior temporal
and their identification (mostly by shape), the other to areas with sparing of parietal regions produced a deficit in
locating objects in space relative to each other and to the a wide range of tasks involving shape recognition but
observer. The expressive power of the system of nouns preserved normal performance in tasks involving object
that identify objects is linked to the shape identification localization and spatial relations.20 Evidence from Levine
submodule; the expressive power of the spatial preposi- et al. (1985) suggests that the imagery system might also
tion system is linked to the submodule governing the carry these distinctions. One of their patients could imag-
location of objects relative to each other. ine object shapes but not spatial relationships or object
layouts; another could imagine spatial layouts, but not the
3.2.1. Nonlinguistic evidence. Our conjecture finds inter- shapes of individual objects.
esting correlative support in neurological, psychological, Finally, evidence from formal modeling of simple
and computational evidence. learning systems supports the distinction. Rueckl et al.
Neurological evidence reviewed by Ungerleider and (1988) found that in a PDP model of a very simple visual
Mishkin (1982), building on previous work by Schneider; system a certain degree of extra efficiency accrues to a
(1969; see also Ingle et al. 1967), suggests that the brains system that strongly separates computation of the "what"
of monkeys contain separate areas specialized for object and "where" functions as long as both subsystems have
identification (the "what" system) and object location (the sufficient computational resources. In their study, Rueckl
"where" system). These specializations have been in- et al. used a stimulus space containing only nine different
ferred from selective deficits following damage to differ- shapes, each of which could occur in nine partially over-
ent areas of the cortex. For example, in one kind of task, lapping locations. Within this tiny system it was found
an animal might learn tofindfood when it is hidden in one that the optimal allocation of resources between the
of a pair of distinctively patterned objects, regardless of its "what" and "where" systems used over three times as

234 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language

many "what" units as "where" units. It is hard to know how tion). A map cannot just specify "wheres": It has to have
this case generalizes to a more realistic system (or to other something to stand in for the objects being located.
different learning systems), but the disparity is reminis- Typically, these stand-ins are points and lines, with some
cent of the one we have found between noun and preposi- conventionalized symbols to distinguish different sorts of
tion systems. Rueckl et al. in fact argue that the disparity objects from each other (big cities vs. small cities, main
can only increase as one moves to a more realistic system. roads vs. subsidiary roads, churches vs. hospitals, etc.). If
Thus, our Design of Spatial Representation Hypothesis a map had to distinguish all the objects by their shapes, it
is suggestively supported on a variety of empirical would be much more complex and quite possibly
grounds. We say "suggestively," because the evidence unusable.
described above will no doubt be subject to revision as We are not suggesting that the "where" system neces-
investigation proceeds. For example, recent evidence sarily encodes something like an internalized map. The
challenges the notion that the "what" and "where" sys- point is only that many of the same design criteria are
tems are as separate and independent as originally applicable, in particular, the need to represent objects as
thought. 21 Another problem concerns how object and tokens in the representation but to compress their encod-
place information is combined. In many cases, object ing by eliminating most information about their form.
recognition will require the assembly of object parts on Based on the evidence and arguments presented here,
the basis of their spatial relations. It is unclear how this we conjecture that the relatively simple shape specifica-
would be accomplished if the what and where systems tions observed in the prepositional system reveal the
were totally separate. We assume that the solution to extent of detail possible in object descriptions in the
these and other problems will be found in increasingly "where" system. These details go somewhat beyond
complex models of how these systems work and how they "thing here," but not much. In particular, as we have
interdigitate. Recent work suggests a considerable degree seen, the way the system works is not by locating objects
of "cross-talk" and the possibility that the systems under- absolutely but by locating one object in terms of another.
lying the identifying and locating objects are considerably The little detail there is in the system's shape descriptions
overlapping at early levels of the visual system (Van Essen is concentrated on the reference object, which defines the
et al. 1992). For the time being, we draw attention to the space in which the figure is located - and even that is
fact that there is evidence for a what and where bifurca- highly restricted. The geometry of the figure object goes
tion, pointing out what we view as an intriguing parallel to beyond "thing here" only in the small class of cases in
the representation of object and place in language. which the issue is its orientation (along, across, around)
or its distribution through a region (all over, throughout).
3.2.2. "What" and "where" in spatial language. How does
the "what" and "where" distinction bear on our observa- 3.2.3. "What" and "where" as universals. If we are right,
tions about spatial language - and especially on their then the structure of object and place systems should be
relation to the Design of Spatial Representation Hypoth- correlated with universals in languages and this universal
esis? Let us consider for a moment the logic of a represen- spatial-semantic system should serve as an important
tation that separates "what" from "where." What informa- constraint on first language learning. Presumably, the
tion does the "where" system have to encode? At the very universal system would exhibit properties at the grain
least, it must have a space of possible locations and a way needed to distinguish among named objects and named
to mark which ones are occupied. But this is obviously not places across languages. Each of these conclusions awaits
enough. It would not do just for the "what" system to considerable further research, but in the meantime we
know that one is seeing a cat and a dog, and for the would like to suggest several hypotheses and lines of
"where" system to know that positions A and B of the investigation that could provide testing ground for our
visual field are occupied: Is the cat at A and the dog at B, claims.
or vice versa? To keep track of which objects are where, The Design of Spatial Representation Hypothesis gives
there must be a liaison between the two representations rise to three different sorts of predictions. One concerns
(see Van Essen et al. 1992 for neurological perspective). the universal structure of language. If the properties of
A simple way to accomplish this liaison formally is by object and place language are caused in part by the
coindexing or linking the object representations in the structure of nonlinguistic representations of object and
two systems (a similar proposal is suggested by Kosslyn place then we should find broad similarities in the expres-
1990). The "where" system could then encode very rudi- sion of object and place across languages. Specifically, we
mentary representations of the objects being located, should find that object names universally draw on de-
perhaps as simple as "thing here." Such extremely sche- tailed descriptions of object shape, but that terms for
matized objects would place only minimal demands on location draw on quite sparse descriptions of object
information-bearing capacity in the "where" system. shape. Although our discussion has focused on English, a
These schematized objects, however, would also be variety of cross-linguistic studies suggest that the broad
linked to or associated with representations in the "what" outlines of our proposal are correct (Bowerman 1989;
system that encode the objects' detailed shape. In other Cienki 1988; Talmy 1983; Vandeloise 1986).
words, the "where" system can get by with including just We have also noted several examples that might pro-
a little object information, as long as it can link its object vide a challenge to our view (sect. 2.6). These are primar-
tokens to those in the "what" system. 22 ily languages in which a range of spatial verbs restricts
An ordinary-life analogy to the situation in the "where" their application to figure or reference objects whose
system is the conventionalized representation in maps. geometries are more detailed than those found among
What is at issue in a map is how to navigate through a English prepositions. We believe there are at least two
region (or perform the many tasks requiring object loca- possibilities consistent with our hypothesis. One is that

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 235

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language
spatial verbs, which are open class and can clearly express principal vertical axis. In addition, however, children
more than just location, are different from English prepo- learning English should be able to learn two separate
sitions, which are closed class elements dedicated to terms for regions that (1) are or (2) are not in contact with
expressing location only. For example, there are many the object (on and above, respectively), whereas children
spatial verbs in English that concern the motion of some- learning Warlpiri should be able to collapse this distinc-
thing into another object, where the reference object is tion into a single term (Bavin 1990).
defined quite specifically for each verb: These are denom- A final set of questions arises from using our evidence
inal verbs such as to bottle, can, house, and so on. The fact on language to make predictions about neuropsychology.
that these verbs (or the spatial verbs in other languages) We are extremely tentative here, as we recognize the
show quite specific restrictions does not invalidate our complexity and extent of the unsettled issues concerning
hypothesis, which is explicitly concerned with those the "what" and "where" systems. If we are right, however,
closed class forms that are dedicated to expressing then the "where" system should incorporate some shape
location. information, specifically, that having to do with objects as
A second possibility is that different languages do in fact volumes, their axes and orientations, and their distribu-
have sets of elements dedicated to expressing location tion in space (sparse/dense/linear). In addition, it should
only and that these elements do incorporate more detail be capable of encoding the kind of qualitative (nonmetric)
about figure and reference objects than is found in En- spatial relations we have discussed, such as distance and
glish. Bowerman's examples from Tzeltal do so (see sect. (axis-based) direction. We would be delighted to find
2.6), although the degree of detail is still radically less neurons capable of responding to such properties. At
than that captured by object count nouns. The question is higher levels of analysis we would be interested to dis-
not whether any languages exist that include some shape cover, for example, whether damage to the "where"
information in place terms; as we have seen, even English system would have repercussions for the use and under-
prepositions include some shape specification of figure standing of spatial prepositions but not, say, nonspatial
prepositions or other prepositions used purely as gram-
and reference object. Rather, the question is what kind of
matical markers.
detail and how much exists across languages.
These questions clearly require additional cross-
linguistic research.23 One possibility is that there are
4. Summary and conclusions
typologies among languages, with certain languages (like
English) drawing on extremely sparse geometric descrip-
tions and others drawing on some additional degree of We began with the goal of using evidence from language
detail. We would predict, however, that no natural lan- to provide insight into the nature of spatial cognition, in
guage exists whose pure locational elements consistently particular to help understand how we represent objects
draw on object geometries even close to the level of detail and places. Our survey has shown that the domains of
seen in the object naming system. language describing objects and places draw on very
A second set of predictions concerns language learning. different kinds of spatial representations. When objects
Given the proposed structural constraints on object are named as belonging to a category, their descriptions
versus place language, children should come to language appear to draw on rather complex representations of
learning prepared to attend only to certain properties shapes and surfaces. When the same objects play the role
when learning namesforobjects versus names for places. of figure or reference object in a locational expression,
In particular, children should attend to object shape for however, their descriptions appear to be highly sche-
object count nouns but only sparse shape elements (or matized, preserving at most the axial structure of the
none at all) for place words. Empirical evidence is consis- object's principal volume. The spatial relationships that
tent with these predictions. Young children learning ob- these objects engage are similarly sparse, including pri-
ject names attend to shape rather than texture or size (see marily qualitative distinctions of distance and direction.
sect. 1 and Landau 1993). In contrast, there are no reports Thus, there are significant limitations on the kinds of
of children overrestricting figure or reference object descriptions represented by language, and the two do-
shape when using prepositions in spontaneous speech or mains of object and place exhibit quite different con-
when learning novel prepositions (see Landau 1993 for straints.
review). Some of these limits appear to reflect partially a prop-
Note that our hypothesis provides a general framework erty of language design: filtering out of metric informa-
for testing other acquisition patterns. Children should be tion. More important, however, these differences in how
predisposed to learn spatial terms representing quite objects and places are represented may be correlated
sparse figure and reference object geometries and non- with a property of neurological design: a separation of
metric regions - roughly the properties listed in Table 2 spatial cognitive systems into "what" and "where." If our
(subject to revision from cross-linguistic evidence). For conjecture is correct, we have found a bifurcation in the
example, they should resist learning a new preposition expressive power of language that corresponds to a bifur-
meaning "plin," "sprough," or "2 inches away from" the cation in the functional and anatomical systems of the
reference object. On the other hand, there should be a brain. To our knowledge, this is thefirsttime a correlation
reasonable degree of flexibility in exactly how these object has been made in cognitive science between a property of
geometries and regions combine to yield spatial meanings grammar - that is, the kinds of things that count nouns
in any particular language, thereby allowing an important and prepositions standardly express - and descriptions of
role for learning. For example, children learning any nonlinguistic systems represented in particular areas of
language should be sensitive to the reference object's the brain.

236 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language

Without this correlation, language would still provide ship between an object's shape and its silhouette: The sign of the
rich and systematic evidence on the character of spatial curvature of points on the silhouette agrees with the sign of the
cognition, for as we argued in the beginning, anything we Gaussian curvature of the corresponding points on the object.
can talk about we must also be able to represent. What is The visual system is apparently designed to register this corre-
exciting about finding this correlation is that whereas spondence, a fact that enters into our ability to parse objects
(Hoffman & Richards 1984.) We thank Don Hoffman for point-
previous studies have documented the tasks performed
ing this out to us.
by the two visual and neurological systems, linguistic 5. Whether geons are truly invariant over all viewpoint
evidence can now provide a new source of insight into the transformations is a question of current interest in theories of
actual forms of information the systems encode and, more object recognition. This technical matter has little impact on our
generally, into the nature of spatial representation. main point that componential theories of object recognition may
play an important role in our understanding of how objects get
named.
ACKNOWLEDGMENTS
A less detailed version of this paper appeared previously (Jack- 6. "Suitable" similarity metrics will no doubt be complex and
endoff & Landau 1991). We would like to dedicate this version to the result will differ depending on the investigator's approach
Lila Cleitman, who through many years of association has been a and goals. For example, one could imagine a bottom-up ap-
constant source of inspiration, wisdom, and good jokes. proach in which one could ask what kinds of discriminations can
be made among different classes of shapes, or a top-down
This research was supported in part by NSF Grant IRI
approach in which one could ask what role is played by world
88-08286 to Brandeis University, by Social and Behavioral
Sciences Research Grant #12-214 from the March of Dimes knowledge in shape classification (e.g., knowing that four-
Birth Defects Foundation to Barbara Landau, and by NIH legged and two-legged animals fall into different classes). Differ-
Crant #l-ROl-HD28675-01 to Linda Smith, Susan Jones, and ent goals might also give rise to different similarity metrics. For
Barbara Landau. We are grateful to many colleagues for critical example, an investigation of the morphological development of
comments and discussion. These include Julian Hochberg, shape (Thompson 1961) and a study of what objects share the
Michael Leyton, and Melissa Bowerman for detailed comments same "basic level" (Rosch et al. 1976) might give rise to different
on previous drafts of this paper, and Don Hoffman, David classes of shapes. Developing such similarity metrics is clearly
LaBerge, David Murray, and Edgar Zurif for essential com- vital to understanding the interaction of shape perception,
ments and references used in this study. classification, and naming.
7. Assignment of these terms can vary cross-linguistically, as
in Hausa, where nonoriented objects are assumed to "face" in
NOTES the same direction as the speaker, with the front assigned to the
1. Clearly, the converse does not hold: We may be able to side farthest from the speaker. [See also Deregowski: "Real
represent notions that are not so naturally encoded by language. Space and Represented Space" BBS 12(1) 1989.] Note also that
In fact, the selectivity of language is the very wedge we use to the criteria for all of these spatial part terms must be stated fairly
gain insight into spatial representations. We take up the ques- carefully; they interact in curious ways depending on the shape
tion of why language does not encode all aspects of spatial and function of the object. For example, a house normally has a
representation in section 3. front, back, and sides, but no ends, because it is not long and
2. Our assumptions here may be oversimplified. For exam- narrow. A wide but shallow office building may be said to have a
ple, there is good reason to suspect that visual, haptic, and front, a back, and two ends (a left end and a right end); the axis
auditory information will connect with more than one motor defining front-back is orthogonal to that defining ends. There
system, for example, one that guides reaching and another that are therefore no axes left over to define sides. By contrast, a bus
guides locomotion. In addition, there may be cases where visual has afront end, a back (rear) end, and two sides; here the criteria
information is directly translatable into motor commands, with- for front/back and end project into the same horizontal axis and
out being translated into amodal spatial terms. Nevertheless, the orthogonal axis remains to define sides. Finally, a rectangu-
translation into a common format is necessary to explain a large lar carton may be said to have two sides and two ends - but no
variety of tasks, including how language is readily learned and front or back, because its axes are symmetric.
used by individuals whose modality of experience differs from 8. Some readers may protest that this theory of object shape
the norm (e.g., see Landau & Gleitman 1985). appears too complicated and full of special-purpose devices -
3. This is not to deny the complications involved in deciding that the way humans understand objects cannot possibly require
what an object should be called, and how this relates to the all this complexity. In reply, we can only challenge such critics to
object's category. We assume that what determines an object's develop a simpler alternative that both answers the many
name is its membership in some category and, to be sure, concerns of object schematization and object constancy ex-
category membership (for adults, at least) is not always deter- pressed by perceptual theorists and makes all the distinctions
mined by object shape. Other considerations can be important, among objects that we have observed in language. We believe
such as function, or, in the case of living things, descent (Carey that any such theory will contain complexity comparable to ours.
1985; Keil 1989; Murphy &Medin 1988). On the other hand, it is 9. Perceptual ground and reference object do not completely
also clear that object shape is often critical to what an object is overlap in the linguistic representation of place. In some cases,
called. A toy bear may not really belong to the category bear they may be the same, for example, in the sentence "The cat is
(depending on one's theory of the nature of a category), but it on the mat," the mat is the perceptual ground and the reference
does share its name - assigned by the head noun - with the object. In other cases they may be different, as in "The cat is
animate versions in virtue of the similarity in appearance. near the mat"; the mat is the reference object but not the
Whatever complex interactions are involved in assigning perceptual ground. Also the encoded reference object is not
objects to categories, our interest for present purposes is in the always a landmark. Landmarks are typically reference objects,
mapping between spatial cognition and language. Limiting but the converse does not necessarily hold. (A piece of furniture
ourselves to these relationships, we note that shape-based might serve as a reference object but would not generally be
representations are critical to object identification and that considered a landmark.) The exact relationship among percep-
object identification is in turn critical to object naming. Hence tual ground, reference object, and landmark is beyond the scope
our focus on object shape and its role in naming. of this target article. We will focus on reference objects unless
4. Recognition is apparently mediated by the lawful relation- otherwise specified.

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 237

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Landau & Jackendoff: Spatial language

10. In Table 1, compound prepositions are combinations of as we can follow their discussion without competence in Navajo).
words that function grammatically as a single preposition, more About the only case that involves a novel shape descriptor for an
or less parallel to compound nouns, such as garbage man and big object, the issue with which we are most concerned, is binikd, a
top. We do not consider compounds that incorporate open class postposition meaning roughly "passing through a hole," as
materials (such as "on the top 3 inches of). Intransitive preposi- through the eye of a needle. It is interesting that the French
tions, often classified as adverbs in traditional grammar, are verb enfiler implies a similar geometry of something slipping
locational and directional words that occur in all the usual through a narrow opening (Jacques Mehler, personal communi-
grammatical positions for prepositional phrases but need not be cation). Thus, although we do not want to claim that English
followed by a noun phrase. Many of them, e.g., outward and exhausts the spatial relations expressible in language, it does
upstairs, contain a preposition as a constituent. In addition, appear to provide a substantial and representative sampling.
many standard prepositions such as below, nearby, and through 19. Functional explanations of this filtering could take one of
can occur with or without an object. By analogy with the familiar a number of forms. For example, one could propose that people
distinction between transitive and intransitive verbs, then, can manage to communicate effectively while expressing only a
these uses of prepositions with and without objects have been small range of spatial relations, so language has evolved to have
termed transitive and intransitive, respectively (see Jackendoff no more such expressions than necessary. This kind of explana-
1973 for more detail). tion runs into problems in trying to explain why there is a
11. What makes a good reference object will vary enor- proliferation of vocabulary in certain apparently "inessential"
mously with context: In the majority of cases, the reference areas (e.g., color) but not in other apparently "essential" areas
object must be "salient" or "distinctive." For example, in the (e.g., spatial terms). What one needs is some reasonable design
context of a city, a good reference object will be large and stable criteria for what would constitute a plausible function. Along
and perhaps have significant cultural or emotional meaning these lines, Pinker and Bloom (1990) have suggested that by
(Lynch 1960). On the other hand, as Miller and Johnson-Laird representing spatial relationships discretely, one can capture
(1976) point out, in a field of grey dots, something as simple as a causal discontinuities in an efficient manner. For example, one
red dot might make a good reference object just because it is is protected from the rain under a ledge regardless of how far
perceptually salient. Extended to the domain of objects, certain under one is; and one is not so protected if one is not under the
features might function as landmarks if they are visible with ledge, regardless of how far away from the ledge. Such qualita-
peripheral vision and serve to distinguish the target from dis- tive causal facts could underlie the separation of continuous
tractors (Hochberg & Gellman 1975). space into discrete regions such as are encoded in languages.
12. These asymmetries are not confined to the spatial do- Our own thought about functional explanations is that they
main; rather, the existence of "cognitive reference points" surely play some role in the evolution of vocabulary (as evi-
(Rosch 1975) can be shown with color, number, and orientation, denced by the existence of technical or specialist vocabularies),
among others. but one would be hard-pressed to come up with a systematic
13. Going across a bridge is a special case, in that one goes account that explains vocabulary limitations across all domains.
from one end of the bridge to the other. Presumably, this is We are happy, however, to admire such attempts from afar.
motivated by the fact that the bridge itself extends from one side 20. Farah et al. (1988) call this a difference between "visual"
to the other of something else, such as a road or a river. and "spatial" capacities. One of us (R. J.) has a somewhat
14. There are a few exceptions to this overall generalization. different interpretation: that we should think of both as sub-
Nautical terms like port and starboard require a boat as refer- systems of the multimodal spatial capacity. Because all the tasks
ence object. Upstairs and downstairs involve levels in a building tested by Farah et al. were exclusively visual - there were no
(though not necessarily stairs, because one can go upstairs in a haptic or motor tasks - the evidence so far does not distinguish
building using only elevators). And the compounds with -ward the two possible interpretations.
such as homeward and shoreward involve reference to the 21. For example, damage to the hippocampus results in
object named by the initial noun. The meaning of these terms impairment on tasks requiring memory for an object's location
may go outside the spatial relation system proper, involving an (Angeli et al. 1988), consistent with evidence reviewed by
interaction with the object shape system. O'Keefe and Nadel (1978). Yet the hippocampus is one of the
15. The adjectival form perpendicular to places constraints termini of the temporal pathway, which is the hypothesized
similar to senses 1 and 2 of across. Opposite used as a preposi- locus of the object system.
tion ("Bill is opposite Harry") means about the same as sense 3 of 22. Formal indexing does not answer the psychological or
across, except that it leaves unexpressed the object that Bill and neurological question of how the link is effected. But this is
Harry are on opposite sides of. altogether parallel to the familiar problem in language of how
16. In Avrutin's data, the prepositions that take the preposi- multiple representations are psychologically or neurologically
tional case include na ("on") and v ("in, inside"). Those that take connected - for example, what it means neurologically for the
the instrumental case include za ("behind"), nod ("above, over"), representation of the sound of a word to be linked to the
pered ("in front of"), pod ("under"), and mezdulsredi ("between, encoding of its meaning.
among"). Those that take the genitive include u ("at, very close 23. An important methodological question is how one de-
to") and okololvozle ("near, not far from, close to"). The one cides which elements in a language to consider. We have
problematic example he has provided is mezdulsredi ("between, restricted our discussion to prepositions - closed-class elements
among"), which we have analyzed as "interior to" and which - and have essentially left untouched spatial verbs, which are
therefore should take the prepositional case. It will take further open class and seem to have much more latitude in what kinds of
research to decide whether this is truly exceptional or whether geometric and nongeometric elements they can represent. We
our feature analysis must be modified. agree with Talmy (1983) that cross-linguistic investigation
17. "Toward" and "away from" are more restricted than "to" should focus on closed-class elements (whether verb markers,
and "from," in that the region they are constructed from is prepositions, postpositions, etc.) that express spatial relation-
always "at X." There are no expressions *toward on X or *away ships.
from under X, parallel to onto X and/rom under X, for instance.
18. Pinxten et al. (1983) make an exhaustive exploration of
Navajo spatial terms. Though they emphasize how different the
Navajo spatial framework is from that of English, there are few
surprises with respect to the parameters discussed here (insofar

238 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Commentary/ Landau & Jackendoff: Spatial language

