Computers in translation

Researchers have been attempting to develop systems to emulate the

human translation process for some forty years. What is it about natural
language that makes this such a daunting challenge? While other software
applications have achieved rapid and lasting success, machine translation
has failed to penetrate the worldwide market to any appreciable extent.
Does this merely reflect a reluctance to adopt it, or does it signal a more
fundamental and intractable problem?

Computers in translation is a comprehensive guide to the practical issues

surrounding machine translation and computer-based translation tools.
Translators, system designers, system operators and researchers present
the facts about machine translation: its history, its successes, its limitations
and its potential. Three chapters deal with actual machine translation
applications, discussing installations including the METEO system, used
in Canada to translate weather bulletins and weather reports, and the
system used in the Foreign Technology Department of the US Air Force.
As a guide for non-specialists seeking to broaden their knowledge of the
practicalities of machine translation, it will be of particular interest to
translators, technical authors, technical publications managers, students
and lecturers in languages and linguistics.

John Newton, a Senior Lecturer in the Department of Languages at

Manchester Polytechnic, also runs his own consultancy business,
providing translation services and distributing linguistic software. He
has undertaken translation-related international consultancy assignments
and has given talks and presentations on machine translation throughout
Europe. He has also supervised a number of machine translation
installations, and is the author of several articles on machine translation
and related topics.
Computers in translation

A practical appraisal

Edited by
John Newton

London and New York

First published 1992
by Routledge
11 New Fetter Lane, London EC4P 4EE

Simultaneously published in the USA and Canada

by Routledge
a division of Routledge, Chapman and Hall, Inc.
29 West 35th Street, New York, NY 10001

Routledge is an imprint of the Taylor & Francis Group

This edition published in the Taylor & Francis e-Library, 2003.

© 1992 John Newton

All rights reserved. No part of this book may be reprinted or

reproduced or utilized in any form or by any electronic,
mechanical, or other means, now known or hereafter invented,
including photocopying and recording, or in any information
storage or retrieval system, without permission in writing from
the publishers.

A catalogue record for this book is available from the British Library.

ISBN 0-203-12897-4 Master e-book ISBN

ISBN 0-203-20273-2 (Adobe eReader Format)

ISBN 0-415-05432-X (Print Edition)

Library of Congress Cataloging-in-Publication Data

Computers in translation: a practical appraisal/edited by John

Newton.
Includes bibliographical references and index.
ISBN 0-415-05432-X (Print Edition)
1. Machine translating. I. Newton, John, 1944–
P308.C67 1992
418'.02'0285–dc20 91–41614
CIP
This book is dedicated to the expertise and
wisdom of its contributors and to
the marvel of natural language.
Contents

Figures ix
Contributors x
Preface xvi
Abbreviations and acronyms xviii
1 Introduction and overview
John Newton 1
2 The story so far: an evaluation of machine
translation in the world today
Jeanette Pugh 14
3 Made to measure solutions
Annette Grimaila in collaboration with John Chandioux 33
4 The Perkins experience
John Newton 46
5 Machine translation in a high-volume translation
environment
Muriel Vasconcellos and Dale A.Bostad 58
6 Esperanto as an intermediate language for
machine translation
Klaus Schubert 78
7 Limitations of computers as translation tools
Alex Gross 96
8 Computerized term banks and translation
Patricia Thomas 131
9 The translator workstation
Alan Melby 147

vii