Open Peer Commentary bits of the syntactic representations. It is here that a difference
emerges between in front ofand at the thick end of. In front of
is appropriately analyzed as a complex preposition: although
originally syntactically complex, it is now well on the way to
Commentary submitted by the qualified professional readership of this becoming "lexicalized," that is, to becoming a single lexical
journal will be considered for publication in a later issue as Continuing
item. (In the case of behind, the lexicalization process has gone
Commentary on this article. Integrative overviews and syntheses are
especially encouraged. even further: unstressed be- is not generally equated with the
preposition by, except by linguists; hind does not occur as a
noun; and there is no following of.) By contrast, no one would
suggest analyzing at the thick end of as a complex preposition.
From a semantic point of view, lexicalization entails the combi-
There is more to location than prepositions nation of separate concepts into a single (albeit complex) con-
cept. The location specified by (2) is best described therefore as
David C. Bennett involving the reference object "the house" combined with the
Department of Phonetics and Linguistics, School of Oriental and African relational concept "in front of." The location in (3), on the other
Studies, University of London, London WC1H OXG, England hand, is indicated by the reference object "the thick end of the
Electronic mail: d.ben@clus1.ulcc.ac.uk cue" combined with the relational concept "at." The question
Landau & Jackendoff (L & J) present an illuminating account of that now arises is why some strings of words become lexicalized
the categorization of physical objects and of prepositional mean- whereas other, syntactically parallel, strings do not. The rele-
ings. They are right, too, in pointing out that languages typically vant factor would seem to be frequency of occurrence. It is
have tens of thousands of names for physical objects and less presumably because objects are so frequently visualized as
than a hundred spatial prepositions. They are wrong, however, occupying the space adjacent to the front of some other object
when they imply that the rich system we have for describing that the phrase "in front of" has acquired the status of a single
things (the "what" system) contrasts strikingly with an impov- concept. By contrast, it is only rarely that we find ourselves
erished system for describing locations (the "where" system). describing something as being located "at the thick end of"
Descriptions of locations can be indefinitely detailed. A moder- something else. In support of this line of reasoning, it may be
ately complex example is given as the italicized phrase in: noted that in an alternative world with very large numbers of
cone-shaped and wedge-shaped objects, the situation would be
(1) The jack must be positioned beneath the jacking points at different, because the meanings "at the thick end of" and "at the
the outer ends of the cross-member. thin end o f would be far more readily applicable. In this case it
Admittedly, much of the information in this phrase is conveyed might happen that the two strings of words in question would
by noun phrases (the jacking points, the outer ends, the cross- become lexicalized and perhaps give rise to the complex prepo-
member), but that does not affect the fact that the phrase as a sitions thick-end a n d thin-end (e.g., A is thick-end B,C is thin-
whole answers a "where" question. Similarly, in the simpler end D).
sentence: This discussion of lexicalization indicates that frequency of
occurrence is one factor influencing the formation of complex
(2) There's a bike in front of the house, locational concepts and is therefore relevant to the question of
the kinds of geometric information that can be incorporated into
the locational information is given by means of a preposition and the meaning of prepositions. It is a mistake, however, to concen-
a noun phrase. In general, locational expressions incorporate a trate on prepositions to the exclusion of other parts of speech.
"reference object." When L & J write (Introduction, para. 6) Faced with verbs of a language such as Atsugewi, which place
"place naming draws on quite sparse elements of object shape," detailed geometric restrictions on the "figure object" (thing to
they are referring only to the prepositional component of loca- be located), L & J write (sect. 3.2.3, para. 3): "Spatial verbs,
tional phrases. The noun phrase component can convey any which are open class and can clearly express more than just
amount of information about object shape. There is no great location, are different from English prepositions, which are
discrepancy therefore between the number of possible things closed class elements dedicated to expressing location only. . . .
and the number of possible places, or between the resources of our hypothesis. . . is explicitly concerned with those closed
our object naming and place naming systems. And if the as- class forms that are dedicated to expressing location." Yet, given
sumed discrepancy does not exist, it is unnecessary to try to that spatial information is expressed by prepositions (in, inside),
explain it by hypothesizing that "what" expressions and "where" adverbs (nearby, inside), verbs (enter, contain), adjectives (long,
expressions are processed by different parts of the brain. thin), and nouns (top, inside), and given the close relationship
I hasten to add that I do not disagree that the number of between such (nonsynonymous) expressions as inside the jar
geometric features of reference objects, and of located objects, and on the inside of the jar, it seems unlikely that the semantic
encoded by spatial prepositions is very limited. Indeed, I shall representations we derive, in the process of understanding
attempt to throw further light on this issue by mentioning one utterances, are segregated into sections deriving from different
factor not noted by L & J. We approach the matter by comparing parts of speech.
(2) with: In view of the opinions expressed above, it is not surprising
(3) There's a crack at the thick end of the cue. that I attach more importance to the problems for the Landau &
Jackendoff hypothesis, which they acknowledge in the final
It seems reasonable to suggest that the reference object in (3) is paragraph of section 3.2.1, than to the claims they make in
not just "the cue" but "the thick end of the cue" - that is where "a section 4, paragraph 2.
crack" is located (simple location being indicated by the preposi-
tion at). Assigning (2) a parallel semantic structure would entail
treating "the front of the house" as the reference object: "a bike"
is located "at-the-interior-of" (in) the space adjacent to "the front
of the house." Such a semantic representation can certainly be
defended. However, in addition to proposing a semantic repre-
sentation for the whole of (2) and (3), it is relevant to indicate
which bits of the semantic representations correspond to which

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 239

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Commentary/Landau & Jackendoff: Spatial language

Spatial and cognitive vision differentiate at tivity into the cognitive system with stroboscopic induced mo-
low levels, but not in language tion (Bridgeman et al. 1981). A surrounding frame was dis-
placed, creating the illusion that a target had jumped, although
Bruce Bridgeman it remained fixed relative to the subject. After all stimuli were
Program in Experimental Psychology, University of California, Santa Cruz, extinguished, the subject pointed open-loop to the last position
Santa Cruz, CA 95064 of the target. Trials in which the target had seemed to be on the
Electronic mail: bruceb@cats.ucsc.edu left were compared with trials in which it had seemed to be on
the right. Pointing was not significantly different in the two
Vision seems to be a single, unified sense - we see things, kinds of trials, showing that induced motion did not affect
describe them, and interact with them in a seamless continuity. pointing.
On closer analysis, however, the unity of vision splinters into In a second condition of the same experiment, information
specialized subsystems. Characterizing these subsystems has was inserted selectively into the motor system. Each subject
become one of the central concerns of vision research; a central adjusted a real motion of the target, which was jumped in phase
source of confusion has been that different methods yield with the frame, until the target seemed stationary, so that the
different bifurcations of visual function. Landau & Jackendoff (L cognitive system specified a stable target. Subjects nevertheless
& J) contrast the rich and varied vocabulary of object names with pointed in significantly different directions when the target was
the sparse geometric description of visual space, concluding that extinguished in the left or the right positions, showing that the
the difference in language may be related to physiological difference in real target positions was still available to the motor
differences between spatial and object systems that have been system. This is a double dissociation: In one condition the target
identified in monkeys (Schneider 1969; Ungerleider & Mishkin displacement affected only the cognitive system and in the other
1982) and in lesioned humans (Farah et al. 1988). I suggest that it affected only motor behavior.
the situation is a bit more complicated; L & J identify two modes Dissociation of cognitive and motor function has also been
of linguistic function in vision, whereas some other contrasts demonstrated by giving the cognitive and motor systems oppo-
identify one system that is accessible to language and another site signals at the same time. Again, the experiment involved
that is not. The two linguistic modes may be part of a broader stroboscopic induced motion; a target jumped in the same
cognitive system. direction as a frame but not far enough to cancel the induced
One requirement for differentiating two modes of vision is the motion. The target appeared to jump in the direction opposite
study of normal humans rather than lesioned humans or other the frame when it actually jumped in the same direction.
primates. There is now extensive evidence that two visual Saccadic eye movements followed the veridical direction even
modes are represented in the normal human brain and that they though subjects perceived stroboscopic motion in the opposite
follow different rules (Bridgeman 1981; 1986; Paillard 1987). direction (Wong & Mack 1981). If a delay in responding was
Following the nomenclature of Paillard (1987), these will be required, however, eye movements followed the perceptual
called cognitive and sensorimotor, respectively. The two sys- illusion, implying that the motor system has no memory and
tems have been applied in many contexts and given many must import information from the cognitive system after a delay.
different names. It is not yet clear whether all of the names refer All of these experiments involve motion or displacement,
to the same neurological machinery, for some of them address leaving open the possibility that the dissociations are related to a
different aspects of behavior. All, however, share a common confounding of motion and position rather than to a representa-
distinction between a uniquely spatial, generally unconscious, tion of visual space per se. A new method tests dissociation of
motor-oriented system, and a more symbolic system whose cognitive and motor function without motion of the eye or the
contents are at least partially conscious, forming the basis for stimuli at any time. The dissociation is based on the Roelofs
perception. effect, a tendency to misperceive the position of a target pre-
In our first experiment on this problem (Bridgeman et al. sented against an off-center background. Subjects were always
1979), two conflicting observations (saccadic suppression of biased by a Roelofs effect in judging which of five possible
perceived target position on one hand and accurate reaching to a positions of a target was presented; an off-center surrounding
target following a saccade on the other) were combined by frame caused judgments to deviate to the opposite side. For half
asking subjects to point to the position of a target that had been of the subjects, however, the frame had no effect on pointing (a
displaced during a saccadic eye movement. Subjects were also measure of the sensorimotor system). Again, if the sensorimotor
asked whether the target had been displaced. Extinguishing the system has no memory, some subjects may have switched to the
target and preventing the subjects from viewing their hands cognitive system to point to the target position after it was
(open-loop pointing) guaranteed that only internally stored extinguished; indeed, with a long enough delay we could force
spatial information could be used for pointing. On some trials all subjects to show a Roelofs effect even in pointing (Bridgeman
the displacement was detected, whereas on others it went 1991).
undetected, but pointing was accurate whether or not the Taken together, these psychophysical experiments show that
displacement was detected. Pelisson et al. (1986) found a similar cognitive and spatial visual systems can be distinguished on a
result. This result implies that visuomotor localization is un- lower level than that of Landau & Jackendoff, a level that
affected by the perceptual detectability of target position. A differentiates linguistic from nonlinguistic coding. Their fas-
further test of this possibility was a two-alternative forced-choice cinating contrasts between object and spatial language may
measure of saccadic suppression of displacement. Even this simply reflect the physical fact that there are a lot of objects, but
criterion-free measure showed no information about displace- only a few Euclidean geometric relations.
ment to be available to the cognitive system under conditions in
which pointing was affected (Bridgeman & Stark 1979).
A more rigorous way to separate cognitive and motor systems
is to stimulate only the motor system in one condition and only The role of cerebral lateralization in
the cognitive system in another. This was done with induced expression of spatial cognition
motion, the apparent motion of a target when its background
moves. We know that induced motion affects the cognitive Halle D. Brown
system because we experience the effect and subjects can make Department of Psychology, Harvard University, Cambridge, MA 02138
verbal judgments of it. The above experiments implied, how- Electronic mail: hb@huelings.harvard.edu
ever, that information used for pointing might come from Although the analysis of linguistic descriptions given by Landau
sources unavailable to perception. We inserted a signal selec- & Jackendoff (L & J) is an innovative approach to the study of

240 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Commentary/Landau & Jackendoff: Spatial language

spatial cognition, I would suggest that only half of the issue has and test, but only in the right hemisphere. That is, although
been considered. In particular, their analysis has informed us "w-o-r-d" is read as the same object type whether it is printed in
about the geometric properties of objects and spatial relations upper or lower case letters, the right hemisphere apparently
that we know to be the specialty of the left cerebral hemisphere. distinguishes tokens (e.g., "WORD" vs. "word") in memory. The
Recent studies, however, suggest that some aspects of spatial physical features or visual details of objects are needed to
cognition are performed by the right hemisphere. These are as distinguish tokens. Why do we not have nouns to express the
important as those performed by the left hemisphere but are not geometric properties of objects that distinguish tokens? Perhaps
readily evident from an analysis of linguistic expressions. The because this analysis of featural variations is carried out more
contributions of both hemispheres to location and object repre- effectively in the right hemisphere, which does not have lan-
sentation will be considered in this commentary. guage output capabilities. The output from the token system
We can certainly express spatial representations verbally, but must cross the corpus callosum into the left hemisphere, where
we also express them in nonverbal behavior. For example, to a discrimination between tokens and types would no longer be
reach for an object, we need to have computed its location in distinguished in the system generating output for "what"
space. Both behavioral studies with humans (Kosslyn et al. 1989) descriptions.
and computer simulation models (Kosslyn et al. 1992) have The alternative hypotheses suggested by L & J regarding
provided evidence for two distinct types of spatial-relation differential linguistic expression of object properties and spatial
representations. Categorical spatial-relation representations, locations can now be reevaluated in light of this consideration of
used to judge whether a dot is above or below a line, can be cerebral lateralization. Their "Design of Language Hypothesis"
distinguished from coordinate spatial-relation representations, may be correct precisely because language-processing systems
needed when judging whether a dot is within 3 millimeters of a in the left hemisphere have access only to generalized object and
line. Kosslyn and his colleagues have found that categorical location representations. Note, however, that language does not
spatial relations are computed more efficiently in the left hemi- "filter" otherwise precise spatial relations; rather, spatial rela-
sphere, whereas coordinate spatial relations are computed more tions computed more efficiently in the left hemisphere serve a
efficiently in the right hemisphere. Only the categorical type of different purpose from the precise, coordinate relations com-
spatial relation underlies the prepositional descriptions listed in puted more efficiently in the right hemisphere. Their "Design of
L & J's Table 2. Spatial Representation Hypothesis" can be better characterized
Coordinate spatial relations are equally important for the in terms of processing in the "what" and "where" pathways in the
study of spatial cognition, yet they are not expressed in as left hemisphere, with computations of object "types" underlying
straightforward a manner with prepositions as categorical spatial count nouns and computations of categorical spatial relations
relations are. Kosslyn and his colleagues have argued that underlying prepositional descriptions.
coordinate spatial-relation representations are most useful to The study of spatial cognition must clearly include an analysis
guide action (see Kosslyn & Koenig 1992). We need to know of cerebral lateralization of function and should avoid research
precisely where the cup is located to guide the trajectory of our tools biased to examine the specializations of only one hemi-
hand toward the handle; a verbal description of the movement sphere. Rather, our tools should be general enough to examine
would be superfluous. The locations of objects in space and of and ultimately characterize the specializations of both hemi-
parts relative to each other may be encoded procedurally rather spheres. Landau & Jackendoff have expanded our understand-
than descriptively, as a set of coordinates to guide motor ing of representations of spatial relations in two different sys-
programming. tems ("what" vs. "where"), but the implications of their analysis
In their analysis of geometric properties conveyed by object are limited to one domain (the left hemisphere). Converging
nouns, L & J have again given us only half the story. Object evidence from other methodologies is needed to complete our
representations that have associated count noun names focus understanding of the range of spatial information humans can
mainly on invariant features of objects. These geometric proper- express.
ties allow us to access the same object name even when the
object is presented in varying forms (e.g., words written in ACKNOWLEDGMENTS
different type fonts are read the same way). Yet we also have the I would like to thank Stephen M. Kosslyn, Chad J. Marsolek, and Kevin
ability to distinguish particular exemplars of a type of object. N. Ochsner for their helpful comments on this manuscript.
Recent studies of cerebral lateralization have indicated that two
fundamentally different object representations can be com-
puted, at least in the visual system. An "abstract-visual-form"
representation is activated by different instances of the same Frames of reference in the spatial
type of object; as long as the inputs have the same invariant representation system
relations among their parts, the same abstract object output will
be computed. Marsolek (1992) has found that abstract-visual- David J. Bryant
form representations are processed more efficiently in the left Department of Psychology, Northeastern University, Boston, MA 02115
hemisphere. On the other hand, form-specific representations Electronic mail: bryant@northeastern.edu
are activated by different instances of the same abstract type of Central to Landau & Jackendoff's (L & J's) argument is the
object. Marsolek et al. (1992) found that the right cerebral assumption that a single spatial representation system underlies
hemisphere distinguishes more effectively between different perception and language. I agree with L & J that human
form-specific representations of the same abstract type of object cognition uses separate "what" and "where" systems, my pur-
than the left hemisphere. pose is not to criticize their analysis. Instead, I will briefly
Why is there no distinction in language for types versus discuss additional evidence for a common perceptual/linguistic
tokens that maps onto this hemispheric difference in the brain, spatial representation system, examine some properties of
as there is with object nouns ("what") versus prepositions the spatial system, and suggest one way in which the analysis of
("where"), which map onto the difference between ventral and spatial language might be extended.
dorsal system processing in the brain? Although object types can Mental models of space. A large body of research has deter-
be distinguished by different spatial relations within and be- mined that people represent texts in situational mental models
tween parts, object tokens can be distinguished more precisely. rather than as a propositional record. It is relevant to the present
For example, in the study by Marsolek et al. (1992), implicit discussion that mental models preserve explicit and inferred
memory for words, measured as stem-completion priming, was geometric relations between objects in described situations. In
reduced when words were presented in different cases at study particular, research has demonstrated that mental models pre-

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 241

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Commentary/Landau & Jackendoff: Spatial language

serve spatial features such as distance (e.g., Glenberg et al. sitions generally do not specify a frame of reference (Retz-
1987; Morrow et al. 1987) and relative direction (Bryant et al. Schmidt 1988). At most, prepositions are constrained by the
1992; Franklin & Tversky 1990). Moreover, these features guide alignment of a major axis of the figure or reference object (L & J).
the retrieval of information, determining the accessibility of Thus, I can say "the computer is behind the filing cabinet," and
objects in mental models. When analyzing a verbal description, mean that the computer occupies a position to the filing cabinet's
readers normally extract information about objects and their intrinsic back. On the other hand, I could say "the computer is
locations and represent it in a form that is spatial and perceptual behind the filing cabinet," and mean that both objects are to my
(Johnson-Laird 1983, pp. 156-62). front, but the computer is farther away. The word behind does
The close correspondence between mental spatial models and not specify whether I am using an egocentric or an allocentric
perception can be seen in a class of mental models called spatial frame of reference. Likewise, most spatial prepositions can be
frameworks (Franklin & Tversky 1990). A spatial framework used in both of these frames of reference (exceptions seem to be
organizes objects and their locations within the frame of refer- words like "through" and "inside," which require an object-
ence created by an individual's three body axes (head/feet, centered interpretation).
front/back, left/right), in the same way that body axes are used Although spatial prepositions do not specify a frame of refer-
in perception to locate objects (Gallistel 1990, pp. 106-18; ence, language at the discourse level does. Levelt (1984) among
Hintzman et al. 1981). Spatial frameworks also render certain others has distinguished between the deictic system of spatial
spatial locations more accessible at retrieval, depending on the reference, in which spatial prepositions are interpreted relative
perceptual and physical asymmetries of body axes and the to one's own egocentric origin and body axes, and the intrinsic
perspective of the observer (Bryant et al. 1992). Recent studies system, in which spatial terms are interpreted with respect to
have demonstrated that people use the same kinds of spatial external axes of a referent object or the environment itself.
frameworks during visual perception of scenes (Bryant & These two systems map onto egocentric and allocentric coordi-
Tversky 1991; Logan 1991), indicating that representations of nate frames respectively, but they coexist at the level of spatial
perceived and described environments have equivalent struc- prepositions. To say that prepositions are interpreted with
ture. respect to egocentric or allocentric referents is crucial because
As further evidence of a common perceptual/linguistic spatial these spatial terms have meaning in both systems. Thus, other
system, peoples spatial representations of descriptions interact cues are needed to specify a frame of reference in language, cues
with perceptual representations. Easton and Bentzen (1987) that can only be included at a level beyond the single word (see
found that performance on a spatial finger-maze task was im- Retz-Schmidt 1988). The deictic system, for example, must
paired when subjects simultaneously verified verbal spatial refer to the speaker's body sides and establish this context in
statements but not when they verified nonspatial statements. discourse for the interpretation of spatial terms. Likewise, the
Similarly, performing a visuospatial tracking task interferes with intrinsic system must establish a set of environmental axes to
a person's ability to form a coherent spatial model from a verbal distinguish between the possible meanings of spatial preposi-
description (Oakhill & Johnson-Laird 1984). In both cases, tions. These discourse-level linguistic systems, like preposi-
interpreting verbal spatial directions and performing a spatial tions, are influenced by the way the spatial system represents
task seem to compete for the resources of the same spatial the perceptual world. For example, Levelt (1984) has noted that
system. perceptual and spatial features such as the orientation of objects,
The spatial representation system. All this reinforces L & J's the world's gravitational axis, and the posture of the observer
assumption that a common spatial system underlies spatial limit what spatial representation can be derived from a verbal
cognition in perception and language. To explore further the description.
geometric properties of spatial language and cognition, how- Conclusion. The purpose of my commentary has been to flesh
ever, it is worth wondering what this spatial representation out the idea of a common spatial representation system as it has
system is like. The purpose of the spatial system is to represent guided the evolution of spatial language. In addition to L & J's
the layout of objects in the environment. To do this, it must use argument for separate "what" and "where" systems, I would
some frame of reference that establishes three spatial axes. The suggest that other vital spatial concepts are embodied in the
spatial system can then determine each object's position along spatial system and constrain language at some level. In particu-
each dimension in the resulting coordinate space. Humans lar, to understand how we represent space, we need to consider
commonly use the egocentric frame of reference, defined by the how frames of reference are used by the spatial system in
three body axes (head/feet, front/back, left/right), and the perception and language.
allocentric frame of reference, defined by orthogonal axes set
outside the observer and anchored on prominent landmarks or
aligned with environmental features. People may also use
object-centered reference frames (see Shepard & Hurwitz Generative versus nongenerative thought
1984).
The allocentric and egocentric reference frames are crucial to Michael C. Corballis
spatial perception. We need allocentric cognitive maps to navi- Department of Psychology, University of Auckland, Private Bag 92019,
Auckland, New Zealand
gate in the world and the egocentric frame of reference to guide
Electronic mail: mikec@ccu1.aukuni.ac.nz
our actions in immediate space. Also, because we can perceive
the world only from our own position, we can create allocentric I should like to argue that many of the distinctive characteristics
representations only through transformations of egocentric rep- of spatial language and spatial cognition described by Landau &
resentations (Gallistel 1990, pp. 106-9). People constantly up- Jackendoff (L & J) are uniquely human and primarily a function
date allocentric maps from egocentric perception and likewise of the left cerebral hemisphere.
direct egocentric interaction with the environment on the basis Like language, human manufacture is characterized by an
of allocentric maps. It is not surprising, then, that both frames of open-endedness that seems to be unique to our species; there
reference are available for the representation of linguistically seems no limit to the number and variety of different objects
described space. Readers create allocentric and egocentric spa- humans can create. Animals do use and make tools (Beck 1980),
tial models to represent described environments and update but in a "one-off" rather than a generative fashion. Moreover,
both types of representation regardless of the perspective of the the procedures by which humans make objects are similar to
description (Taylor & Tversky 1992). those by which they generate prepositional language, and in-
Frames of reference and spatial language. Although frame of volve some of the same principles (Corballis 1991; Greenfield
reference is clearly necessary to represent space, spatial prepo- 1991). Just as phonemes are combined to form words, words to

242 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Commentary /Landau & Jackendoff: Spatial language

form phrases, phrases to form sentences, and so on, so standard evidence that the representation of partwise or manufactured
parts are combined in hierarchical and recursive fashion to objects may be primarily left-hemispheric, whereas the repre-
create complex objects. Most of those parts have names - sentation of more naturalistic ones is more bilateral or right-
wheels, axles, handles, knobs, shafts, blades, nuts, bolts, hemispheric (Farah 1991; Warrington & McCarthy 1987; War-
screws, nails, bulbs, sockets, sprockets, switches, transistors, rington & Shallice 1984; but see also Farah & McClelland 1991).
and so on (and on). However, I do not think that the distinction between the GADly
This combinatorial, recursive mode may characterize not only and the unGADly coincides precisely with that between the
the language and objects we produce but also many of our manufactured and the naturalistic. GADly principles have no
cognitive representations. Biederman (1987) drew attention to doubt invaded our representations of animals or plants, cul-
the parallel between his recognition-by-components theory of minating in the sciences of zoology and botany, and there are few
object representation and the combinatorial nature of language. more GADly sights than the models of molecules that lie around
Although his "geons" are not directly or obviously related to chemistry laboratories. Conversely, the craft of the sculptor,
manufacture, and indeed are readily applied to nonmanufac- potter, or painter may be based on holistic rather than combina-
tured objects (such as a penguin, to take one of his own torial principles.
examples), I think it reasonable to suppose that both may be I suggest that the contrast between the language of objects
linked to the same underlying, distinctively human cognitive and the language of places is neither exclusively a property of
system. Elsewhere, I have referred to this system as a genera- language nor of spatial cognition, but was bestowed by GAD.
tive assembling device or GAD (Corballis 1991).
GAD may be distinguished from an (evolutionarily) much
older system that represents things in a nonsymbolic way.
Because this older system is a direct product of gradual selective
adaptation, it is much more finely tuned to the subtle nuances of Are spatial representations flatfish?
the natural environment. I do not think that a geon-based theory
could ever account for our ability to recognize the faces of the J. B. Deregowski
individuals we know; geons or generalized cones might capture Department of Psychology, University of Aberdeen, King's College, Old
Aberdeen AB9 2UB, Scotland
the prototypical shape of the human body (Marr 1982), but could
Electronic mail: psyO22@aberdeen.ac.uk
not easily represent the idiosyncrasies of shape or gait that single
us out as individuals. Similarly, as L & J point out, language Landau and Jackendoff(L & J) find a great disparity between the
permits only crude statements about the locations of things in way the English language describes objects and the way it
space, yet our behavior in reaching or navigation displays a fine describes spatial relationships amongst objects. The latter de-
spatial tuning that belies ourclumsy words. As a rough analogy, I scriptions, they observe, are markedly more rudimentary.
like to think of the GADly system as comparable to a Lego Although naming an object acknowledges its spatial pres-
construction set, whereas the unGADly one might be compared ence, it does not necessarily imply that the object's three-
to Play-Doh (Corballis 1992), or for those who like to play with dimensionality is a dominant characteristic of its mental repre-
adult toys, to a digital computer and a connectionist network, sentation. It may be that this characteristic, although important
respectively. in the context of immediate perceptual activity, is less important
I have speculated elsewhere on the possible time course of the for encoding and retention. Perception obviously has to take
evolution of GAD (Corballis 1991; 1992). Here, however, it is into account the spatial attributes of objects and their mutual
pertinent to ask why it was adaptive to superimpose a new spatial relationships. This is necessary for survival; but it can be
representational system on the old. Despite the open- questioned, perhaps, whether the encoding of objects in mem-
endedness afforded by GAD, I do not think the answer hinges ory by reference to their three-dimensional characteristics is of
on memory capacity - we can, after all, recognize hundreds if equal importance in this context. Recognition of objects may not
not thousands of faces, and many other natural objects that do rely primarily on their three-dimensionality, which may consti-
not readily lend themselves to geonetric (or geonological?) tute a secondary factor, present but not salient, akin to, say,
principles. A network with as many connections as a human surface colour; and, like colour, it may in certain circumstances
brain would have enormous storage capacity without having to gain greater importance than it normally has; apart from such
resort to symbols or recursive principles. circumstances, however, its importance is small. The hypothesis
The advantage of GAD may lie partly in communication. A that three-dimensionality of objects is not of prime importance
channel transmitting categorical distinctions is intrinsically less for their encoding will be examined using a source of evidence
noisy than one transmitting metric distinctions. L & J draw largely ignored by L & J: pictures, that is, images on flattish
attention to the fact that even in sign language, where the surfaces.
potential for analog mapping is high, a continuous dimension The axes that pass through Marr's (1982) cylinders, threading
like speed is reduced to two or three categories - a reflection no them together to form a figure (Figure 2A in L & J's target
doubt of the limit on our ability to make absolute judgments article), themselves constitute an easily recognisable pin-figure.
(Miller 1956). A more important consideration, though, may This mannequin and its perceptual kith and kin have been much
have been the need for improvisation. The unGADly system is used by artists for millenia. A band of pin-man hunters is shown
the product of slow evolutionary adaptation, and as such was chasing deer in a Mesolithic picture in Cueva de Ios Caballos
poorly equipped to deal with the rapid changes in environment and the chase continues in numerous Bushman shelters of South
that we hominids inflicted on ourselves, partly through persis- Africa (see, e.g., Pager 1972). This type of portrayal was readily
tent emigration to different terrains, but more conspicuously accepted by illiterate Mekan (Me'en) living in a pictureless
through the manufacture of objects, such as dwellings, tools, culture (Deregowski et al. 1972) and is universally popular with
clothes, and so forth. Arguably, only a recursive geon-like children. It is unlikely, therefore, that the use of pin-figures calls
system could keep up with cultural and manufacturing change, for great pictorial sophistication. This is confirmed by data
much of which occurs within an individual's life span. obtained by Fortes (1940; 1981; Deregowski 1978), which show
It may please the more theologically inclined to know that that men who have never drawn before use pin-figures quite
GAD seems to reside primarily in the left cerebral hemisphere. spontaneously when requested to draw. [See also Deregowski:
We all know this to be true of language, but Kosslyn et al. (1989) "Real Space and Represented Space: Cross-Cultural Perspec-
have also shown that in making judgments about the relative tives" BBS 12(1) 1989.]
locations of objects, the left hemisphere seems to rely on a The notable feature of pin-figures is that they readily convey
categorical code, the right on a metric one. There is also their meaning without any attempt to show directly that the

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 243

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Commentary/Landau & Jackendoff: Spatial language

depicted objects are three-dimensional. Silhouettes and outline Causal models of spatial categories
figures (which can be thought of as unfilled silhouettes) also
represent objects without any direct hint of depth. Unlike pin- Jacob Feldman
figures, these figures are constructed by portraying selected Center for Cognitive Science, Rutgers University, New Brunswick, NJ
contours of a solid's surfaces. These "typical contours" have been 08903
found to be of great importance in children's drawings as well as Electronic mail: jacob@ruccs.rutgers.edu
in drawings of certain artistic schools (Deregowski 1990a; Landau & Jackendoff "s (L & J's) discussion of the verbal corre-
Dziurawiec & Deregowski 1992). The use of silhouettes in road lates of spatial categories is a welcome addition to research on
signs is likewise evidence of the efficacy of this type of pictorial categorization phenomena, filling out the sometimes vague
communication (Deregowski 1990c). territory between perceptual models and verbal conceptual
Pictures that have no illusory element of depth, in which the categories. It dovetails nicely, for example, with my own at-
depicted objects are seen as flat (although the portrayals are tempt to classify spatial categories on purely perception-
readily recognisable as showing three-dimensional objects) have theoretic grounds, with the added benefit of cross-fertilization
been termed epitomic (Conley 1985; Deregowski 1980; 1990b; from linguistic (or at least lexical) concerns. L & J's enumeration
Parker & Deregowski 1990). Pin-figures, silhouettes, and out- of the possible meanings of prepositions, in particular, has an
line drawings clearly belong to this category. This has implica- impressive aura of completeness. As they argue, these simple
tions for L & J's analysis because it shows that the recognition of spatial categories are apparently just those that are so percep-
depicted objects proceeds quite well in the absence of percep- tually salient and have such general utility in our descriptions of
tual cues as to their three-dimensionality; it also raises a ques- the relative positions of things that we are actually willing to
tion: If solids can be adequately depicted by treating indications assign them each fixed words - closed-class words, no less.
of their three-dimensionality as distinctly secondary, is it neces- At an underlying formal level, though, categories of structure
sary to postulate that mental representations must treat such and categories of position are more intimately related than L & J
indications otherwise? are willing to let on, notwithstanding the differences they
Chwistek (1960; see Parker & Deregowski 1990) observed discover in the way they are eventually expressed - in fact
that there is a type of figure that represents objects as they are making these differences all the more interesting. My own
and as the artist knows them to be. A uniformly red vase, say, is account of these types of categories, though differing in motiva-
painted in such a picture as uniformly red without any reference tion, methodology, and scope, results in a list of "natural"
to shadows or reflection caused by the ambience in which it is categories that L & J would find quite familiar. Hence, I think
placed and the direction of falling light, or to the artist's experi- my account tends to reinforce their position, laying a somewhat
ence of the object, or indeed to the artist's mental condition. firmer mathematical foundation under the sometimes vague
Such a picture shows the vase so that its typical contour appears conceptual arguments from which they get their enumerations.
in the picture's plane. In particular, my account suggests a justification, on formal
If the line of argument followed above is valid, then spatial epistemological grounds, of this amazingly constant list of ubiq-
qualities are not central to the encoding of objects. Many objects uitous spatial categories. I will get back to that account below.
can be readily encoded, given names, or otherwise classified "What" Is more complex than "where"? L & J express surprise
purely on the basis of their encoding attributes: axial arrays or that encoding shape turns out to be so much more complex than
typical contours. The encoding attributes, however, are not encoding (relative) position. Their surprise seems puzzling,
equally readily derivable from all objects in the environment however, when one considers that structure is in fact intrin-
although they are easily derivable from a large majority of sically more complex than location. The position something
objects present when languages were formed. Most animals, for occupies, after all, is at most a three-dimensional entity (because
example, have readily detectable, typical, idiosyncratic con- space is 3-D); what complexity there is comes from the compli-
tours that generally run along the animals' spine so that the side cated ways in which coordinate systems for encoding a 3-D
view of an animal is the most typical one (but see Dziurawiec & position can be affixed to something, namely, to a gravity or
Deregowski 1992). In contrast, man-made objects often lack viewpoint vector, or to vectors defined by aspects of some
such distinctiveness of shape (there is little difference between object's shape (above a surface, near a vertex, along an axis, and
the essential shape of a television set and a packet of cigarettes), so forth).
although for reasons of commerce attempts are often made to The shape of an object, on the other hand, is a very high-
introduce shape distinctions (e.g., in the shape of containers for dimensional construct; just how high-dimensional depends on
liquid, where none is needed functionally). one's model of shape. (This is in fact why Rueckl et al. [1988] feel
The argument is therefore that spatial qualities are not as certain that the discrepancy they find between the computa-
important in the encoding of objects, as would appear to be the tional difficulty of "what" and "where" would only increase with
case prima facie. The disparity between descriptions of objects more realistic object models. More realistic locational models
and the description of spatial relationships is therefore not as never get more complex than three dimensions - only object
great as L & J suggest, simply because the naming of objects models do.) Intrinsically, that is, with a maximally dumb model,
does not imply that their spatial qualities are of prime impor- structural models could be infinite-dimensional, if every nuance
tance for mental representation. of shape were taken to be a plausible, category distinction. What
It may also be worth observing that many terms used to is impressive about, say, Marr's (1982) generalized cone theory is
describe spatial qualities of objects are those that are sensu that shape could be meaningfully captured with so few dimen-
stricto applicable to flat objects. Boxes, for example, are said to sions, that shape could be reduced to such a simple description
be square and no distinction is made between the roundness of without losing much descriptive power vis-a-vis the categories of
an orange and the roundness of a coin. This usage is not confined things actually extant in the world. The power of a shape theory
to the hoi polloi but is found in psychological writings as well. is in constraining the useful dimensionality of shape without
Greenfield et al. (1967) are agreed, for example, that an orange giving up too many desirable distinctions among classes of
and an alarm clock are round. This observation, by suggesting things. But you can never get it down to as low as three
that the third dimension is not of prime importance in linguistic dimensions, like position - even as simple a class of geometric
descriptions, strengthens the notion of the relative unimpor- objects as 4-geons has four intrinsic shape dimensions. (See
tance of the third dimension in the mental representation of Feldman [1991] or Kendall [1989] for an indication of how
objects. complex this problem can get when the purely mathematical
aspects are investigated more deeply.) Hence, we would always
expect the linguistic system for encoding shape to be larger and

244 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Commentary/Landau & Jackendoff: Spatial language

richer than that encoding location; again, the surprise is that a mathematically as just a translation along an edge (say), with
finite but useful representation of shape is possible at all. translations away from the edge measuring zero and hence
Good categories as cues to causal structure. The more diffi- algebraically dropping out. Enumerating all such models we get
cult question, though, is: What elevates the particular set of a lattice (a kind of hierarchy; see Jepson & Richards 1992) of
relative position categories that people evidently find natural subcategories of location, each of which putatively has a privi-
(and that they express in prepositions) above the host of unnatu- leged status as a cue to a distinct causal model. Because each
ral but mathematically similar alternative categories? L & J causal model ends up corresponding to a distinct formal object,
repeatedly suggest that relative-location categories are simply this scheme attaches a very literal definition to the idea of
coarser than pure metric information - that the "bins "of relative qualitative equivalence with respect to causality. "Meaning"
position picked out by prepositions are distinguished simply by thus accrues to the relative location of object and referent, in a
being big, and borders between them, by being blurry. It is easy way that makes the lexicalization Landau & Jackendoff have
to see that it is not quite so simple. For example, the regions discovered seem particularly intriguing.
"above" and "below" the ground are separated by a surface that is
very fine compared to the size of the two regions. The region
"near" or "at" a point is infinitely smaller than the complemen-
tary region (the rest of the universe). Perhaps most interesting
(for reasons elaborated below), the region "in line with" or On places, prepositions and other relations
"along" an axis or an edge picks out a subspace of lower dimen-
sion than the 3-D space in which it is embedded. Conversely, a Angela D. Friederici
coarse metric grid does not automatically produce natural spatial Cognitive Science Lab, Berlin Institute for Psychology, Free University
concepts: "Less than 100 meters away" and "less than 200 meters Berlin, Berlin, Germany
Electronic mail: friederici@zedat.fu-berlin.dbp.de
away" are both lexicalized equally often, namely, never. Rather,
the distinction between good and bad spatial concepts has a Landau & Jackendoff (L & J) propose a theory in which the
subtle structure. The outlines of this structure are probably best linguistic and conceptual systems are viewed as isomorphic and
looked for, as Marr (1982) famously prescribed, in a consider- sharing a common neural basis - at least for spatial aspects. In
ation of the essential constraints governing a useful representa- the following, I will focus on the isomorphism assumption and
tion of (in this case) the locations of things. leave the neuropsychological claim undiscussed.
In particular, consider the following general constraint on Concerning the isomorphism assumption, L & J claim in
perceptually natural concepts about relative location. Given the particular that the linguistic and the conceptual representations
regularities extant in the particular world being described, each of objects and places are identical with respect to the (geometri-
such concept should contain a set of relative positions, all of cal) aspects they encode. Representations of objects lin-
which are qualitatively interchangeable with respect to the guistically encoded in count nouns are rich in detail concerning
causal relationship between object and referent. For example, the object's shape. Places are linguistically encoded in preposi-
the objects "in" the refrigerator tend to be perishable food items tions, and these represent information about the object only
that were all intentionally placed there for the same reason; they sparsely. The evidence L & J provide in support of this claim is
differ systematically from objects "outside of" the refrigerator. sparse itself, however.
Similarly, to take a more violent example, the objects "along" the Linguistically seen, prepositions encode spatial relations (a
line of a rifle sight have in common with each other a particular definition used in sect. 2.1) between objects, between objects
potential causal relationship with the rifle. That is, the preposi- and places/regions (e.g., the tall building at Washington
tions that capture these spatial relationships are doing some- Square), as well as between places/regions (e.g., Washington
thing more than simply encoding them, in the sense of reducing Square is close to Kennedy Square) but they do not encode
image data to a more compact representation; rather, they are places/regions as such. Because there is only a small number of
particularly tuned to encapsulating plausible hypotheses about possible spatial relations, the number of linguistic elements that
causal stories. encode these relations is limited. L & J use the fact that the
It turns out that the above aim can be satisfied, more or less, in number of nouns is much larger than the number of prepositions
a formal scheme such as the one I described briefly in Feldman as one of the arguments that objects and places are not only
(1992). Leaving the details aside, we can treat a spatial position represented differently linguistically, but that this difference
as the result of a sum of translation operations in some coordi- reflects a contrast between objects and places at a more general
nate frame, in much the same way that Leyton (1992) treats cognitive level. The fact that a language contains more nouns
objects as the result of some sequence of group-theoretic opera- than prepositions is not taken as a reflection of the few existing
tions, and for much the same reason: The position of an object, spatial relations that must be encoded, because the number of
with respect to the origin of the coordinate frame, is then seen as prepositions could be larger in principle, they claim, if preposi-
having a causal history behind it. (That is the sense in which tions were to encode properties of the objects whose relation
place is a special case of structure: Structure can be modeled as they describe. But why should a linguistic system do this? If
the result of a sequence of arbitrary generative operations, but prepositions are considered to encode relations similar to formal
with position it is always translation.) The coordinate frame is in logical relations (e.g., <, >, =), the linguistic sign to encode
turn defined in terms of critical values (right angles, equal such a relation can be used most efficiently if it abstracts from
lengths, etc.) of commonplace structural parameters (orienta- the parts whose relation it encodes. A very general principle for
tion, length, etc.) that are liable to have causal significance the structure of relational expressions could be: Specify the
because of the regularities of the physical, mechanical, and parts in any necessary detail, but do not encode these details in
biological laws governing our environment. These critical values the element expressing the relation itself.
tend, over and over again, to mark critical distinctions between L & J use a lot of space in their target article to list linguistic
regions of configuration spaces of objects in the environment - observations (and experimental evidence) in support of their
that is, to mark boundaries between causally interchangeable claim. Although their final statement is that a linguistic analysis
objects. would be sufficient to model the nature of spatial representa-
The critical formal trick is to represent location as a sum of tions, they do not use this approach themselves. Rather than
transverse translations in some coordinate frame; then discrete staying within the linguistic domain, they switch from the
subcategories correspond to discrete, algebraically distinct sub- linguistic level to the cognitive-psychological level and back,
sets of the full inventory of translations. For example, one of the just as they need it. The argumentation, therefore, becomes
models would be the set of positions that can be expressed circular. Given the isomorphism assumption, this might be

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 245

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Commentary/Landau & Jackendoff: Spatial language

tempting, but given that they try to provide evidence for the maxims of Grice (1975), which extend the hypotheses of "design
isomorphism claim in the first instance, this is not a valid of language" and "design of spatial representation" presented in
procedure. the target article. By this hypothesis, we stress the two separate
The particular experimental data they cite in support of their communicative goals served by count nouns and spatial preposi-
claim are not convincing. Take, for example, the study by tions. Count nouns communicate object identity, providing
Landau and Stecker (1990) in which children and adults had to access to the "what" system, whereas spatial prepositions com-
learn new words, in this case nouns and prepositions. The very municate location information, providing access to the "where"
elegantly designed experiments show that children at the age of system. Furthermore, in normal conversation, spatial commu-
3 years are able to use syntactic information (prepositional nication does not exist outside of either a defining context or the
phrase vs. noun phrase) to interpret a given phonological shape perceptual-cognitive loop, be it explicit or implicit.
as encoding either an object or a location. This is an exciting Why are there so many count nouns and so few spatial
result. Children, like adults, are able to generalize over differ- prepositions? Count nouns provide a token to refer to represen-
ent object shapes when the new word appears in a prepositional tations in a very powerful pattern recognition apparatus. As they
phrase, but they are sensitive to shape when they have to learn a are normally used, however, count nouns are not intended to
new noun. L & J, however, take this finding as evidence that tell the listener anything new about the spatial properties of the
children and adults do not represent details about object shape objects. For example, consider the phrase "gorilla on a moun-
when they learn prepositions. The fact that children and adults tain." Our spatial system has stored the complex geometry of a
are able to generalize over different shapes when it comes to gorilla. We can use this knowledge to recognize a gorilla or to
learning the use of a preposition does not mean that they imagine one when the word "gorilla" is used. However, there is
represent the object in a sparse way during these learning no new information being communicated about the shape of the
instances. A necessary test for this claim would be to compare gorilla in the linguistic act. The rich spatial knowledge of the
recognition memory for object shape under both the noun pattern recognition system is isolated from the language system.
learning and the location learning situation. It may well be that The preposition "on" is the only word in this example that
even in preposition learning details of the object's shape are provides new spatial information. If we want to communicate
represented but these are just not relevant to the use of a new spatial information about the geometry of a gorilla, we need
preposition. to use spatial prepositions (or verbs) to convey the new
It seems that there is no solid empirical evidence to support L information.
& J's otherwise interesting claim that objects are sparsely Our language apparatus does not give us a very powerful
represented when a spatial relation is encoded. What is shown is mechanism to talk about shape. When we say "gorilla," there is
that object shape information is more relevant in noun use than no implicit shape built into the syntax of the word. The word
in preposition use but not that the underlying representation is provides access to a powerful pattern recognition system. The
of equal sparseness or richness, respectively. It may well be that system does not seem to lend itself to the general build-up of a
there is a rich underlying representation for given spatial scenes linguistic description of gorillas from its parts. There are many
and that only certain features are extracted when a spatial count nouns because we cannot rely on any system other than
relation is linguistically encoded. rote matching of complex shape to "word." Each count noun is a
It appears that the advantage and the drawback of language in token for some fuzzy set of shapes around some prototype.
general is that a given linguistic form generalizes over different Spatial prepositions do something very different. They tell
instances, objects, and even actions (people have different you where things are and they do so in a very structured
habits of drinking), but that at the same time this means a manner. When I say one object is "in front" of another, I am
linguistic form is underspecified with respect to a particular telling you something new. Furthermore, the context of the
instance, object, or action. situation and interaction with the perceptual apparatus often
resolve any ambiguity. If I say, "The pencil is near the tele-
ACKNOWLEDGMENT phone," it is most likely that the pencil is within a relatively
This work was supported by the Alfried Krupp von Bohlen und Halbach small area, say between 2 cm and 20 cm away from the phone. If
Stiftung. I say that the "mailbox is near the restaurant" or the "airport is
near the city," the figure is certainly more than 20 cm away from
the reference object. In addition, the spatial information con-
veyed by the term "near" in the example of the pencil is enough
Is spatial information imprecise or just to locate the object. In contrast, the purpose of the linguistic act
coarsely coded? in the airport example is not to specify an exact location but
convey class information about the metropolis (e.g., "We could
P. Bryan Heidorn & Stephen C. Hirtle hold the conference there, as there is an airport . . ."). Thus,
Department of Information Science, University of Pittsburgh, Pittsburgh, PA spatial prepositions transmit enough information to locate an
15260 object, which is the purpose of the linguistic act.
Electronic mail: pbh@lis.pitt.edu; sch@lis.pitt.edu
The interesting thing here is the different underlying system
Landau & Jackendoff's (L & J's) target article makes a strong case that is being used: It is not the pattern recognition system, but
for the assertion that there are different cognitive systems for the perceptual system. If one draws the listener's attention to
representing "what' and "where" in spatial language and spatial the correct region, the listener's perceptual system can resolve
cognition. We generally agree with this assertion and the argu- any minor ambiguity. Likewise, the geometry of the subject and
ments made for the usefulness of comparing language and object of a spatial preposition is unimportant for the communica-
cognition data. We also concur with the premise that any aspect tive task of putting things in "close enough so you can find them"
of space that can be expressed in language must also be present locations in space. The "where" system is powerful enough to
in nonlinguistic spatial representations. Our main difference is resolve the ambiguity.
with the conclusion that the representation of geometry of Furthermore, if we consider the type of coarse coding often
objects is "rich" relative to the representation of place, as argued used in neural networks (sect. 3.2.1; Hinton et al. 1986), we find
from the language data, and, as a result, the representation of that sufficient, and often quite accurate, spatial information can
place is imprecise. be encoded using relatively few "where" units. It appears that L
To explain the differences in spatial properties of count nouns & J have equated the coarseness of the representational code
and spatial prepositions we argue for a "functional design of with the coarseness of the spatial information conveyed by the
language" hypothesis, consistent with the conversational code.

246 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Commentary/Landau & Jackendoff: Spatial language

The argument that the different word types use separate but we will at least hit the rim.) And in the NBA it is often the
underlying cognitive apparatus and serve different functions spatial relationships of objects around and above the rim that are
also supports the idea that there are separate and different changing all the time.
"what" and "where" systems. According to this "functional Objection: Though other things may change position, the
design of language hypothesis" language develops to support a basket does not, so this is still a very special case of "where"
set of specific functions. The mechanisms that develop in the ability, and therefore does not imply a more generally rich
language are matched to the system being used. There are few "where" representational system. Perhaps. But Jordan also
spatial prepositions because that is all that is needed to support made a number of assists, and the "where" targets for his passes,
the spatial-perceptual system. There are many count nouns as well as those trying to defend against them, were, well, all
because we need to communicate the many object distinctions over the place and not standing still. It is hard to imagine a much
(but not shape distinctions) that need to be communicated. One richer environment in which to exercise and display your
can no more say that the "shape" system is rich and the "spatial" "wheres." To get to Pippin you must go through Drexler and
system is weak than that human pattern recognition is rich and around so and so and over so and so and . . . well, no wonder his
human perception is weak. tongue hangs out. Most spatial relationships for which there are
prepositions and, I claim, many more for which there are none,
are being exercised.
Perhaps there are objections that are fatal to the Jordan
example. It seems certain, though, that many other plausible
No perception without representation examples can replace it - examples from other sports, everyday
physical activities, experiments in stereo probe placements or
Donald D. Hoffman structure-from-motion probe placements (Braunstein et al.
Department of Cognitive Science, University of California, Irvine, CA 92717 1992), or other psychophysical experiments. And that is the
Electronic mail: dhoffman@orion.oac.uci.edu problem with the task that Landau & Jackendoff have set
Whatever we can talk about we can also represent. This is a key themselves. They are trying to set upper bounds on the com-
principle behind the arguments of Landau & Jackendoff (L & J), plexity of our "where" representations, but the data they collect
and, stated at this level of generality, it is certainly hard to take can really only set lower bounds. And in this regard they have
exception. No claim is made about the form of the representa- done a great service. I come away from their target article with a
tion - it could be abstract symbols, single neuron activity, or new respect for and understanding of the representational
population firing patterns. The claim is simply that language is capacity of our "where" system. It is almost surely no less
not magic. Each linguistic ability requires some underlying sophisticated than they have described. And many NBA fans
representational ability. count on it being much more.
What is true here for language is, most likely, true more
generally. There is no perception without representation. In-
deed, there is no perceptual or motor ability without represen-
tation. Denoting such an ability A, and a representation R, we
can write this dictum as A —* R. Evolution and physiology of "what" versus
L & J actually use the converse of this principle: -<A —> ->R. "where"
They show, with very thorough and illuminating examples, that
David Ingle
our linguistic abilities for describing "where" are very limited.
Department of Brain and Cognitive Science, Massachusetts Institute of
Thus, ->A. From this they conclude —>R, namely, our "where"
Technology, Cambridge, MA 02139
representations are very limited.
Airtight? Only if L & J have really established —*A, and they The target article by Landau & Jackendoff (L & J) documents
have not. To establish —>A it is not enough to show a limited and explores important differences between spatial and object-
ability with "where" in our language system; one must show this based representations as expressed in language. Because their
limitation in all of our perceptual and motor systems. If there is article necessarily skimps on the evolutionary background and
even one perceptual or motor system in which we have a richer physiological underpinnings of these two realms of perception,
ability with "where," then, by A —» R, we must posit the my comments will highlight those issues.
corresponding richer representation. It might be noted that the first explicit discussion of "what-
I think there are such systems. Consider Michael Jordan, for versus-where" systems was a four-man symposium published in
example. During the NBA playoffs he sank six 3-point shots and 1967 in Psychologische Forschung. Schneider (1967) dealt
amassed 35 points in a single half- from every position on the mainly with an attempt to distinguish tectal and cortical visual
court and every possible orientation (or so it seemed) of his body functions as orienting versus discriminating, and one could
in space. Put Jordan somewhere on the court and ask him to argue that his stripe-discriminations do not constitute "object
describe where the basket is. Ten to one, you could not make vision." The other three participants (Trevarthen [1968], Held
even a Iayup based on his description. Now hand him a basket- [1968], and Ingle [1967]) actually focused more on perceptual
ball and ask him where the basket is. Ten to one, you will soon distinctions between spatial and object vision, using examples
believe he knows, and without a word spoken. After asking him from humans, monkeys, and fish.
repeatedly from different positions and orientations, you will It would be of great evolutionary interest to devise tests
soon believe he has a very rich ability with "where" - and thus, comparing the number of spatial distinctions versus the number
by A —» R, a very rich representation of "where." of object discriminations animals make. Most animal psycholo-
Objection: This is a highly trained ability in a talented man. gists would probably support my prediction that there are
True. But most of us can at least hit the rim most of the time, and smaller differences between rodents and primates (including
swishes are riot needed to make this point. man) in spatial route-finding or object-retrieval than in object
Objection: Baskets are simple objects, hardly a case in which discrimination skills. Thus, an evolutionary explosion of object-
complex "where" interrelations are needed. True. But remem- classification abilities in the higher mammals probably precedes
ber that those 35 points came with 9 other men on the court (5 language capacities, and may be a precondition of language.
desperately trying to stop him by any means the refs could not L & J seem to fall into an error common among physiologists
see), and with thousands of fans not exactly in quiet meditation. in discussing the relationship of low-level cortical processing
"Where" relations were changing rapidly even after he began a (color, orientation, motion coding) to the shape/space dichot-
shot. (Most of us, once again, will not score in such conditions, omy. It is quite clear that lines and edges can define the three-

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 247

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 CommentarylLandau & Jackendoff: Spatial language

dimensional layout of surfaces as well as the shapes of objects. ties. Is this multiplication of theoretical entities justified? I will
More dramatically, it is clear from the "biological motion" consider each distinction separately.
demonstrations of Johannson (1973) that rather complex arrays The linguisticlsublingulstic distinction. In partially accepting
of moving spots (with no contours visible) can yield highly the Design of Language Hypothesis, L & J explain that the
specific identification (e.g., a man hammering, a couple danc- language faculty has access only to part of the general spatial
ing) as well as yielding percepts of three-dimensionality via representational system. This part is a "filtered" version of the
parallex on expansion effects. There are many examples to show richer metric descriptions at the sublinguistic levels accessible
that motion and contour processes can feed into either spatial or to perceptual and motor systems. Filtering, they say, is likely to
object shape recognition systems. Milner and Goodale (1992) be "a design feature required by any system that must collapse
have recently made a similar argument in discussing parietal numerous complex distinctions onto a finite set of elements."
(spatial) mechanisms involved in grasping oriented objects. These elements are still geometric categories (such as geons,
L & J do recognize that spatial vision involves unconscious directed axes, and angles), but are more qualitative and sche-
computations that do not require language, but the extent of matic than the corresponding sublinguistic encodings.
these hidden computations exceeds the examples they have This filtered/unfiltered distinction brings to mind such vener-
listed. In my own recent work, it is clear that humans (even able dichotomies as conceptual/perceptual, discrete/analog,
young children and mental retardates) accurately localize re- and descriptive/depictive. Indeed, the filtered representations
membered targets on the floor of a large room after walking and are more abstract, less visualizable. However, both the linguis-
turning along a disjointed route without vision. What is con- tic (filtered) and sublinguistic (unfiltered) systems seem to be
sciously perceived is the target's stable location in space, not the conceptual, discrete, and descriptive - modifications of, not
distances walked or angles turned; those computations remain major departures from, data structures such as Marr's (1982) 3-D
hidden. model. Both are based on a "finite set of elements" and make
Other examples include the prediction of trajectories of "numerous complex distinctions." The filtered and unfiltered
missiles (baseballs in our era, spears in Paleolithic times) that systems, it appears (though L & J do not go into detail), differ
must be caught or avoided. No outfielder can describe to a only in the amount of detail they can encode, not in the basic
trainee how he judges whether a fly ball will carry beyond him. mechanism of representation (as, for example, the dis-
Somewhat more complex patterns (not yet analyzed in the crete/analog distinction standardly implies).
laboratory) include estimating the timing and constraints of a It is therefore unclear why language is restricted to using only
human arm swinging a weapon so as to efficiently duck and the set of "filtered" representations. For example, what barrier
counterattack. In general, the fighting skills of carnivores, birds, prevents a lexical entry for a count noun from referring directly
and even lizards (where jaw fencing is a skill comparable to that to a geometric model constructed in the fine-grained, sub-
of the Three Musketeers) require some impressive computa- linguistic medium, given that it can refer to one in the coarse-
tions at an unconscious level. As we begin to analyze spatial grained medium? Such a barrier is necessary to make plausible
skills involving complex motion of both subject and object, we the major claim of L & J's target article - that the linguistic
will perhaps realize that a large part of the mammalian brain is where system is a much coarser medium than the linguistic what
devoted to largely unconscious (at least routinized) computa- system - because the sublinguistic what and where systems are
tions for acting within a spatial framework. both admittedly quite fine-grained and capable of capturing
Object and spatial vision should be compared not in their metric details of objects and configurations. But independent
implicit complexity, but in terms of which representations justification for, and explanation of, this linguistic/sublinguistic
normally become part of conscious deliberation. This is a dichot- barrier is needed.
omy within the recognition mode as well (e.g., recognizing up to Such analytic support could come out of a recent movement
10,000 different faces is accomplished without verbal analysis, for theoretical restructuring within cognitive science: "situ-
whereas identifying the kind of emotion or social implication in a ativity theory" (Greeno 1992; Greeno & Moore, in press). As a
generic facial expression does map onto subtle verbal labeling). way of reconciling traditional information-processing psychol-
Finally, I suggest that Landau & Jackendoff's ideas be further ogy with ecological psychology, these theorists argue that lin-
validated by counting the number of object categories that can guistic and other symbolic knowledge should be analyzed in a
be remembered by very young children before and after lan- fundamentally different way from sublinguistic knowledge. For
guage becomes differentiated. Is the richness of object memory example, much basic knowledge, such as a mammal's under-
more related to chronological age or to verbal performance in standing of Euclidean space, is better characterized as based on
cases where the two can be dissociated? "attunements" (Greeno et al., in press) or "internalized con-
straints" (Shepard 1984) rather than "representations." Instead
of being used ubiquitously as in traditional cognitive science,
"representation " is reserved for a particular kind of attunement,
mediated by symbol systems explicitly used and assigned mean-
Distinguishing the linguistic from the ing by the agent. Verbally expressible knowledge (such as that
sublinguistic and the objective from the measured via think-aloud protocols) falls into this category.
configurational That this distinction between attunement-knowledge and
representation-knowledge could usefully elaborate a theory
Scott D. Mainwaring such as L & J's is an exciting possibility.
Psychology Department, Stanford University, Stanford, CA 94305-2130 The what/where distinction. Whereas the sublinguistic/
Electronic mail: sdm@psych.stanford.edu
linguistic distinction categorizes geometric knowledge based on
Landau & Jackendoff's (L & J's) analysis centers on two key something like degree of abstraction, the what/where distinc-
distinctions, what versus where and (what I will label) linguistic tion divides it according to subject matter. According to the
versus sublinguistic. Conjoining these, a four-way classification Design of Spatial Representation Hypothesis, the what system
of geometric representational systems underlying spatial underlies knowledge of objects, including complexes composed
thought and language results. Not only are there separate what of multiple parts; the where system underlies knowledge of the
and where systems, but also distinguishable linguistic and sub- spatial relationships between objects comprising a configura-
linguistic partitions within each. This represents a significant tion. The principal claim is that, at the linguistic level, the what
elaboration of the "3-D model structure" in Jackendoff's (1987a) system encodes part shapes and interpart connections in detail,
theory of the interface between the spatial and linguistic facul- whereas the where system encodes object shapes and interob-

248 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Commentary /Landau & Jackendoff: Spatial language

ject relationships more schematically. (How the what and where above nose is what makes a drawing of a face a face; but mole
systems differ at the metric, sublinguistic level is less clear.) above the eyes is irrelevant to its being a face, so the location of
Loosely, the object system is to the configuration system as a set the mole relative to fixed features has to be worked out anew for
of aerial photographs of cities is to a low-resolution road map: each face.
One directly encodes detailed shape information of delimited Again, a face and a farmyard are both "objects" differing only
areas, the other compresses and schematizes shape and location in scale: The nose at the front/the house at the front; the hair at
information to convey regional spatial relationships succinctly. the back/the barn at the back; the nose in the middle (of the
By linking the two via "indexing," integrated environmental front/the well in the middle; the mole at the side of the nose/
knowledge emerges (see sect. 3.2.2). the garage at the side of the barn; and so on. What makes the
As L & J point out, what and where knowledge is complexly location of an object relative to some reference frame complex,
interdigitated. Both object and configuration knowledge consist we suggested, was that it is often variable, whereas the location
of elements associated together spatially. The heart of the of spatial features of objects relative to each other is fixed.
proposal, as I read it, is that the way this association is accom- Variable locations may be difficult, not because they involve new
plished in the object/u>/jat system differs from that in the spatial predicates, but because they have to be constructed
configuration/if/iere system. This suggests that the same physi- anew.
cal complex could be encoded as an object or as a configuration, These two issues may also be related. Perhaps what makes the
resulting in two very different ways of understanding it. It would representation of variable location possible is the possession of a
be interesting to work out an analysis of environmental learning, set of concepts, concepts that would permit an object or feature
typically glossed as progressing from landmark to route to of an object to be represented (sense 2) as being round or being
survey knowledge (e.g., Moore 1976), as involving initial devel- at X. Indeed, it is reasonable to expect that representing (sense
opment of direct but loosely bound configuration knowledge 2) the location of a mole as being to the left of the nose is, in
followed by reconceptualization of the environment as a single principle, no different from representing the eyes as being
complex object that, like a map, can compactly though indirectly above the nose. Neither is required for recognizing a face or
orient one within the actual environment. recognizing Aunt Maggie.
Recent work by Neisser (1989; 1992) presents an additional
perspective on the kind of information used by the what versus
where systems. He argues that the what system operates by
recognition, comparing gradually accumulating evidence with
stored representations (criteria for belief). The where system Is spatial language a special case?
operates by "direct perception," picking up invariants from the
rich flux of sensory information (Gibson 1966). Considering Dan I. Slobin
such epistemological issues in object recognition versus configu- Department of Psychology, University of California at Berkeley, Berkeley,
ration perception may be a fruitful direction for further elabora- CA 94720
tion of Landau & Jackendoff s thought-provoking model. Electronic mail: slobin@cogsci.berkeley.edu

Landau & Jackendoff (L & J) describe two different kinds of

phenomena and try to put them together into a single story:
Information about objects and locations is separately repre-
sented in the mind/brain, and information about objects and
Spatial development locations is encoded by different means in language. But the
connective tissue is missing from their argument. Why should
David R. Olson objects be encoded by nouns and spatial relations (in some
Ontario Institute for Studies in Education, 252 Bloor St. IV, Toronto, languages) by prepositions? And why should nouns point to
Ontario, Canada M5S 1V6
richly detailed objects, whereas prepositions point to schematic
Electronic mail: internetd.olson@oise1.oise.on.ca
relations? I submit that the linguistic division of labor reflects
Landau & Jackendoff's (L & J's) impressive synthesis leaves the way in which the cognitive distinctions are mobilized online
unresolved two of the issues that so puzzled us (Olson & for communicative purposes.
Bialystok 1983) in our theory of children's spatial development. L & J pose the question in the following terms: Why is it that
First, what is the relation between representing an object object names represent the geometric richness of shape whereas
such as a face in terms of spatial properties including roundness, spatial prepositions represent a limited and coarse encoding of
and representing a face as round. L & J use the terms "descrip- "place"? They say that the naming of objects requires "detailed
tion" and "representation" interchangeably, as does much of the descriptions of shape;" however, object names do not describe
cognitive science literature, to mean both a property that figures the shapes of objects; rather, the name references the object.
in some way in the representation of an object (sense 1), and a When I say "frog," I do not describe the shape of a frog, but
concept, category, or word that represents that property .as a evoke whatever representation of frog you have in your mind.
member of a class (sense 2). Preverbal children recognize balls But when I saw "the frog is in the jar" I invite you to represent a
on the basis of their roundness, but they represent them as spatial relationship between a frog and ajar. The question is not
round, and thereby include them in the class of round things, why some object categories are definable in terms of shape, but
only when they are around 4 years of age. We argued that spatial rather why there are so few grammaticized terms for spatial
cognitive development consisted largely of translating the spa- relationships, and why these terms tend to ignore details of the
tial features implicit in the representation (sense 1) of objects figure and of the reference object.
and events into conceptual spatial representations (sense 2). To First consider the meanings conveyed by spatial terms. Note
fail to mark this distinction is to lose, I suggest, the possibility of that (in languages like English) spatial prepositions do not occur
explaining cognitive development. alone but are accompanied by nouns that reference figure and
Second, L & J's willingness to adopt the "what" versus ground. The words "frog" and "jar" already tell you a good deal
"where" distinction from neuropsychology leads them to over- about the possible relations between these two objects. This is
look an important aspect of the relation between objects and why many languages do without the precision of English spatial
locations. Objects are merely collections of parts with fixed prepositions when describing canonical or highly predictable
locations relative to each other. Conversely, locations specify relations between two objects. Spanish, for example, makes do
variable relations between parts, often of macro "objects." Eyes with a single preposition, en, to describe both frog-m-jar and

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 249

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Commentary/Landau & Jackendoff: Spatial language

frog-on-table. Turkish does not make use of available locative propose that this is precisely because the human mind/brain is
postpositions at all to describe these scenes; an all-purpose constructed to conceive of object destruction or manner of
locative inflection is sufficient: "frog jar-LOC" and "frog table- movement or whatever just in terms of those features that can be
LOC." Both language types have recourse to more specific used to describe these closed class lexical sets. Convenient, and
relational expressions for noncanonical locations - the equiva- maybe true, but, if so, not due to the difference between the
lents of "the frog is on top of the jar, under the table," and so on. "what" and the "where." It is, of course, a truism that whatever is
These facts suggest that a good deal of inferencing is carried out, expressed in language must also be present in nonlinguistic
making use of representations of the objects associated with representations. This initial "premise" of L & J couldn't be
locative terms to more fully interpret the grammaticized spatial otherwise. But why are there relatively closed and relatively
representation. open sets of lexical items?
Other languages, as noted by L & J in their discussions of L & J propose that one open class - "count nouns that label
Atsugewi and Tzeltal, provide verb-marked information about different kinds of objects" - is based on our capacity to represent
characteristics of figure and/or ground. It is interesting that shapes and they explain that the visual system is constructed to
such languages make sparse use of nouns in their discourse. represent many different kinds of shapes. But the capacity to
In Atsugewi (Talmy 1975), for example, a verb meaning represent shape hardly accounts for the range of object names in
"don't-cause-dirtlike-substance-to-move-downwards-into-liquid" language. L & J state that "descriptions must be potentially fine-
is sufficient, without the use of any nouns at all, to command you grained enough that one can decide which objects are to be
not to flick your cigarette ashes into my coffee, or not to dump named horse and which donkey, or which dog and which wolf."
the coffee grounds into the river, because the relevant object Note, however, that these "descriptions" are not part of lan-
representations can be determined from visual evidence alone guage; they are descriptions of how the visual system and brain
(i.e., eontextually), given the schematic object representa- might recognize objects. The reason we have words for horses
tions encoded on the verb. Tzeltal (Levinson 1991) gets along and donkeys is that we can distinguish them - by whatever
with only one spatial preposition, an all-purpose term simply means (just as we have words such as mistress, lover, and wife,
meaning "located with regard to." Thus, "sits-bowl-like which cannot be distinguished by shape at all). L & J confuse the
LOCATIVE-PREPOSITION table gourd" conveys the same fact that we can recognize shapes and the fact that we can name
information as the English the gourd is on the table. Further- objects, some of which can be characterized by shape. Lan-
more, given the information in the verb, one or both nouns may guages have a large number of nouns because people categorize
be omitted. the world in a myriad of ways. Shape recognition is a fascinating
Across all languages it seems that information about spatial scientific topic but it in no way explains why people need so
relations cannot be calculated without knowing what kinds of many different ways to refer to objects of experience.
objects/substances are being related. Why, however, is the list Why do we use nouns instead of descriptions? There are at
of object and spatial-relational characteristics so limited? L & J least two answers: (1) Fully specified descriptions are impossible
suggest that this is due to a separation between the "what" and (e.g., there is no way to substitute the noun dog with a series of
the "where" systems of the mind/brain. But this division does words that fully, or even adequately, describes a dog), and (2)
not dictate that the "where" system be so limited and sche- partial descriptions are cumbersome. Our minds are so con-
matized. One reason the "where" system can be linguistically structed as to quickly and easily access a particular mental
limited is because of what can be inferred from the accompany- representation in association with a particular word - a sound
ing nouns (or the physical presence of their referents). This pattern that is totally arbitrary with regard to the associated
suggestion introduces a factor not considered by L & J: economy concept. When necessary, we use word combinations to name
of information processing. objects - if they are new or complex or unfamiliar - but we
In making such a proposal, it is appropriate to include spatial reduce them to shorter expressions and ultimately to words
verbs and verb particles as well as adpositions (prepositions and when they become frequently used (intercontinental ballistic
postpositions). L & J exclude spatial verbs on the grounds that missile to ICBM, analysis of variance to ANOVA, etc.). One
they are "open class elements" (i.e., elements of a large, unre- could imagine a language with a small number of words in which
stricted set that can be added to by speakers). Although it is true every referring expression was phrase-like or longer. Such a
that verbs, in general, constitute one of the major open classes language would be far more transparent than our languages, but
(along with nouns, and, in some languages, adjectives and it would require a different kind of processor. The human
adverbs), languages also have relatively small and relatively processor has to make fast decisions: We can integrate verbal
information only over short stretches of time. We economize by
closed subclasses of verbs that function to schematize asemantic
packing a great deal of information into brief bursts of sound,
domain in the same way as prepositions (or personal pronouns,
such as nouns.
modal auxiliaries, or tense/aspect markers, etc.). For example,
one such closed subclass of verbs deals with object destruction. It is not surprising, then, that we also economize in our
In English, which has a fairly elaborate collection, we make relational expressions. Such economy is evidenced not only in
distinctions of the nature of the object to be destroyed (e.g., the spatial prepositions on which L & J focus; it is characteristic
break, tear, smash), force dynamics (e.g., tear vs. rip), the of all grammatical choices. We might just as well ask why
degree of destruction (e.g., cut vs. shred), the texture or constit- Russian has only six grammatical cases, or why Bantu languages
uency of the object (e.g., crumple, crumble, shatter), and so have a dozen or so noun classes, or why the English aspect
forth. The list of such verbs, however, is fairly short: I estimate system has only perfect and progressive. Every utterance is only
that there are no more than 50 such verbs in English - fewer a schematic sketch of the content to be conveyed. The "open
than the prepositions enumerated by L & J. There are many class" content words - nouns and/or verbs - point to the entities
such "closed subclasses" of verbs, for example: verbs of manner and events evoked by the utterance and the "closed class"
of movement (walk, run, jump, swim, fly), manner of talking grammatical elements suggest a framework of temporal, causal,
(shout, scream, whisper, mumble, mutter), posture (sit, stand, spatial, and social relations in which those entities can be
lie, crouch), and so forth. The Tzeltal spatial verbs seem no situated. It is characteristic of languages to have a collection of
different: Though there may be about 300 of them (Levinson obligatory grammatical elements that must be used repeatedly,
1991), still they are a closed set, characterizable by a small generally in every sentence. A grammatical/conceptual distinc-
collection of systematic distinctions. In each instance, we might tion that must be accessed so frequently operates under several
ask, following the model of L & J: Why are there only 50 verbs of
constraints: (1) It must be maximally general, so as to apply to
object destruction, or 25 verbs of manner of movement, or 6
most instances. Thus, languages do not inflect nouns for the
verbs of posture? And, continuing their example, one might
color of the referent object, perhaps because most objects do not

250 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Commentary /Landau & Jackendoff: Spatial language

have distinctive colors. (2) It must correspond to a way in which position of having to amend the theory in an ad hoc manner.
humans recognize and store events. Thus, for example, no Second, because they confuse the different roles spatial rela-
known language marks its verbs to distinguish whether an event tions play in visual recognition and representation, there is no
took place on a rainy or a dry day, perhaps because this distinc- clear mapping between spatial prepositions and spatial
tion is not relevant to the ways in which we represent most relations.
events. (3) It must contrast maximally within a small class of The volumetric primitive approach (Biederman 1987; Marr &
distinctions. Here there seem to be two types of online con- Nishihara 1978) to object representation is attractive for many
straints: (a) If a grammatical distinction is obligatory, that is, if it reasons, not the least of which is its computational elegance.
has to be expressed in every applicable sentence, the choice Symptomatic of the problematic nature of this approach, how-
tends to be binary (e.g., perfective/imperfective, singular/ ever, is L & J's need to "enrich" it by incorporating con-
plural) or ternary (e.g., singular/dual/plural, here/there/ straints gleaned from linguistic evidence. It is not that three-
yonder). This is probably because online decisions cannot enter- dimensional part-based descriptions will not have some role in
tain many alternatives. Such categories must be general, sa- visual cognition but that these descriptions are fundamentally
lient, and minimally partitioned so as to allow for constant, rapid inadequate in and of themselves for explaining a great deal of
allocation of every noun or verb to the appropriate subcategory. human visual behavior. This is dearly evidenced by numerous
Such obligatory markers lay out the basic framework of proposi- experimental results on human object recognition that are not
tional content and relation: tense/aspect/modality, transitivity/ accounted for by volumetric models (Bulthoff & Edelman 1992;
causativity, person/number, speech act, and so on. (b) Spatial Jolicoeur 1985; Tarr 1989; Tarr & Pinker 1989). Indeed, propo-
prepositions represent a different kind of online problem, in nents of the volumetric approach (Biederman 1987; Biederman
that every sentence does not require specification of spatial
& Cooper 1991) have been careful to state that such representa-
relations. If a sentence does have a prepositional phrase, it
tions are restricted to basic-level access and that other forms of
requires a particular preposition. (Similarly, if it does have a
object representations and recognition mechanisms subserve
modal phrase, it requires the choice of a particular modal, and so
"complex" recognition. The inherent need for these representa-
forth.) Here, it seems, languages provide relatively small sets of
choices: a handful of modals, several handfuls of prepositions, a tions is indicated by L & J's introduction of surface representa-
small basket of spatial verbs. tions to account for objects such as paper, phonograph records
(compact discs?), and lakes. Unfortunately, as incorporated by L
The reason that sets of grammatical options are small is not & J, this extension is ad hoc in that it is conceived solely to
because human beings are incapable of conceiving of more account for extant linguistic data. This same theoretical weak-
detail. We express more detail whenever we need to, with the ness occurs in a variety of guises: orienting axes, directed or
extended possibilities of building phrases and clauses. The symmetric axes, negative object parts, containment, and so on.
"where" system can be linguistically expressed in exhaustive Essentially, whenever a spatial property is linguistically repre-
detail, but not in the grammatical component. I propose that the sented, a corresponding property is posited in the spatial
constraint on what is grammaticized in human languages is not representation.
due to the structure of cognition alone, but is also due to what is L & J attempt to sidestep the "special-purpose" characteris-
useful, quick, and easy to access online: to enable us to program tics of their theory by challenging "critics to develop a simpler
utterances while speaking and to make use of received ut- alternative" (Note 8). Yet even if a better alternative could not be
terances while listening in order to build up a mental represen- offered, it is crucial that elements of any theory of spatial
tation of the communicator's intent. A processor of the human representation play a functional role in spatial reasoning, recog-
sort must make rapid decisions with regard to all grammaticized nition, and navigation. As formulated by L & J, the proposed
notions - not only notions of spatial relationship. Such decisions additions serve a purely explanatory role. Moreover, there are
cannot require attention to a great deal of specific detail. To be plausible alternative theories of object representation that im-
sure, the schematization reflects the most natural predisposi- plicitly include many of the spatial properties that are to be
tions of the human mind, but these are cognitive systems with a found in linguistic structure.
dynamic component. I suggest that the interface between gram- Recently, both computer scientists (Koenderink 1987; Seibert
mar and cognition is influenced by our ability to schematize & Waxman 1992; Ullman & Basri 1991) and psychologists
experience for communicative purposes just as much as it is (Bulthoff & Edelman 1992; Tarr 1989; Tarr & Pinker 1989) have
influenced by our biologically determined predispositions to proposed multiple-views theories of object representation. One
schematize experience as we do. important element of such theories is that shape representations
are egocentric and thus encode the relative position of the object
to the observer. That is, each view of an object depicts the
appearance of the visible surfaces of that object from a small
range of orientations. Multiple-views models include many of
From perception to cognition
the properties that L & J have argued are necessary in theories of
Michael J. Tarr spatial representation. First, because properties such as left and
right are inherently egocentric, they are implicitly encoded in
Department of Psychology, Yale University, New Haven, CT 06520-7447
Electronic mail: tarr@cs.yale.edu
each orientation-specific representation (Tarr & Pinker 1989).
Second, qualitative spatial relationships between views (e.g.,
I am sympathetic to Landau & Jackendoff's (L & J's) proposal front/back and top/bottom, as well as explicit encoding of
that linguistic structure may serve as a window on the nature of left/right) provide sufficient information to infer adjacency be-
human spatial cognition, reflecting the "deeper" structural con- tween views, whereas qualitative spatial relationships between
straints of spatial representations. Indeed, mechanisms of hu- object parts provide category information (the role Biederman
man visual cognition may have evolved concurrently with those [1987] has suggested for part-based models). Third, surfaces and
used for language (Tarr & Black 1991), and, in all likelihood, as containment are easily represented without modifications or
the result of similar adaptive pressures (White 1989). Despite additions to the theory. Both are simply encoded in the number
my generally favorable reaction to this approach, however, L & and variety of views necessary to completely represent an object
J's treatment of theories of spatial representation 1 raises con- at a given scale (Koenderink 1987; Kriegman & Ponce 1990). In
cerns that may detract from their effort to draw plausible contrast to L & J's proposal, therefore, multiple-view theories of
connections to linguistic structure. First, in adopting an ap- spatial representation have principled (e.g., functional) reasons
proach to object representation of somewhat limited explana- for incorporating may of the properties identified as elements of
tory power - that of volumetric primitives - they are put in the spatial language.

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 251

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Commentary/Landau & Jackendoff: Spatial language

My second concem is that in their mapping from spatial objects, it offers only a meager means for talking about places.
prepositions to spatial relations, L & J have equated the so- Moreover, they argue that these differences in language have
called "what/where" distinction with the recognition of objects parallels in the "nonlinguistic disparity between the representa-
and the recognition of places. In fact, the role of spatial relations tions of'what' and 'where " in the brain. In the course of their
in visual cognition is far more complex. There exist at least three argument, L & J take us through an insightful review of both
levels at which spatial relations may be used to encode language: spatial cognition and the language of objects and location. In the
(1) between-part relationships within an object (i.e., "her head end, however, their argument collapses. Let us see how.
was on top of her body"); (2) between-object relationships within L & J's first claim rests on the premise that objects are
a scene (i.e., "the calculator was on top of the desk"); and, (3) encoded by nouns and places by prepositions. "In English,"
between-place relationships within the environment (i.e., "the they say, "objects are represented by count nouns and places are
apartment on top of the Chinese restaurant"). Object location represented canonically by prepositions or prepositional
(e.g., "where" in sense 2) is computationally distinct from place phrases." Although they add the qualification "or prepositional
location (e.g., "where" in sense 3). For the most part, the phrases," they proceed in their argument as if they hadn't. They
"where" system as discussed by L & J refers to object location - a then note that nouns are much more highly developed in
problem equivalent to the problem of segmentation - for exam- languages than prepositions. Languages have many more nouns
ple, determining the position of an object independently of its than prepositions, nouns have more elaborate conventional
identity. Both senses of "where," along with within-object rela- meanings than prepositions, and so on. It follows, they main-
tions, are used in spatial language, hence the parallel drawn tain, that languages have a much richer means for talking about
between the bifurcation of language (count nouns and preposi- objects than places. But this only follows if the initial premise is
tions) and spatial cognition (what and where) is inadequate for correct, and it does not seem to be.
capturing the complexity of the mapping. Prepositions are neither necessary nor sufficient to represent
An alternative bifurcation of spatial cognition may provide a places or locations. Verbs like "support," "hold," and "contain,"
clearer path for making comparisons. One possibility is that the for example, take subjects that denote locations, without prepo-
visual system distinguishes between metric information and sitions, as in "The table supported the statue," "The vase held
qualitative information. When placed in this framework, many the jewels," and "The boat contained illicit cargo.' Likewise,
levels of spatial representation can be shown to incorporate both there are no locative prepositions in "Chicago is windy," "His
kinds of information. As touched on earlier, there is experimen- shoes are muddy," "His garden is full of flowers," or the proto-
tal evidence that object representations may include metric typic expressions of place, "here" and "there." Prepositions are
information stored as views and qualitative information, as the merely relational terms. In "The statue is on the table," "on"
relations between the views and object parts (Tarr & Pinker denotes a relation between an object (the statue) and a place (the
1990). Likewise, neurophysiological studies (Paillard 1991) table) just as the verb "support" denotes a relation in "The table
and implementations of exploration algorithms (Engelson & supported the statue."
McDermott 1992) both suggest that representations of place Places, therefore, are entities, not relations between entities,
also include both metric and qualitative information. No doubt and are regularly represented by count nouns. "Table," "child,"
the same holds true for the relations of objects in a scene. Given
and "apple" can be used to express either objects or places,
this sort of distinction, the connection between spatial preposi-
depending on your perspective. If you "move a table," "watch a
tions and spatial relations may be revised, prepositions being
child," or "pick up an apple," you are treating the table, child,
used to capture qualitative relational information between
and apple as objects. But if you put "a chair at the table," "a hat
parts, objects, or places, and count nouns being used to capture
on the child," or "a knife through the apple," you are treating
shape information about parts, objects, or places.
them as places. Yet places are entities we can relate other
In summary, L & J raise many interesting and important entities to, as expressed in locative predicates such as preposi-
points about the connection between spatial language and spa- tional phrases (prepositions plus noun phrases). In English, the
tial cognition. Bridges between the two domains must be built answers to "where" questions (like "Where did you put the
upon a sound foundation, however, one in which there are chair?") are ordinarily locative predicates ("At the table"). They
principled reasons for the connection at both ends. Indeed, such can never be prepositions alone ("At") unless the location can be
cross-fertilization is essential if we are to understand how human inferred. In locative predicates, the place information carried in
information-processing progresses from perception to cogni- the noun phrases is essential.
tion. Prepositions, L & J argue, make reference to only a limited
set of properties of locations. "On" in "on the couch" calls on only
NOTE
1. Marr's (Marr & Nishihara 1978) and Biederman's (1987) theories a few of the many properties of couches. From that, L & J seem
are both limited to visual representation and essentially only to object to conclude that people's representations of locations must
shape. In contrast, a complete theory of spatial representation would themselves be limited. This argument has serious flaws. First,
include models of the environment (Paillard 1991). For example, audi- prepositions pick out only a fraction of the locative information
tory perception in humans contributes only location information (e.g., that is actually exploited in locative predicates. Consider "There
for controlling eye movements, Jay & Sparks 1987) but tells us nothing is/are X on the couch" and think of all the subtle inferences you
specific about object shape (although identification may be made make about location when X is "an adult," "two lovers," "a
through auditory features). cushion," "a spider," "dirt," "a label," "four legs," "anti-
macassars," or "a bug (hidden listening device)." This locative
information, we suggest, is in principle unlimited. Second, L &
J seem to imply (in Whorfian fashion) that the poorer the
prepositional system in a language, the poorer the representa-
Prepositions aren't places tions of location. By this logic, speakers of Tseltal should have
empty representations of entities viewed as places because
Barbara Tversky and Herbert H. Clark Tseltal has only a single all-purpose preposition. On the con-
Department of Psychology, Stanford University, Stanford, CA 94305-2130 trary, speakers of Tseltal have an extraordinarily rich system for
Electronic mail: bt@psych.stanford.edu & herb@psych.stanford.edu
describing places. It just happens not to rely on prepositions
People have extensive knowledge of objects and places, what (Levinson, in press). In short, the poverty of the semantics of
things are and where things are, and they are able to talk about prepositions cannot be used to argue for a poverty in the
both. Landau & Jackendoff (L & J) argue, however, that al- representation of place.
though language offers people a rich means for talking about Intricate location information underlies our very conceptions

252 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Commentary /Landau & Jackendoff: Spatial language

of many objects. As L & J argue, languages have large vocabul- From observations on language to theories
aries for classes of objects, and many of these terms are under- of visual perception
stood as referring to shapes. Yet critical to shape, as L & J note
later, is location. An object gets its shape from its parts, but the Johan Wagemans
parts must be arrayed in a particular configuration (Tversky &
Laboratory of Experimental Psychology, University of Leuven, B-3000
Hemenway 1984). The difference between a T and a +, or Leuven, Belgium
between a coherent object and a collect of parts, is precisely how Electronic mail: fpaas10@blekul11.bitnet
the parts are located with respect to each other. Place, then, is
critical to mental representations of objects. Landau & Jackendoff (L & J) explore the language of objects and
Place must also be represented in the semantics of object places to learn more about their underlying representations.
names. Terms for parts, such as "back," "fender," and "leg" (of a They conclude that objects and places differ considerably in the
person, chair, or table), must represent the location of the parts geometric richness with which they are encoded. Object nouns
with respect to the wholes they are parts of. One large class of imply more detailed representations that those that are cur-
nouns must be represented in part by the potential locations of rently suggested in theories of visual object recognition,
their denotata, for example, "blankets" (they go on beds), "roofs" whereas spatial prepositions imply much sparser spatial infor-
(they go on buildings), and "plugs" (they go in holes). That makes mation than must be available for complex visual tasks such as
it easy to turn them into denominal verbs that express subtle eye-hand coordination and navigation. Although this linguistic
locational relations: "We blanketed the bed, roofed the garage, perspective on mental representations provides some challeng-
and plugged the holes" (Clark & Clark 1979). Another large class ing hypotheses, there are many gaps to bridge when going from
of nouns must be represented in part by the locative relations of observations on language to theories of visual perception. I
other things with respect to their denotata, for example, "dock," think L & J jumped to unwarranted conclusions in some cases. I
"lodge," and "list." These are also useful as denominal verbs will discuss two of them, formulating a question in relation to
describing special locations: "We docked the boat, lodged the each of them in turn.
guests, and listed the groceries we needed." So place is inherent First of all, L & J propose some enrichments of spatial
to - and richly represented in - the understandings of countless representations of objects that are needed to allow names for
nouns. spatial parts (such as top, bottom, front, back, sides, and ends),
A second important claim in L & J's target article is that "there for objects best described as surfaces (such as records and lakes),
are significant differences in the geometric richness with which for "negative" object parts (such as ridges and grooves), and for
objects and places are encoded," specifically, that "detailed containers. These enrichments do not seem too remote from
geometric properties [of the object] - principally its shape. . . . current models of object representations, but I have my doubts
- are represented," but for places, "only very coarse geometric whether all the fine distinctions needed can be derived from the
properties are represented, primarily its [the object's] main available visual input. I am strongly convinced by Marr's (1982)
axes." This could be a claim about encoding in language or a argument that there are two constraints determining the con-
claim about encoding in the brain, or a claim about both. Clearly, tent of a representation: further task requirements (such as
language can be used to describe both objects and places being able to speak about it), heavily stressed by L & J, and the
vaguely or specifically. Objects can be referred to by nominals possibility of perceptual recovery from images or intermediate
that convey minimal geometric properties, like "this," or "in- representations, unwarrantably neglected by L & J. For exam-
strument," or "thing." Moreover, labels may be understood ple, I can easily imagine that the visual system can compute
differently from their conventional meanings. In a given situa- object and part axes based on elongation or symmetry, but I do
tion, we may know exactly what "thing" is being referred to, and not see any obvious computational scheme to distinguish top
encode quite specific features, yet in cases where terms and from bottom or front from back. I believe that perceptual
objects are unfamiliar, we may encode only vague features from representations are what they are because of bottom-up as well
specific terms. Also, there is no simple correspondence be- as top-down constraints (available input information and task
tween what the perceptual system finds easy to represent and requirements, respectively). Perhaps later nonperceptual rep-
what speakers of a language find easy to talk about. Yes, people resentations are more fully based on requirements from lan-
are good at identifying shapes and good at describing them for guage and other cognitive faculties only. In sum, my question to
others to pick out. However, people are also good at identifying L & J concerns whether or not they would consider their
faces, and yet bad at describing them for others to pick out. enriched formats of object representation to be perceptual. To
Similar to linguistic representations, mental representations answer this question L & J will need some dividing line, which
of objects and locations may be rich or coarse. Given our ability might be on- versus off-line, bottom-up versus top-down, or
to catch fly balls, sink holes in one, return tennis serves, play the some other criterion they like more.
piano or violin, avoid obstacles in familiar environments in the A second concern is about the criteria for distinguishing
dark, and much more, it would be rash to claim that places are between those task requirements that have implications for the
mentally represented any less richly than objects. Success at any underlying representations and those that do not. For example,
of these activities seems to depend on detailed representations L & J explicitly mention that metric information must be
of geometric and other properties of objects with respect to encoded to support performance in fine motor tasks such as
locations. The neuropsychological literature on separate sys- reaching and grasping, whereas most of that information seems
tems for "what" and "where" does not, as we understand it, to be filtered out in the representations underlying spatial
include evidence for differences in coarseness of representation
expressions. Only the very gross geometry of the coarsest level
in the brain.
of representation of the object seems to be preserved. Even
L & J's unifying thesis, then, remains unproven. Nouns and within the domain of linguistic expressions, some observations
prepositions may very well differ in richness, as L & J have are used to derive statements about the representations of
argued, but our linguistic and mental representations of objects objects and places, whereas others are done away with as having
and places do not differ in parallel ways. Representations of both to do with historical and pragmatic issues rather than principles
objects and places may be rich or coarse, in language, in the of spatial representation (sect. 2.9, para. 3). For example, large
brain, and in the mind. vehicles (such as buses and yachts) are conceptualized either as
containers that one is in or sorts of platforms that one is on,
whereas small vehicles (such as cars and rowboats) are only
conceptualized as containers. Although this seems to suggest
that the representation of the reference object must include a

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 253

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Commentary/Landau & Jackendoff: Spatial language

Target tilted 20 deg right of vertical Target tilted 20 deg right of vertical
B

II 1, 1 |

1
1' '
1' 1 1 1
C . D

1/ 1/ i i / / s\

Target tilted 20 deg right of vertical Target vertical

Figure 1 (Wolfe). How do we talk to ourselves about orientation? It is easy to find a target that is oriented 20 deg away from vertical in
a homogeneous array of vertical items (la). If the background items are heterogeneous it is quite hard (lb) even though most of the
distractor lines differ from the target by more than 20 deg. It is hard even if there are only two distracting orientations (lc) and it is hard
even if the target is vertical (Id). We cannot ask ourselves about "20 deg" or even about "vertical."

size parameter (which can be quite metric!), L & J refer to this as ent representations of the what and where of objects. Although
a special situation that involves conventionalized concepts. I the language the authors discuss is the language between indi-
would therefore like to have a principled account of when to use viduals, we can also consider the language we use to talk to
an observation on language (or any other perceptual or cognitive ourselves. Of course, this is not a language in the normal sense
capacity) to derive predictions about the underlying representa- but rather the internal vocabulary used by one part of us to tell
tions. Or do L & J believe in an abundance of representations of another part what to do.
objects and places, one for each task to the limit? In visual perception, this internal speech is used routinely in
In general, I believe that L & J have formulated some certain visual search tasks. In a visual search task, subjects look
challenging hypotheses to be tested by perceptual research. For for a target item among distractor items. In one standard
example, the representation of a container (such as a cup) as a version, we measure reaction time (RT), the time it takes to say
thickened surface that encloses a cylindrical space (sect. 1.6, "yes" there is a target or "no" there is not. In some searches, RT
para. 1) makes its functional use or affordance (Gibson 1979) is roughly independent of the number of distractor items (set
much more explicit and its perceptual relevance might well be size). These searches make use of parallel mechanisms that can
tested with more primitive "containers" such as bowls made process all items in the display at once. Other searches appear to
from leaflets. Similarly, L & J's basic conjecture that the rela- involve serial examination of objects at a rate of about 1 item
tively simple shape specifications observed in the prepositional every 50 msec. It is widely believed that the visual system can
system reveal the extent of detail possible in object descriptions be broadly divided into a preattentive stage, in which a limited
within the "where" system (sect. 3.2.2, para. 5), has much set of basic features (color, orientation, motion, etc.) can be
intuitive appeal and seems quite testable. L & J's case would be processed in parallel across the visual field and a subsequent
much stronger, however, if they clarified some of the conceptual stage, in which more sophisticated processes can do elaborate
fuzziness identified above. tasks of object recognition, but only over a limited portion of the
field (see Treisman 1986; 1988; Wolfe 1992 for more detail).
ACKNOWLEDGMENT To be useful, these parallel processors of basic features must
The author is supported by a grant from the National Fund for Scientific be "spoken to" by other, higher processes. For example, con-
Research (Belgium). sider a visual field full of spots of different colors. You can see all
those colors all the time. If you want to determine whether this
multicolored mass contains any green spots, however, you need
Talking to yourself about what is where: to be able to ask the parallel color process to mark the green
spots in some fashion. It is certainly possible to use parallel
What is the vocabulary of preattentive mechanisms to find one color in a heterogeneous array (Duncan
vision? 1989; Wolfe etal. 1990). For present purposes, it is interesting to
compare this internal request from "you" to a feature processor
Jeremy M. Wolfe with your overt speech about objects.
Center for Clinical Cataract Research, Harvard Medical School & Brigham Our clearest data on this matter come from visual search for
and Women's Hospital, Boston, MA 02115
oriented lines. If one searches for a target of one orientation
Electronic mail: wolfe@cccr.bwh.harvard.edu
among homogeneous distractors of another orientation, search
Landau and Jackendoff (L & J) argue that our language about is very efficient with slopes of RT X set size functions near zero,
what something is and about where something is reflects differ- reflecting the contribution of parallel orientation processing

254 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Response /Landau & Jackendoff: Spatial language
10° targets are the only "steep" items 10° targets NOT the only "steep" items
B
>
\
\ • \ -A
>S i

\
^\ ^
\
\ \ / A"'\^
/ f 1 \ \ \
*-•
\
\ 1 .«•"
\ •^ \
N 0

\ ^ \
Figure 2 (Wolfe). So . . . how do we talk to ourselves about orientation? Search is quite easy if the targets possess a unique
categorical attribute ("steepness" in 2a). Even if all geometrical relations are preserved, search is more difficult if the target is not
categorically unique (2b). We can ask ourselves about "steep," "shallow," "left," and "right."

(Foster & Ward 1991). Search can become painfully inefficient real world. If we can use simple parallel processing to restrict
when one searches for the same target among multiple distrac- our attention to "steep, brown things with green parts," we can
tor orientations or even among just two flanking orientations afford to use capacity-limited "what" processes to determine
(Moraglia 1989; Wolfe et al. 1992). This is shown in Figure 1. In whether the attended item is a redwood or a maple.
la-c, the target is tilted 20 deg to the right of vertical. When Categorical terms like "steep" and "shallow" in the preatten-
there is one distractor orientation, search is easy. When there tive vocabulary probably reflect underlying perceptual limita-
are two or more, search becomes inefficient, even though the tions and not perceptual learning. There is little evidence, for
distractor orientations are quite different from the target. Our example, for an ability to learn orientation categories like "2
internal "language" apparently does not allow us to ask for "20 o'clock." It would be interesting to determine how much of the
deg" as we would ask for "red" or "green." In Id, the target is "where" aspect of spoken language is constrained by these
vertical but it is still not easy to find, suggesting that even properties of preattentive vision.
"vertical" may not be in the internal vocabulary.
What is in the vocabulary? It would be unfortunate to have an
orientation processor that only worked when faced with a
unique item on a homogeneous background, as that situation is
fairly rare in the real world. Fortunately, there is a vocabulary
with which we can talk to ourselves about the orientation of Authors' Response
lines. Like the "where" system descriptions proposed by L & J,
that vocabulary appears to be extremely restricted. Specifically,
it is restricted to categorical descriptions of the target. In
orientation, it is not possible to talk to yourself about 20 deg lines Whence and whither in spatial language and
but it is possible to ask the orientation processor for "steep," spatial cognition?
"shallow," "left," or "right." This is seen most convincingly in
stimuli like those in Figure 2. In each panel, there are four Barbara Landau8 and Ray Jackendoff3
targets. In 2a, the 10 deg targets are the only "steep" items 'Department of Cognitive Science, University of California, Irvine, CA
among +50 and - 5 0 deg shallow distractors. They are relatively 92717 and bLinguistics Program, Brandeis University, Waltham, MA 02254
easy to find. In 2b, the same 10 deg targets are the "steepest" Electronic mall: 'blandau@orion.uci.edu; bjackendofl@brandeis.bitnet
items among - 3 0 deg and +70 deg distractors. These are not
easy to find because the - 3 0 deg distractors share "steepness" We are grateful to all the commentators for their consider-
and the +70 deg distractors share "lightness" with the targets. ation and gratified that so many of them found our
Note that in both 2a and 2b, the targets are 40 deg different from argument at least partly persuasive. Considering their
one type of distractor and 60 deg different from the other. critiques, we find ourselves raising questions that are
Efficient search is possible only in 2a, however, where the target sometimes only hinted at in the commentaries them-
is categorically unique (Wolfe et al. 1992). selves. We apologize in advance for schematizing com-
In size, we have found that the "vocabulary" is limited to big mentators' arguments down to points and lines, a neces-
and small. It does not appear to be possible to search for the sary strategy for reasons of everyone else's patience; and
target of medium size (Wolfe & Bose, unpublished). In color, the we hope allowance will be made for misunderstandings on
vocabulary is probably restricted to categorical color names:
Thus, a command for "olive drab" cannot restrict search to drab
our part. In the remarks to follow, we summarize the logic
olives among olives of other shades of green and brown. Turning of our argument in such a way as to highlight our re-
to spatial relations, it is much harder to find the "red thing near sponses to the commentaries.
the vertical thing" than it is to find the "red vertical thing"
(Grabowecky & Khurana 1990; Wolfe et al. 1990). It is very hard
to find a "thing that has red and green parts" but comparatively R1. The codings relevant to our theory and the
easy to find a "red thing with a yellow part" (Wolfe & Friedman- nature of their interfaces
Hill 1992).
As with L & J's "where" language, this internal language of In asking how language and spatial representation map
preattentive commands appears to be profoundly limited. Nev- onto each other, we first sketched a model of how we see
ertheless, it seems to be rich enough to get the work done in the spatial representations interfacing with other compo-

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 255

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Response/Landau & Jackendoff: Spatial language
nents of the representational system (Figure 1 in the representation, which encodes the spatial layout of these
target article). These are the particular codings relevant to objects in the environment.
our article, and their interfaces (or partial homologies) We propose that these two codings interface by individ-
with other codings: uating objects the same way (the same number of objects
1. Linguistic representations: Most prominently for us, in both codings) and by schematizing objects in terms of
syntactic structures represent the grammatical structure one or two principal axes and a reference frame. Other
of linguistic utterances. These interface with phonological than that, they diverge: (a) Object representation elabo-
structure for access to the sensorimotor periphery and rates object shape in much greater detail and ignores
with "propositional" semantics (as described, for exam- location, (b) Object location representation (i) uses sche-
ple, by Jackendoff's 1987b conceptual structure) for ac- matized objects to help define regions and trajectories in
cess to spatial representation. terms of which other schematized objects are located,
2. Spatial representation: a modality-independent cod- ignoring much of the detailed coding of object shape that
ing in geometric form that contains the elements of spatial is represented in the object system. Object location
understanding. Some of the key elements include figure representation can also (ii) encode location in relatively
and reference objects, regions, places, paths, and trajec- coarse terms, including qualitative distinctions in dis-
tories. Mainwaring calls this system "sublinguistic," as tance and direction.
though it were on a plane of understanding lower than Notice that it is impossible to view any of these systems
language. We would prefer to think of it as a parallel as "filtered" or "unfiltered," in any absolute sense, as
system of equal status. In our target article, we assumed Mainwaring implies. Each is filtered relative to the other.
that spatial representation interfaces with the following: Object shape representation encodes schematic and de-
codings proprietary to the visual, haptic, and auditory tailed shape but nothing about location; object location
systems, so that one can use vision, touch, and audition to representation encodes schematic shape and location but
understand one's spatial environment; not detailed shape. Language encodes aspects of shape
codings involved in formulating action, so that one can but also abstract concepts such as quantification, totally
use one's understanding of spatial layout to locate objects absent in spatial representation. In fact, we are interested
or to navigate; in just what these representations have in common so that
codings proprietary to the linguistic system, in particu- an adequate interface can exist.
lar the level of meaning, so that we can talk about space. Spatial representation has separate interfaces with the
Critical to this scheme is the notion of the interface motor system and the linguistic system. This means that
between representational systems. A familiar example of different properties of spatial representation encoding
an interface can be found within the linguistic system. could in principle appear in each of these other two. As we
There, information is represented in terms of (among point out in the target article, there are great differences
other things) a phonological and a syntactic coding. These between the aspects of spatial representations that filter
codings are partially homologous, in that they both in- through into the motor system and those that filter into
volve a segmentation into words arranged in linear order. language. Hoffman points out the need to account for
On the other hand, other aspects of the two codings precision dribbling and shooting; Bridgeman describes
diverge. Phonological coding divides words into syllables elegant experimental evidence attesting to the fact that
and syllables into individual speech sounds; these further the motor system is encapsulated, at least from the
divisions do not appear in syntax. Syntactic coding assigns processes involved in explicit perceptual judgments (see
to each word a part of speech (noun, verb, etc.) and groups also Ingle).
words into larger syntactic constituents (noun phrase, Quite a number of commentators (Hoffman, Tversky &
verb phrase, etc.); these features are not reproduced in Clark, Bridgeman) object that we have neglected the
phonological coding. precision required by the motor system. In section 3.1,
Such partial homologies among distinct codings consti- however, we point to the precision of the motor system,
tute the vocabulary they have in common. They are and we conjecture that the interface between spatial
therefore the only means by which different codings can representation and the linguistic system filters out metric
communicate with one another. For example, the phono- precision. As Figure 1 of our article shows, such precision
logical and syntactic codings of language are kept in synch and filtering are a natural possibility of the model as
by virtue of the correspondences in the linear order of presented and it becomes even more natural with possi-
words. (To put this in the terms of current neuroscience, ble additional structure as suggested to us by Brown
the partial homologies are the informational links used in (whose commentary we discuss later).
binding different codings together.) Because the homo-
logy is only partial, though, the interface behaves like a R2. The logic of our argument
filter: Only certain aspects of each coding can filter
through to the other. To bring linguistic evidence to bear on spatial representa-
The starting point of our target article is that there tion, we made use of a basic premise: If linguistic judg-
needs to be a partial homology between spatial represen- ments can be based on spatial properties of objects then
tations and language that enables us to talk about what we the information involved in the linguistic judgment must
see (or touch, or hear). be able to pass through the interface between spatial
At the end of the paper, we explore the possibility that representation and language.
there are at least two distinct components of the spatial To pass through the interface, the information must be
representation system: (1) object shape representation, available in spatial representation. Thus, for example, if
which encodes the three-dimensional shapes of individ- language can characterize a "hole" as wide or narrow,
ual objects in the environment, and (2) object location having sides, top, and bottom, spatial representation had

256 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Response/Landau & Jackendoff: Spatial language

better contain an entity with which the term "hole" can be More central to our target article, the variety of ways in
associated, and this entity had better have properties which language describes the spatial layout of stationary
relevant to determining width or picking out sides, top, and moving objects requires us to include in spatial
and bottom. representation the abstract notions of region and trajec-
This is a point on which we believe Slobin misses our tory, which are not directly present in perceptual input.
argument. He writes that object names do not refer to The richness of linguistic description sets a lower bound
shapes of objects; rather, they refer to the objects. We on the richness of spatial representation: We must have at
agree. The issue of shape enters in the following way: In least those elements. There may, of course, be more.
agreeing or disagreeing with the statement, "This is an The second aspect of our argument is that the primitive
X," where "X" is an object name, one has to evaluate the notions found in the spatial language for location have no
referent of "this." If this evaluation is carried out visually, evident explanation in terms of language itself. They do,
then, as we have pointed out, one of the crucial factors however, have an obvious explanation in terms of geomet-
must be its shape. We conclude from this that names for rical notions that appear altogether natural in spatial
physical objects must be mentally linked to, among other representation. Such notions include, for example, the
things, spatial encodings of their shapes. Therefore, spa- figure/ground opposition, the schematic axes of objects in
tial representation must be detailed enough to distinguish different frames of reference (viewer-centered and
all the kinds of objects that can be distinguished by shape. object-centered), and the notion of relative distance from
Should every distinction appearing in language be an object (interior vs. contact with surface vs. proximal vs.
recoverable from the visual image? (Here we distinguish a distal). Because these notions seem quite natural exten-
visual image from the amodal spatial representation sions of many current approaches to object and place
linked to it.) Wagemans points out that parts such as top encoding, we claim that the degrees of freedom found in
and bottom of an object may not be recoverable from the language are homologous to parallel degrees of freedom
image, and this prompts him to ask whether our "en- in the description of spatial representation. That is, we
riched" representations are perceptual or not. Actually, are claiming that the partial homology between language
top and bottom (differentially marked ends) can often be and spatial representation includes at least these notions.
recoverable as the opposite ends of the object's vertical No commentator seems to object seriously to this part of
axis. Exactly which end is the top and which is the bottom the argument.
will also often correspond to information recoverable in The more controversial part of our argument was based
the image (e.g., Fillmore [1975] suggests that the front of on the extremely limited number of parameters involved
an animate object is determined by the positioning of its in distinguishing preposition meanings in conceptual
principal perceptual organs). structure, resulting in such a small number of preposi-
More broadly, however, we see spatial representation tions. Why should there be so few? We offered two
as a relatively central coding that draws on - but is not possibilities. The Design of Language Hypothesis attri-
equivalent to — information from all the perceptual sys- butes the limitations of language to the interface: In
tems. Our point is that the distinctions uncovered by present terms, spatial representation has a much richer
language must be represented within the spatial repre- selection of parameters, but only a limited number of
sentational system. Some of these properties can be them enter into the homology with language. The Design
mapped from the visual image, for example, the object's of Spatial Representation Hypothesis says that the inter-
principal axis. Others cannot be recovered from the visual face is in fact pretty much as rich as it can be, and that the
image alone. For example, there is nothing in the image limitation actually lies in spatial representation itself.
(that we can think of, anyway) that would correspond to Note that with each of these hypotheses we explicitly
the spatial notion of region (or, in fact, to the principal axis reject the notion of isomorphism between language and
of a nonrotating sphere). But each of these - whether or spatial representation, despite Friederici's suggestion
not it is mappable from the image - must be encoded in that we do claim such an isomorphism.
spatial representation. It is an empirical problem to We are the first to admit that (as Hoffman points out)
determine which elements of spatial representation can there is nothing in the linguistic evidence per se that can
filter through interfaces all the way up from (or down to) decide between these possibilities, or, as we suggested,
the image, which elements can be linked to some inter- some combination of the two. What intrigues us is that,
mediate form of representation (e.g., Marr's [1982] 2£-D with the exception of metrical precision, the location
sketch), and so forth. In other words, the nature of the system in language has just about the right properties to
interface between spatial representation and strictly vi- interface with the "where" system described by neuro-
sual representation is an issue for further research in science. We therefore conjecture (a weaker term than
computational vision (and similarly for the other modal- "claim") that what we find in the language of places has a
ities that interact with spatial representation). fairly strong homology with the coding of objects and
Following from our basic premise, we argue two points. places situated in the "where" system of the brain. In
First, the richness of spatial language points to a richer particular, one kind of object description gets through the
conception of spatial representation than is usually imag- interface between spatial representation and language for
ined by visual theorists. For example, terms like "hole" naming (the "whats"), and another kind of object descrip-
require us to include in spatial representation not just tion does so for locating the "wheres."
objects and their parts, but also "negative parts"; terms We are pleased that quite a number of commentators
like "top," "side," and "front" require us to include sche- saw our hypothesis as a good start, or better yet, as
matizations of objects in terms of orienting axes; terms plausible, even providing converging evidence. Bryant
like "edge" require us to include schematizations of ob- agrees with our general approach to the mapping of space
jects in terms of thickened bounded planes. and language, emphasizing that the representations de-

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 257

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Response/Landau & Jackendoff: Spatial language
rived from larger linguistic units - discourse - also map circularity in using psychological evidence to help
onto spatial representations. Wolfe provides evidence for sharpen our hypothesis, apparently not understanding
categorical coding, parallel to that in language, involved the notion of converging evidence. She also feels that our
in preattentive processes, for example, distinctions be- empirical data are not convincing, as they only show
tween steep and shallow, or large and small. He suggests absence of shape representation when filtered by a spatial
that the rough search accomplished with these processes preposition task (Landau & Stecker 1990). But that is
can efficiently support the more effortful, capacity- precisely the point: In our view, it is the preposition
limited search for more detail as is needed in the identi- context that elicits the kind of representation that ignores
fication of particular object kinds. We regard such evi- detailed object shape.
dence as critical in illustrating the existence of categorical Ingle tells us that we should compare object and spatial
coding schemes derived from empirical studies quite (locational) vision not in terms of their implicit complexity
different from our own. Similarly, Brown provides con- but in terms of which representations normally play a role
verging evidence for categorical coding, although she in conscious deliberation. We see no reason for this
argues that it is a characteristic of the left hemisphere restriction. Both aspects must be addressed: what infor-
rather than language per se (more later). Feldman agrees mation the mind processes, and how this information
that certain nonmetric regions will be more naturally contributes to conscious experience.
represented than others, although he proposes a func-
tional reason why (again, more later).
A number of commentators disagree with our meth- R4. Our approach to the linguistic representation
odology, our approach to the linguistic description of of objects and places
objects and places or our approach to spatial representa-
tion. In what follows, we will do our best to answer their R4.1. Objects. It is surprising that we did not have many
objections. complaints about our emphasis on the importance of
shape in object naming. We say it is surprising because we
have often encountered versions of this objection during
R3. Our interdisciplinary methodology colloquia based on this material As an anonymous BBS
referee put it,
The basic stance of our target article is that evidence from Space enters only very indirectly into naming. Terms
language can provide insight into the structure of human like "dog" or "clothespin" seem to apply to kinds of
spatial cognition that can extend our understanding as objects, regardless of what it is that determines mem-
derived from standard psychological and neuropsycho- bership in a kind. . . . Spatial representations dis-
logical techniques. Our assumption is that linguistic an- cussed here may enter directly into visual object recog-
alyses should dovetail with and supplement these other nition, but they seem to enter only weakly and
approaches. As with any scientific problem, all available derivatively into naming.
tools should be brought to bear on spatial cognition. The In just this way, Slobin objects to our emphasis on
analysis of spatial language vis-a-vis nonlinguistic spatial object shape as an important diagnostic for the variety of
representations seems a fruitful approach. human object categories expressed by language. We reit-
A number of commentators seem to have misun- erate that shape is involved in naming to the extent that
derstood our methodology. Brown, coming from the one must establish by perceptual means the truth value of
viewpoint of vision research, criticizes us for using tools some statement, "This is an x." We believe that Slobin
that are specific to the left hemisphere, given that spatial misrepresents our position when he objects that the word
cognition involves both hemispheres. This criterion "frog" does not describe the shape of a frog, but rather
would exclude the use of linguistic evidence altogether. references the object: "When I say frog,' I do not describe
We agree with Brown that right hemisphere evidence is the shape of a frog, but evoke whatever representation of
equally important in providing an overall resolution of the frog you have in your mind." We agree. And, although we
problem, and in fact her comments about hemispheric showed in sections 1, 1.1, and 1.2 that shape recognition
specialization, if correct, add an important clarification to is crucial for object naming, we do not mean to exclude
our position (see below). But at the same time, we think it the other criteria Slobin mentions (see Note 3 in our
is critical that the hemispheric differences she describes target article for explicit acknowledgment of these).
are found in the context of different linguistic instruc- Once again, our concern is the mappings between
tions: Left hemisphere superiority is found with instruc- spatial cognition and language. We are therefore forced to
tions using locations encoded by English prepositions deal with the fact that object shape is a key index of object
such as "whether a dot is above or below a line," whereas kind. The importance of other factors should not diminish
right hemisphere superiority is found with locations en- the significance of object shape. Likewise, the impor-
coded by phrases, such as "whether a dot is within 3 tance of object shape in naming object kinds does not
millimeters of a line." Analysis of the language itself automatically mean that shape should or should not count
suggests the same kind of differential encoding of space for other natural concepts. For example, at the moment,
(categorical in the "left hemisphere" instructions, metric there is no independent theory of what makes crumpling
in the "right hemisphere" instructions) that Brown de- perceptually different from crushing (to use Slobin's ex-
scribes. Thus, evidence from language and visual cogni- ample); but perhaps someday there will be, in which case
tion complement each other, as we believe they should. we would want to draw on those representations in estab-
Other commentators seem to misinterpret our ap- lishing links to those verbs. In contrast, there will never
proach from the other end, seeing it more as an isolated be a theory of what makes a wife perceptually different
exercise in linguistic description. Friederici accuses us of from a mistress (impossible, as Slobin observes), and we

258 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Response/Landau & Jackendoff: Spatial language

have no intention of claiming that object shape — or any pressed in language, one that has a distinctly functional
perceptual property - should distinguish these. flavor. (We actually anticipated this objection in our dis-
cussion of the Design of Language Hypothesis.) The idea
R4.2. Places. There were far more objections to our is that spatial representation of location is very rich, but
linguistic treatment of location, in particular, to our use of that spatial language is only detailed enough to get the
English prepositions as the domain upon which to build message across - the rest of the detail is taken care of by
our theory. It is curious that the commentators objected the visual system or by pragmatic understanding of the
for very different (often incompatible) reasons. Some situation. This way, language can be processed on line
commentators (Tversky & Clark, Slobin, Bennett) objec- more efficiently.
ted that the description of location is actually infinitely There are two problems here. First, although Slobin
detailed and that our restriction to prepositions (English suggests that pragmatics derived from the scene can take
yet!) had the effect of artificially restricting the range of care of disambiguating the spatial relationship, we believe
meanings that could be expressed. A partially overlapping he wildly overestimates the usefulness of such prag-
set of commentators (Bridgeman, Bryant, Bennett, Feld- matics, especially for child learners. He says, "The words
man, Friederici, Heidorn & Hirtle) objected that there 'frog' and jar' already tell you a good deal about the
are not more prepositions just because there are no more possible relations between these two objects." They do?
spatial relationships to be described! They say, in effect, The frog could be in the jar, but it could also be on the jar,
"Of course there should not be many spatial relations in next to the jar, behind, below, or above the jar. . . and
language. There are not many spatial relations, period. these are only the lexically encoded concepts! What about
All you need is a frame of reference that establishes three 3" from the left front of the jar, 2.5 mm. into the jar? Or it
spatial axes, and you have everything you need." might have just its head and its left hind leg in the jar and
Considering the latter objection first, we think such the rest out. Although pragmatics of the sort Slobin
limitations on possible spatial relations are actually not so invokes might help adults who already know how the
self-evident. First, as the study of geometry shows, nu- language encodes spatial relationships, such information
merous spatial relationships are possible that simply do could not possibly allow the learner to narrow down the
not get encoded by the basic vocabulary in English or target meaning on the basis of observation alone. On the
other languages. (All you have to do is look in a not-so- other hand, perhaps Slobin means to suggest that humans
elementary geometry book to discover that there are a are predisposed to represent only certain spatial relation-
number of formally established geometries - e.g., Eu- ships by language, and so, given that predisposition, one
clidean, similarity, projective, topological - that form a can more readily narrow in on some target relationship. If
hierarchy, with Euclidean containing the finest-grained so, we definitely agree.
metric information.) The second problem concerns the very notion of rely-
Although there are many well-formed geometric prop- ing on functional explanations of why we have so few
erties and relationships, however, not all of them are prepositional meanings. We do acknowledge functional
lexically encoded. For example, there is no word for "3 requirements as a possibility, but we are concerned that
inches away from" despite the fact that this is a well- this kind of argument raises the danger of begging the
formed spatial relationship in metric geometry. On the question: How do you tell what counts as "enough detail,"
other hand, there are words that represent nonmetric especially when it varies from language to language (as
relationships, such as contact or attachment ("on"), con- Slobin and Tversky & Clark point out)? As an example,
tainment ("in"), and so on. Our question is: Why should Bennett suggests that "in front of" is natural, whereas "at
language express (as single morphemes) the particular the thick end of" is not, just because of the general
relations it does, some of which (e.g., "along" and usefulness of the former. But why should one be useful
"throughout") are fairly eccentric, while not expressing and the other not? What could "usefulness" possibly
others (e.g., our hypothetical relations "plin," "sprough," mean, in this context? Our own explanation of this partic-
and "betwaft") that are prima facie no more complex? ular case has nothing to do with usefulness; "at the thick
Our conjecture is that the relations we find in language end of" is unnatural as a lexicalized spatial relation be-
are built up from an exceedingly sparse primitive vocabu- cause, like our hypothetical example "sprough," it refers
lary that happens to allow the former relations but not the to the shape of a geon.
latter. The particulars of this primitive vocabulary are We are interested in why these particular limitations
empirical issues, which is why we devoted more than half exist and not others. The Design of Spatial Representa-
of our target article to running through the boring details tion Hypothesis strikes us as an interesting direction to
of English prepositions. Precisely because we have this pursue. It may be wrong, but the potential payoff if it is
internal vocabulary, it seems self-evident that there are right encourages us to stick our necks out.
not many spatial relations "in the world." We are re- Now back to the first objection: Why did we focus on
minded of a passage from Fodor (1980, p. 333): English prepositions in the first place? Maybe it is really
From in here it looks as though we're fit to think PPs (prepositional phrases) of potentially infinite com-
whatever thoughts there are to think. . . . It would, of plexity that describe location. First of all, can you imagine
course, precisely because we are in here. But there is how long the target article would have been if we had
surely good reason to suppose that this is hubris bred of described the semantics of English PPs?
an epistemological illusion. No doubt spiders think that More seriously, our basic argument was that it is possi-
webs exhaust the options. ble to use evidence from language to provide insights into
Other commentators (Bennett, Heidorn & Hirtle the structure of human spatial cognition. We therefore
again, Slobin) propose a somewhat different (though re- examined English prepositions not because we believe
lated) reason why there are so few spatial relations ex- they are isomorphic to spatial representation, but be-

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 259

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Response/Landau & Jackendoff: Spatial language

cause, after carefully filtering out the nonspatial senses mixture balances out empirically, and how it should bal-
(an onerous task in the linguistic literature that we have ance out in principle. For a simple example, there seems
spared BBS readers), they provide a relatively pure ex- no principled explanation of why in American English one
pression of spatial relations. We argued at length that the makes a decision but in British English one takes a
meaning of a spatial preposition - its representation in decision. It is just a learned fact of language. In the area of
terms of "propositional" structure - is a function that prepositions alone, such works as Herskovits (1986) and
maps the reference object, canonically expressed by the Vandeloise (1986) are sizable studies concerned with this
object of the preposition, into a region. The verb of the question. So, we suggest, there can at the moment be no
sentence in turn locates the figure object in this region. principled answer to Wagemans's question. In our target
(Incidentally, although Friederici supposes that "preposi- article we attempted to extract what we judge to consti-
tions are considered to encode relations similar to formal tute the regularities; we acknowledge there may be dis-
logical relations," we know of no formal logical relations agreement. We think, though, that the general order of
that behave like prepositions.) complexity in the meanings of spatial prepositions is not
Bennett and Tversky & Clark observe that the preposi- too much greater than our analysis suggests.
tion alone cannot specify location; it is the preposition
plus its object that specifies location. As just pointed out,
we agree. The preposition expresses only a function. Our R5. Our approach to spatial representation
question is, What properties of the reference object can
play a role in such functions, so that they serve to differen- A number of commentators wonder whether our charac-
tiate the resulting regions? Our answer is: very few. terization of spatial representation is altogether on the
More technically, we disagree with Tversky & Clark's mark. Deregowski suggests that object representations
assertion that count nouns such as "the table" can repre- are actually more "flattish" than three-dimensional. Un-
sent places. Canonically, it is the prepositional phrase that likely, we think, but if so, we do not see what the
represents the place (they call the PP a "locative predi- consequences are for explaining how we talk about what
cate"; for detailed arguments against this common ap- we see. Tarr proposes that a 3-D representation can be
proach, see Jackendoff 1983, Chapter 4). We do agree dispensed with, and our analysis purged of "ad hoc"
with them that the meanings of many nouns and verbs elements, by adopting a "multiple egocentric views"
include locative relations; however, formal analysis (Jack- theory of object representation. Feldman proposes
endoff 1983; 1990; Talmy 1980) discloses that such locative that location is represented in terms of potential causal
relations are of the same sort as we have described, relationships. Corballis suggests that spatial represen-
namely, mapping objects into regions in terms of which tation is an analog medium, more like Play-Doh than a
other objects can be located. For example, the verbs Lego set.
"enter" and "exit" encode the two different directions in A number of commentators worry about our hypoth-
and out, each of which has the spatial properties we eses of how spatial representation and language connect.
describe in our inventory of English prepositions. Hoffman, Bridgeman, and Tversky & Clark are con-
Thus, when Bennett says it is a mistake to concentrate cerned that we have ignored the precision metric coding
on prepositions and Slobin points out the unilluminating evident in the motor system. They worry that if spatial
quality of the preposition systems in Turkish and Tzeltal, representation were as coarse as our analysis of language
we think they miss the point. We agree that a full treat- indicates, there would be no way to account for the
ment of this topic would deal with the total variety of basketball stars of the world (not to mention the ability of
expressions for spatial location in all the languages of the 4-month-olds to "catch" moving targets - see our sect.
world. We have just found English prepositions a conve- 3.1). Brown suggests that the phenomena we describe can
nient and telling entree into the subject. No doubt addi- be explained by the properties of left versus right hemi-
tional rigorous studies will provide information to sphere processing.
sharpen or modify our hypotheses. We have certainly As to the general componential framework for objects
provided enough references to languages other than En- we have adopted, we agree that it is not the only possible
glish and to parts of speech other than prepositions to see way to characterize high-level vision. We submit, how-
where further investigation might lead. And we welcome ever, that any alternative theory must address the tradi-
detailed theoretical descriptions of spatial meanings that tional problems of object constancy and spatial layout; in
differ from our own. For example, Tzeltal is often cited as addition, it must now connect with the linguistic data of
a language characterized by geometrically rich spatial the sort we have presented. We are open to detailed and
terms (which would not fit comfortably into our scheme). comprehensive alternatives.
A careful reading of Levinson's (1992) analysis, however, In particular, we find Tarr's view (on the face of it)
shows that the system of terms fits remarkably well. insufficient: (1) Multiple egocentric views of single ob-
A separate issue is the important methodological ques- jects presuppose particular viewing distances and lighting
tion that Wagemans raises regarding our linguistic anal- conditions and must therefore be linked somehow to
ysis. As we observed, languages often have idiosyncrasies represent the same object over time and space. In any
of prepositional use that are not well predicted by our case, we do not get the punch line in his objection to our
analysis. For example, in English one gets in cars but presentation. Would adopting his approach make any
either in or on buses. He requests a principled way of substantive difference to our theory? (2) Because haptic
determining which uses are regular and which idiomatic. representations are clearly not based on the same surface
Our reply here is of necessity a hedge. Language is a characteristics as visual ones, Tarr's proposal provides no
complex mixture of rule-regulated and idiomatic ele- way to interface with the sense of touch, as our approach is
ments; much dispute among linguists concerns how this designed to do. (Admittedly, none of the other specific

260 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 Response/Landau & Jackendoff: Spatial language

visual approaches is currently designed to work for hap- Ingle), including object-focused tasks such as reaching
tics.) (3) It is not clear how multiple egocentric views of and grasping a bottle versus a bowl and environment-
individual objects can be integrated into a representation focused tasks such as navigating through a town. We
of spatial layout. (4) In particular, an observer anticipating suggested that the Design of Language Hypothesis re-
his own path through a scene requires a representation in quires that all lexical items level out some degree of
which the scene remains constant while the observer geometric detail and that the Design of Spatial Represen-
moves - clearly not an egocentric representation in Tarr's tation Hypothesis further levels out detail in the loca-
sense at all. tional system. These two means of leveling result in an
Finally, we do not find our proposed augmentations of asymmetry in the detail with which objects are encoded
the componential framework especially ad hoc; this is when talking about "what" versus "where."
evidently a matter of taste. For example, without using We acknowledge that a problem still remains: how to
something equivalent to our feature "hollow" versus integrate metric properties with the coarse coding, de-
"solid," how does Tarr represent the understanding of a pending on the task. For example, although we propose
closed box as hollow, and therefore potentially containing that objects are represented as points and lines in the
other objects? In short, although we find Tarr's approach "where" system, it is obvious that more detailed geomet-
potentially useful as a theory of the highest level of purely ric information would be needed to grasp accurately a
visual representations of shape (though not spatial layout), Ming vase versus a bowling ball. One possibility is that
it can hardly be satisfactory as a comprehensive theory of the "where" system is drawn on for the general trajectory
spatial cognition. of one's reach, whereas the "what" system is drawn on for
Feldman's causal approach to location has some of the the grasp. Another is that the "whats" and "wheres" are
same problems (although we find the basic thrust of his described in two different ways, both coarsely and in fine
approach very interesting). We observed that some prep- detail, and that these are drawn on differentially for
ositional meanings incorporate forces (e.g., "run into the different tasks.
tree" and "lean against the wall"). But in general, a A version of the latter hypothesis is offered by Brown; it
preposition like "in" expresses a spatial dependency, not a is essentially a more elaborate version of the Design of
causal one. Food is not in the refrigerator for the same Spatial Representation Hypothesis. The idea is that the
reasons that the comets are in the solar system or the fish linguistic system for location is closely homologous to
are in the sea. The objects along the sight of a rifle are in spatial coding by the left hemisphere "where" system,
different peril from that of houses along the beach or the which, like the linguistic coding, lacks metrical informa-
pedestrians, trees, and tunnels along the road. What tion. All metrical information is coded in the right hemi-
these cases share are their spatial relations; the various sphere "where" system. That is, the loss of metrical
causal relations follow as inferences from knowing about precision and the introduction of discrete categories of
the objects thus spatially related. And in any case, we location are properties of the interface between the left
wonder what independent means Feldman intends to use and right hemisphere spatial systems, not of the interface
to discover which spatial relationships are causally related between the spatial system and language. Tarr suggests a
and which are not. It seems, from our vantage point, that similar segmentation of spatial representation, although
his theory depends quite strongly on analyses such as ours without proposing brain localization. Bridgeman suggests
to discover which of the locations are causally related. It a division of spatial cognition into "sensorimotor" and
seems to us that these are just the ones that are lexically "cognitive," the former incorporating metrical informa-
encoded. tion, the latter, not. If such further divisions of spatial
Insofar as we can understand Corballis's nonsymbolic cognition come to be supported independently, it will be
"unGADly" system, we do not see how it suffices for the a stunning result for us, as it will bring the linguistic facts
tasks the spatial system must perform. Play-Doh cannot even closer to those of spatial understanding.
structure itself with precision. It requires an external We will end with a serious problem for our approach
agency with precise intentions to mold and interpret it. It that has been raised by Olson, Tarr, and Tversky & Clark,
has no internal structure over which similarity metrics can as well as by many colleagues at colloquia where this
be defined; there is no way to establish a homology to material has been presented. The language used to ex-
other representations so that they can communicate with press the relation of parts to objects is identical to that
each other. In adopting something like the Marr (1982) used for configurations of independent objects. To adapt
theory of spatial representation, we have assumed that one of Olson's examples, we speak both of a nose on one's
spatial representation contains some more or less analog face and a fly on one's face. If, as we have claimed, the
degrees of freedom: the exact shape of generalized cones "where" system is implicated in the latter case, then the
and the exact proportions, orientation, and curvature of linguistic evidence suggests it is implicated in the former
axes. On the other hand, the degrees of freedom we have as well. This in turn appears to undermine our claim that
been most concerned with, such as the number of axes the "what" system encodes all the details of object shape
and the attachment of parts, are combinatorial in the (here the shape of a face), including their decomposition
usual sense. So in a way, we (like Marr) have weakened the into parts. How does one reconcile our analysis with this
strict analogue/digital distinction that Corballis and pos- evidence?
sibly Mainwaring assume.
We see three possible solutions. One is just to abandon
As for the existence of metric precision in the motor the connection to the "what" and "where" systems and fall
system, we agree that this is a critical part of the puzzle. back on the Design of Language Hypothesis, as many of
Spatial representation must clearly have adequate geo- the commentators have urged us to do. Even in the face of
metric power to account for metric capabilities that un- such a retreat, our evidence would still leave ample
derlie motor tasks (while remaining "hidden," following implications for the organization of spatial cognition. We

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 261

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 References/Landau & Jackendoff: Spatial language

are reluctant to give up so easily, however, for reasons Biederman, I. & Cooper, E. E. (1991) Priming contour-deleted images:
already detailed in the present reply. Evidence for intermediate representations in visual object recognition.
Cognitive Psychology 23(3);393-419. [MJT]
A second possibility is to take the bull by the horns and Binford, O. B. (1971) Visual perception by computer. Presented at IEEE
claim that the "where" system indeed encodes the rela- Systems, Science, and Cybernetics Conference, Miami, FL. [aBL]
tion of object parts to the body of an object. As we Bornstein, M. (1985) Color-name versus shape-name learning in young
understand it, this is a problem being encountered by all children. Journal of Child Language 12:387-93. [aBL]
who attempt to understand the cortical representation of Bowerman, M. (1989) Learning a semantic system: What role do cognitive
predispositions play? In: The teachability of language, ed. M. L. Rice &
objects (e.g., Kosslyn 1990). This encoding could be R. C. Schiefenbusch. Brooks. [aBL]
redundant with the encoding of the "what" system, or it (1991) The origins of children's spatial semantic categories: Cognitive vs.
might even supplant it. In the latter case, the details of linguistic determinants. In: Rethinking linguistic relativity, ed. J. J.
the "what" system would be devoted primarily to a de- Gumperz & S. C. Levinson. Cambridge University Press
(Cambridge). [aBL]
tailed description of the geons making up objects, with
Braunstein, M. L., Saidpour, A. & Hoffman, D. (1992) Interpolation in
the "where" system putting the parts together. That is, structure from motion. Perception ir Psychophysics 51(2):105-
the articulation of object shape might involve as elaborate 17. [DDH]
an interweaving of "what" and "where" information as the Bridgeman, B. (1981) Cognitive factors in subjective stabilization of the visual
description of spatial layout does. world. Ada Psychologica 48:111-21. [BB]
(1986) Multiple sources of outflow in processing spatial information. Acta
A third possibility is that the "what" system operates Psychologica 63:35-48. [BB]
more or less the way the componential approaches have (1991) Complementary cognitive and motor image processing. In:
envisioned, with objects articulated into connected Presbyopia research: From molecular biology to visual adaptation, ed.
geons. When one specifically wishes to mention or other- G. Obrecht & L. Stark. Plenum Press. [BB]
Bridgeman, B., Kirch, M. & Sperling, A. (1981) Segregation of cognitive and
wise draw attention to an object part, however, it is motor aspects of visual function using induced motion. Perception ir
necessary to regard it as a separate object with its own Psychophysics 29:336-42. [BB]
location. Under this condition, when the part is regarded Bridgeman, B., Lewis, S., Heit, G. & Nagle, M. (1979) Relationship between
as figural, its shape still falls under the "what" system, but cognitive and motor-oriented systems of visual position perception.
its location falls under the "where" system; hence the Journal of Experimental Psychology: Human Perception and Performance
5:692-700. [BB]
usual prepositional expressions apply. We do not think
Bridgeman, B. & Stark, L. (1979) Omnidirectional increase in threshold for
evidence from language alone can distinguish among image shifts during saccadic eye movements. Perception & Psychophysics
these three possibilities, or any others that might be 25:241-43. [BB]
developed. We leave the resolution to further research in Brugman, C. (1981) Story of over. Indiana University Linguistics
other paradigms. Club. [aBL]
Bryant, D. J. & Tversky, B. (1991) Locating objects from memory or from
If nothing else, the linguistic evidence is instrumental
sight. Paper presented at the 32nd annual meeting of the Psychonomic
in pointing out the distinctions among these hypotheses Society, San Francisco, CA. [DJB]
and demonstrating what is at stake. In this respect we Bryant, D. J., Tversky, B. & Franklin, N. (1992) Internal and external spatial
stand firm in our original argument - that language can frameworks for representing described scenes. Journal of Memory and
and should play a role alongside vision in our understand- Language 31:74-98. [DJB]
ing of the organization of spatial cognition. Bulthoff, H. H. & Edelman, S. (1992) Psychophysical support for a two-
dimensional view interpolation theory of object recognition. Proceedings
of the National Academy of Sciences USA 89:60-64. [MJT]
ACKNOWLEDGMENTS Carey, S. (1985) Conceptual change in childhood. Bradfood Books/MIT
Preparation of this commentary was supported by the granting Press. [aBL]
agencies cited in our target article as well as by the University of Chwistek, L. (1960) Wieloid rzeczywistoici w sztuce. Warsaw:
Pennsylvania Institute for Research in Cognitive Science and Czytelnik. [JBD]
Cienki, A. (1988) Spatial cognition and the semantics of prepositions in
Department of Psychology, who jointly sponsored a Visiting
English, Polish, and Russian. Ph.D. dissertation, Brown
Professorship for Barbara Landau during 1992-93. University. [aBL]
Clark, E. V. & Clark, H. H. (1979) When nouns surface as verbs. Language
55:767-811. [BT]
Clark, H. (1973) Space, time, semantics, and the child. In: Cognitive
development and the acquisition of language, ed. T. Moore. Academic
Press. [aBL]
References Conley, G. (1985) Theories of pictorial representation: Goodman's realism and
the similarity theory. Ph. D. dissertation, Department of Philosophy,
Letters a and r appearing before authors' initials refer to target article and University of Minnesota. [JBD]
response respectively. Cooper, L. (1989) Mental models of the structure of visual objects. In: Object
Anderson, S. & Keenan, E. (1985) Deixis. In: Language typology and perception: Structure and process, ed. B. Shepp & S. Ballesteros.
syntactic description, ed. T. Shopen. Cambridge University Press Erlbaum. [aBL]
(Cambridge). [aBL] (1992) Probing the nature of the mental representation of visual objects.
Angeli, S. J., Murray, E. A. & Mishkin, M. (1988) The hippocampus and Seminar on Cognitive Neuroscience, session IV: Memory. American
place memory in rhesus monkeys. Society of Neuroscience Abstracts Association for the Advancement of Science. [aBL]
14:232. [aBL] Corballis, M. C. (1991) The lopsided ape: Evolution of the generative mind.
Au, T. K. & Markman, E. M. (1987) Acquiring word meanings via linguistic Oxford University Press. [MCC]
contrast. Cognitive Development 2:217-36. [aBL] (1992) On the evolution of language and generativity. Cognition 44:197-
Bavin, E. (1990) Locative terms and Warlpiri acquisition. Journal of Child 226. [MCC]
Language 17:43-66. [aBL] Coucelis, H., Golledge, R. G., Gale, N. & Tobler, W. (1987) Exploring the
Beck, B. B. (1980) Animal tool behavior: The use and manufacture of tools by anchor-point hypothesis of spatial cognition. Journal of Environmental
animals. Garland STPM Press. [MCC] Psychology 7:99-122. [aBL]
Bennett, D. (1975) Spatial and temporal uses of English prepositions: An DeLoache, J. S. (1987) Rapid changes in the symbolic functioning of very
essay in stratificational semantics. Longman Press. [aBL] young children. Science 238:1556-57. [aBL]
Biederman, I. (1987) Recognition-by-components: A theory of human image DeLoache, J. S., Strauss, M. & Maynard, J. (1979) Picture perception in
understanding. Psychological Review 94(a):115-47. [aBL, MCC, MJT] infancy. Infant Behavior and Development 2:77-89. [aBL]

262 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 References/Landau & Jackendoff: Spatial language.

Deregowski, J. B. (1978) On examining Fortes' data: Some implications of Greenfield, P. M. (1991) Language, tools and brain: The ontogeny and
drawings made by children who have never drawn before. Perception phylogeny of hierarchically organized sequential behavior. Behavioral and
7:479-84. [JBD] Brain Sciences 14:531-95. [MCC]
(1980) Illusions, patterns and pictures. Academic Press. [JBD] Greenfield, P. M., Reich, L. C. & Olver, R. R. (1987) On culture and
(1990a) lntercultnral search for the origins of perspective. In: Contemporary equivalence: II. In: Studies in cognitive growth, ed. J. S. Bruner, R. R.
issues in cross-cultural psychology, ed. N. Bleichdrodt & P. J. D. Olver & P. M. Greenfield. Wiley. [JBD]
Drenth. Swets & Zeitlinger. [JBD] Greeno, J. G. (1992) The situation in cognitive theory. Paper presented to the
(1990b) On two distinct and quintessential kinds of pictorial representation. American Psychological Society, San Diego, CA, June. [SDM]
In: Ecological perception research, visual communication and aesthetics, Greeno, J. G. & Moore, J. L. (in press) Situativity and symbols: Response to
ed. K. Landwehr. Springer-Verlag. [JBD] Vera and Simon. Cognitive Science. [SDM]
(1990c) Traffic signs as embodiments of characteristic views. Zeichen Greeno, J. G., Smith, D. R. & Moore, J. L. (in press) Transfer of situated
(Theorie) in der Praxis. Paper presented at the Biennial Congress of the learning. In: Transfer on trial, ed. D. Detterman & R. Sternberg.
German Semiotic Association, Passau. [JBD] Ablex. [SDM]
Deregowski, J. B., Muldrow, E. S. & Muldrow, W. F. (1972) Pictorial Grice, H. P. (1975) Logic and conversation. In: Syntax and semantics, vol. 3:
recognition in a remote Ethiopian population. Perception 1:417— Speech acts, ed. P. Cole & J. L. Morgan. Seminar Press. [PBH]
25. [JBD] Gruber, J. (1976) Lexical structures in syntax and semantics. North-
Downs, R. M. & Stea, D. (1973). Image and environment. Aldine. [aBL] Holland. [aBL]
Duncan, J. (1989) Boundary conditions on parallel processing in human vision. Hawkins, B. W. (1984) The semantics of English spatial prepositions. Ph.D.
Perception 18:457-69. [JMW] dissertation, Department of Linguistics, University of California at San
Duncker, K. (1929) Uber induzierte Bewegung. Psychologische Forschung Diego. [aBL]
12:180-259. (Translated and condensed in Ellis, W. [1950] Source book of Heibeck, T. H. & Markman, E. M. (1987) Word learning in children:
Ccstalt psychology. Humanities Press.) [aBL] An examination of fast mapping. Child Development 58:1021-
Dziurawiec, S. & Deregowski, J. B. (1992) Twisted perspective in young 34. [aBL]
children's drawings. British Journal of Developmental Psychology 10:35- Hein, A. & Jeannerod, M. (1983) Spatially oriented behavior. Springer-
49. [JBD] Verlag. [aBL]
Easton, R. D. & Bentzen, B. L. (1987) Memory for verbally presented routes: Held, R. (1968) Dissociation of visual functions by deprivation and
A comparison of strategies used by blind and sighted people. Journal of rearrangement. Psychologische Forschung 31:338-48. [DI]
Vistial Impairment and Blindness 81:100-105. [DJB] Herskovits, A. (1986) Language and spatial (.ignition: An interdisciplinary
Engelson, S. P. & McDermott, D. V. (1992) Passive robot map building with study of the prepositions in English. Cambridge University Press
exploration scripts. YALEUIDCSITR-898, Department of Computer (Cambridge). [arBL]
Science, Yale University. [MJT] Hinton, G. E., McClelland, J. L. & Rumelhart, D. E. (1986) Distributed
Farah, M. J. (1991) Patterns of co-occurrence among the associative agnosias. representations. In: Parallel distributed processing: Explorations in the
Cognitive Neuropsychology 8:1-19. [MCC] microstructure of cognition, vol. 1, ed. D. E. Rumelhart, J. L.
Farah, M. J., Hammond, K., Levine. D. & Calvanio, R. (1988) Visual and McClelland & the PDP Research Group. Bradford Books/MIT
spatial mental imagery: Dissociable systems of representation. Cognitive Press. [PBH]
Psychology 20:439-62. [aBL, BB] Hintzman, D. L., O'Dell, C. S. & Arndt, D. R. (1981) Orientation in
Farah, M. J. & McClelland, J. L. (1991) A computational model of semantic cognitive maps. Cognitive Psychology 13:149-206. [DJB]
memory impairment. Journal of Experimental Psychology: General Hirtle, S. & Jonides, J. (1985) Evidence of hierarchies in cognitive maps.
120:339-57. [MCC] Memory and Cognition 13:208-17. [aBL]
Feldman, J. (1991) Perceptual simplicity and modes of structural generation. Hochberg, J. & Brooks, V. (1962) Pictorial recognition as an unlearned ability:
Proceedings of the Thirteenth Annual Conference of the Cognitive Science A study of one child's performance. American Journal of Psychology
Society, Chicago, IL, August. [JF] 75:624-28. [aBL]
(1992) Constructing perceptual categories. Proceedings of the I.E.E.E. Hochberg, J. & Gellman, L. (1975) The effect of landmark features on mental
Conference on Computer Vision and Pattern Recognition, Champaign, IL, rotation times. Memory and Cognition 5:23-26. [aBL]
June. [JF] HoBman, D. & Richards, W. (1984) Parts of recognition. Cognition 18:65-
Fillmore, C. (1975) Santa Cruz lectures on deixis, 1971. Indiana University 96. [aBL]
Linguistics Club. [arBL] Hooper, K. (1978) Perceptual aspects of architecture. In: Handbook of
Fodor, J. A. (1980) Reply to Putnam. In: Language and learning: The debate perception, vol. 10, ed. E. C. Carterette & M. P. Friedman. Academic
between Jean Piaget and Noam Chomsky, ed. M. Piatelli-Palmerini. Press. [aBL]
Harvard University Press. [rBL] Huttenlocher, J. & Strauss, S. (1968) Comprehension and a statement's
(1983) Modularity of mind. MIT Press. [aBL] relation to the situation it describes. Journal of Verbal Learning and
Fortes, M. (1940) Children's drawings among the Tallensi. Africa 13:293- Verbal Behavior 7:300-304. [aBL]
95. [JBD] Ingle, D. (1967) Two visual mechanisms underlying the behavior offish.
(1981) Tallensi children's drawings. In: Universals of human thought, ed. Psychologische Forschung 31:44-51. [DI]
B. Lloyd & J. Gay. Cambridge University Press (Cambridge). [JBD) Ingle, D., Schneider, G., Trevarthen, C. & Held, R. (1967) Locating and
Foster, D. H. & Ward, P. A. (1991) Asymmetries in oriented-line detection identifying: Two modes of visual processing: A symposium. Psychologische
indicate two orthogonal filters in early vision. Proceedings of the Royal Forschung 31:42-43. [aBL, DI]
Society of London B 243:75-81. [JMW] Jackendoff, R. (1973) The base rules for prepositional phrases. In: A festschrift
Franklin, N. & Tversky, B. (1990) Searching imagined environments. Journal for Morris Halle, ed. S. Anderson & P. Kiparsky. Holt, Rinehart &
of Experimental Psychology: Ceneral 119:63-76. [DJB] Winston. [aBL]
Fussell, S. R. & Krauss, R. M. (1989) The effects of intended audience on (1976) Toward an explanatory semantic representation. Linguistic Inquiry
message production and comprehension: Reference in a common ground 7:89-150. [aBL]
framework. Journal of Experimental Social Psychology 25:203-19. [aBL] (1983) Semantics and cognition. MIT Press. [arBL]
Gallistel, C. R. (1990) The organization of learning. MIT Press. [aBL, DJB] (1987a) On beyond zebra: The relation of linguistic and visual information.
Cibson, E. (1969) Principles of perceptual learning and development. Cognition 26:89-114. [aBL, SDM]
Appleton-Century-Crofts. [aBL] (1987b) Consciousness and the computational mind. MIT Press. [arBL]
Gibson, J. J. (1966) The senses considered as perceptual systems. Houghton (1990) Semantic structures. MIT Press. [arBL]
Mifllin. [SDM] (1992) Languages of the mind. Bradford Books/MIT Press. [rBL]
(1979) The ecological approach to visual perception. Houghton Jackendoff, R. & Landau, B. (1991) Spatial language and spatial cognition. In:
Mifflin. [JW] Bridges between psychology and linguistics: A Swarthmore Festschrift for
Clenberg, A. M., Meyer, M. & Lindem, K. (1987) Mental models contribute lila Gleitman, ed. D. J. Napoli & J. Kegl. Erlbaum. [aBL]
to foregrounding during text comprehension. Journal of Memory and Jay, M. F. & Sparks, D. L. (1987) Sensorimotor integration in the primate
Language 26:69-83. [DJB] superior colliculus, II: Coordinates of auditory signals. Journal of
Coodale, M. A. & Milner, A. D. (1992) Separate visual pathways for Neuwphysiology 57:35-55. [MJT]
perception and action. Trends in Neurosdence 15(l):20-25. [DI] Jepson, A. & Richards, W. (1992) What makes a good feature? In: Spatial
Grabowecky, M. & Khurana, B. (1990) Features were meant to be integrated. vision in humans and robots, ed. L. Harris & M. Jenkins. Cambridge
Investigative Ophthalmology and Vision Science 31:105. [JMW] University Press (Cambridge). [JF]

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 263

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 References/Landau & Jackendoff: Spatial language

Johansson, G. (1973) Visual perception of biological motion and a model for its (in press) Relativity in spatial conception and description. In: Rethinking
analysis. Perceptual Pschycophysics 14:201-11. [DI] linguistic relativity, ed. J. J. Gumperz & S. C. Levinson. Cambridge
Johnson-Laird, P. N. (1983) Mental models: Towards a cognitive science of University Press (Cambridge). [BT]
language, inference, and consciousness. Harvard University Leyton, M. (1989) Inferring causal history from shape. Cognitive Science
Press. [DJB] 13:357-89. [aBL]
Johnston, J. (1984) Acquisition of locative meanings: Behind and in front of. (1992) Symmetry, causality, mind. MIT Press. [aBL, JF]
Journal of Child Language 11:407-22. [aBL] Livingstone, M. & Hubel, D. (1989) Segregation of form, color, movement,
Jolicoeur, P. (1985) The time to name disoriented natural objects. Memory 6- and depth: Anatomy, physiology, and perception. Science 240:740-
Cognition 13:289-303. [MJT] 49. [aBL]
Jones, S., Smith, L. & Landau, B. (1991) Object properties and knowledge in Logan, G. D. (1991) Linguistic and conceptual control of visual spatial
early lexical learning. Child Development 62:499-516. [aBL] attention. Paper presented at the 32nd annual meeting of the
Keil, F. (1989) Concepts, kinds, and conceptual development. MIT Psychonomic Society, San Francisco, CA. [DJB]
Press. [aBL] Lowe, D. (1985) Perceptual organization and visual recognition.
Kendall, D. (1989) A survey of the statistical theory of shape. Statistical Kluwer. [aBL]
Science 4:87-120. [JF] Lynch, K. (1960). The image of the city. MIT Press. [aBL]
Koenderink, J. J. (1987) An internal representation for solid shape based on Marr, D. (1982) Vision. Freeman. [arBL, MCC, JBD, JF, SDM, JW]
the topological properties of the apparent contour. In: Image Marr, D. & Nishihara, H. (1978) Representation and recognition of the spatial
understanding 1985-86, ed. W. Richards & S. Ullman. Ablex. [MJT] organization of three-dimensional shapes. Proceedings of the Royal
Kosslyn, S. M. (1990) Components of high-level vision: A cognitive Society of London B 200:269-94. [aBL, MJT]
neuroscience analysis and accounts of neurological syndromes. Cognition Marr, D. & Vaina, L. (1982) Representation and recognition of the movements
34:203-77. [arBL] of shapes. Proceedings of the Royal Society of London 214:501-
Kosslyn, S. M., Chabris, C. F., Marsolek, C. J. & Koenig, O. (1992) 24. [aBL]
Categorical versus coordinate spatial representations: Computational Marsolek, C. J. (1992) Abstract-visual-form representations in the left cerebral
analysis and computer simulations. Journal of Experimental Psychology: hemisphere. (Submitted.) [HDB]
Human Perception and Performance 18:562-77. [HDB] Marsolek, C. J., Kosslyn, S. M. & Squire, L. R. (1992) Form-specific visual
Kosslyn, S. M. & Koenig, O. (1992) Wet mind. The Free Press. [HDB] priming in the right cerebral hemisphere. Journal of Experimental
Kosslyn, S. M., Koenig, O., Barrett, A., Cave, C. B., Tang, J. & Cabrieli, Psychology: Learning, Memory, and Cognition 18:492-508. [HDB]
J. D. E. (1989) Evidence for two types of spatial representations: Miller, G. A. (1956) The magical number seven, plus or minus two: Some
Hemispheric specialization for categorical and coordinate relations. limits on our capacity for processing information. Psychological Review
Journal of Experimental Psychology: Human Perception and Performance 63:81-97 [MCC]
15:723-35. [HDB, MCC] Miller, G. & Johnson-Laird, P. (1976) Language and Perception. Harvard
Kriegman, D. J. & Ponce, J. (1990) Computing exact aspect graphs of curved University Press. [aBL]
objects: Solids of revolution. International Journal of Computer Vision Milner, A. D. & Goodale, M. A. (1992) Visual pathways to perception and
5(21):119-35. [MJT] action. In: The visually responsive neuron, ed. T. P. Hicks, S.
Kuipers, B. (1978) Modeling spatial knowledge. Cognitive Science 2:129- Molotchinkoff&T. Ono. Elsevier. [DI]
53. [aBL] Moore, G. T. (1976) Theory and research on the development of
Lakoff, G. (1987) Women, fire, and dangerous things. University of Chicago environmental knowing. In: Environmental knowing, ed. G. T. Moore &
Press. [aBL] R. G. Golledge. Dowden, Hutchinson & Ross. [SDM]
Lakoff, G. & Turner, M. (1980) Metaphors we live by. University of Chicago Moraglia, G. (1989) Display organization and the detection of horizontal line
Press. [aBL] segments. Perception 6 Psychophysics 45:265-72. [JMW]
Landau, B. (1986) Early map use as an unlearned ability. Cognition 22:201- Morrow, D. G., Greenspan, S. L. & Bower, G. H. (1987) Accessibility and
23. (aBL] situation models in narrative comprehension. Journal of Memory and
(1991) Spatial knowledge of objects in the young blind child. Cognition Language 26:165-87. [ DJ B]
38:145-78. [aBL] Murphy, G. & Medin, D. (1985) The role of theories in conceptual coherence.
(1993) Object, word and name in first language learning. In: Special issue Psychological Review 92:289-316. [aBL]
on the lexicon, ed. L. R. Gleitman & B. Landau. Lingua (in Neisser, U. (1989) Direct perception and recognition as distinct perceptual
press). [aBL] systems. Paper presented to the Cognitive Science Society, Ann Arbor,
Landau, B. & Gleitman, L. R. (1985) Language and experience: Evidence MI, August. [SDM]
from the blind child. Harvard University Press. [aBL] (1992) Distinct systems for "where" and "what": Reconciling the ecological
Landau, B., Jones, S. & Smith, L. (1992) Syntactic context and object and representational views of perception. Paper presented to the
properties in early lexical learning. Journal of Memory and Language (in American Psychological Society, San Diego, CA, June. [SDM]
press). [aBL] Newport, E. (1988) Constraints on language learning and their role in
Landau, B., Smith, L. & Jones, S. (1988) The importance of shape in early language acquisition: Studies of the acquisition of American Sign
lexical learning. Cognitive Development 3:299-321. [aBL] Language. Language Sciences 10(l):147-61. [aBL]
Landau, B., Spelke, E. & Gleitman, H. (1984) Spatial knowledge in a young Oakhill, J. V. & Johnson-Laird, P. N. (1984) Representation of spatial
blind child. Cognition 16:225-60. [aBL] descriptions in working memory. Current Psychological Research b
Landau, B. & Stecker, D. (1990) Objects and places: Syntactic and geometric Reviews 3:52-62. [DJB]
representations in early lexical learning. Cognitive Development 5:287- O'Keefe, J. & Nadel, L. (1978) the hippocampus as a cognitive map. Oxford
312. [arBL, ADF) University Press (Oxford). [aBL]
Landau, B., Stecker, D. & Lederer, A. (forthcoming) Asymmetry in children's Olson, D. & Bialystok, E. (1983) Spatial cognition: The structure and
and adults' judgements of near. [aBL] development of the mental representation of spatial relations.
Lederman, S. & Klatzky, R. (1987) Hand movements: A window into haptic Erlbaum. [aBL, DRO]
object recognition. Cognitive Psychology 19:342-68. [aBL] Oppenheimer, F. (1934) Optische Versuche fiber Ruhe und Bewegung.
Levelt, W. (1983) Some perceived limitations on talking about space. Psychologische Forschung 20:1-46. [aBL]
Unpublished manuscript, Max Planck Institute for Psycholinguistics, Pager, H. (1972) Ndedema. International Book Services. [JBD]
Nijmegen. [aBL] Paillard, J. (1987) Cognitive versus sensorimotor encoding of spatial
(1984) Some perceptual limitations on talking about space. In: Limits in information. In: Cognitive processes and sfiatial orientation in animal and
perception, ed. A. J. van Doom, W. A. van de Grind & J. J. Koenderink. man, ed. P. Ellen & C. Thinus-Blanc. Martinus Nijhoff. [BB]
Utrecht: Coronet Books. [DJB] ed. (1991) Brain and space. Oxford University Press. [MJT]
Levine, D., Warach, J. & Farah, M. (1985) Two visual systems in mental Parker, D. M. & Deregowski, J. B. (1990) Perception and artistic style. North
imagery: Dissociation of "what' and "where" in imagery disorders due to Holland. [JBD]
bilateral posterior cerebral lesions. Neurology 35:1010-18. [aBL] Parsons, L. H. (1987) Imagined spatial transformations of one's hands and feet.
Levinson, S. C. (1991) Relativity in spatial conception and description. Cognitive Psychology 19(2):178-241. [aBL]
Working paper no. 1, Cognitive Anthropology Research Group, Max Pelisson, D., Prablane, C , Goodale, M. & Jeannerod, M. (1986) Visual
Planck Institute for Psycholinguistics, Nijmegen. [DIS] control of reaching movements without vision of the limb. II. Evidence
(1992) Vision, shape, and linguistic description: Tzeltal body-part for fast nonconscious process correcting the trajectory of the hand to the
terminology and object description. Paper delivered to the Workshop on final position of a double-step stimulus. Experimental Brain Research
Spatial Description in Mayan Languages. Nijmegen. [rBL] 62:303-11. [BB]

264 BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 References/Landau & Jackendoff: Spatial language

Pinker, S. (1989) Leamability and cognition: The acquisition of argument Tarr, M. J. fit Black, M. J. (1991) A computational and evolutionary
structure. MIT Press. [aBL] perspective on the role of representation in vision. YALEUIDCSIRR-899,
Pinker, S. & Bloom, P. (1990) Natural language and natural selection. Department of Computer Science, Yale University. [MJT]
Behavioral and Brain Sciences 13('l):707-84. [aBL] Tarr, M. J. fit Pinker, S. (1989) Mental rotation and orientation-dependence in
Pinxten, R., van Dooren, I. fit Harvey, K. (1983) The anthropology of space: shape recognition. Cognitive Psychology 21(28):233-82. [aBL, MJT]
Explorations into natural philosophy and semantics of the Navajo. (1990) When does human object recognition use a viewer-centered
University of Pennsylvania Press. [aBL] reference frame? Psychological Science l(42):253-56. [MJT]
Retz-Sehmidt, G. (1988) Various views on spatial prepositions. AI Magazine Taylor, H. A. fit Tversky, B. (1992) Spatial mental models derived from survey
9:95-105. [DJB] and route descriptions. Journal of Memory and Language 31:261-
Rieser, J. J. fit Heiman, M. (1982) Spatial self-reference systems and shortest 92. [DIB]
route behavior in toddlers. Child Development 53(2):524-33. [aBL] Thompson, D. (1961) On growth and form. Cambridge University Press
Rosch, E. (1975) Cognitive reference points. Cognitive Psychology 7(4):532- (Cambridge). [aBL]
47. [aBL] Treisman, A. (1986) Properties, parts, and objects. In: Handbook of human
Rosch, E., Mei-vis, C. B., Gray, W. D., Johnson, D. M. & Boyes-Braem, P. perception and performance, ed. K. R. Boff, L. Kaufmann & J. P.
(1976) Basic objects in natural categories. Cognitive Psychology 8:382- Thomas. Wiley. [JMW]
439. [aBL] (1988) Features and objects: The 14th Bartlett memorial lecture. Quarterly
Rueckl, J., Cave, K. & Kosslyn, S. (1988) Why are "what" and "where" Journal of Experimental Psychology 40A:201-37. [JMW]
processed by separate cortical visual systems? A computational Trevarthen, C. B. (1968) Two mechanisms of vision in primates.
investigation. Journal of Cognitive Neuroscience l(2):171-86. [aBL, JF] Psychologische Forschung 31:299-337. [Dl]
Sadalla, E., Burroughs, W. J. & Staplin, L. J. (1980) Reference points in Tversky, B. fit Hemenway, K. (1984). Objects, parts, and categories. Journal of
spatial cognition. Journal of Experimental Psychology: Human Learning Experimental Psychology: General 113:169-93. [BT]
and Memory 6(5):516-28. [aBL] Ullman, S. fir Basri, R. (1991) Recognition by linear combinations of models.
Schneider, C. E. (1967) Contrasting visual functions of tectum and cortex in IEEE Transactions on Pattern Analysis and Machine Intelligence
the golden hamster. Psychologische Forschung 31:52-62. [DI] 13(10):992-1006. [MJT]
(1969) Two visual systems. Science 163:895-902. [aBL, BB] Ungerleider, L. G. fie Mishkin, M. (1982) Two cortical visual systems. In:
Seibert, M. fit Waxman, A. M. (1992) Adaptive 3-D object recognition from Analysis of visual behavior, ed. D. J. Ingle, M. A. Goodale fit R. J. W.
multiple views. IEEE Transactions on Pattern Analysis and Machine Mansfield. MIT Press. [aBL, BB]
Intelligence 14(2):107-24. [MJT] Vandeloise, C. (1986) L'espace en francais. Editions du Seuil. (English
Shepard, R. N. (1984) Ecological constraints on internal representation: translation: Spatial prepositions. University of Chicago Press,
Resonant kinematics of perceiving, imagining, thinking, and dreaming. 1991.) [arBL]
Psychological Revicu) 91:417-47. [SDM] Van Essen, D., Anderson, C. fit Felleman, D. (1992) Information processing
Shepard, R. & Cooper, L. (1982) Mental images and their transformations. in the primate visual system: An integrated systems perspective. Science
MIT Press. [aBL] 255:419-23. [aBL]
Shepard, R. N. & Hurwitz, S. (1984) Upward direction, mental rotation, and von Hofsten, C. (1980) Predictive reaching for moving objects by human
discrimination of left and right turns in maps. Cognition 18:161- infants. Journal of Educational and Child Psychology 30:369-
93. [DJB] 82. [aBL]
Smith, L., Jones, S. fie Landau, B. (1992) Count nouns, adjectives, and Warrington, E. K. fie McCarthy, R. (1987) Categories of knowledge: Further
perceptual properties in novel word interpretations. Developmental fractionation and an attempted integration. Brain 110:1273-
Psychology 28(2):273-86. [aBL] 96. [MCC]
Stephens, A. fit Coup, P. (1978) Distortions in judged spatial relations. Warrington, E. K. fit Shallice, T. (1984) Category-specific semantic
Cognitive Psychology 10:422-37. [aBL] impairments. Brain 107:829-54. [MCC]
Stiles-Davis, J., Kritchevsky, M. & Bellugi, U. eds. (1988) Spatial cognition: White, R. (1989) Visual thinking in the ice age. Scientific American 261(1):92-
Brain bases and development. Erlbaum. [aBL] 99. [MJT]
Supalla, T. (1990) Structure and acquisition of verbs in motion and location in Williams, P. (1991) Children's and adults' understanding of across.
A.S.L. Bradford Books/MIT Press. (aBL] Undergraduate honors thesis, Department of Psychology, Columbia
Talmy, L. (1975) Semantics and syntax of motion. In: Syntax and semantics, University. [aBL]
vol. 4, ed. J. Kimball. Academic Press. [DIS] Wolfe, J. M. (1992) The parallel guidance of visual attention. Current
(1978) The relation of grammar to cognition. In: Proceedings ofTlNLAP-2: Directions in Psychological Science 1(4): 125-28. [JMW]
Theoretical issues in natural language processing, ed. D. Waltz. Wolfe, J. M. fit Friedman-Hill, S. R. (1992) Part-whole relationships in visual
University of Illinois Press. [aBL] search. Investigative Ophthalmology and Visual Science 33:1356.
(1980) Lexicalization patterns: Semantic structure in lexical forms. In: [JMW]
Language typology and syntactic description, vol. 3, ed. T. Shopen. Wolfe, J. M., Friedman-Hill, S. R., Stewart, M. I. & O'Connell, K. M. (1992)
Cambridge University Press. [rBL] The role of categorization in visual search for orientation. Journal of
(1983) How language structures space. In: Spatial orientation: Theory, Experimental Psychology: Human Perception and Performance 18:34-
research, and application, ed. H. Pick & L. Acredolo. Plenum 49. [JMW]
Press. [aBL] Wolfe, J. M., Yu, K. P., Stewart, M. I., Shorter, A. D., Friedman-Hill, S. R.
(1985) Lexicalization patterns: Semantic structure in lexical forms. In: it Cave, K. R. (1990) Limitations on the parallel guidance of visual
Language typology and syntactic description. Vol. 3: Grammatical search: Color X color and orientation X orientation conjunctions. Journal
categories and the lexicon, ed. T. Shopen. Cambridge University Press of Experimental Psychology: Human Perception and Performance 16:879-
(Cambridge) [aBL] 92. [JMW]
Tarr, M. J. (1989) Orientation dependence in three-dimensional object Wong, E. fit Mack, A. (1981) Saccadic programming and perceived location.
recognition. Ph.D. dissertation, Department of Brain and Cognitive Ada Psychologica 48:123-31. [BB]
Sciences, MIT. [MJT]

BEHAVIORAL AND BRAIN SCIENCES (1993) 16:2 265

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733
 VISUAL
NEUROSCIENCE
An International Journal
for Experimental and
Theoretical Research
EDITOR:
James T. Mcllwain, Brown University

Visual Neuroscience is devoted to the

publication of research and theoretical
articles in basic visual neuroscience, with
primary emphasis on retinal and brain
mechanisms that underlie visually-guided
behaviors and visual perception. The
major goal of Visual Neuroscience is to
bring together in one journal a broad
range of studies which reflect the exciting
diversity and originality of contemporary
research in this field. Methodologies are
drawn from neuroanatomy,
neurophysiology, neurochemistry,
neuroimmunology, and behavioral science,
as well as computational models and
computer-assisted formulations. Molecular,
cellular, local-circuit, and systems-level
analyses in both vertebrate and invertebrate
species are presented.

Visual Neuroscience (ISSN 0952-5238) is published

bi-monthly. Subscription to Volume 10, 1993:
$320.00 for institutions, $125.00 for individuals,
single parts $55.00.

Send orders to:

CAMBRIDGE UNIVERSITY PRESS

40 West 20th Street
New York, NY 1001M211

Outside the USA, Canada, and Mexico:

CAMBRIDGE UNIVERSITY PRESS
The Edinburgh Building
Shaftesbury Road
Cambridge CB2 2RU, England

Downloaded from https://www.cambridge.org/core. INSEAD, on 30 Jul 2018 at 05:31:26, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1017/S0140525X00029733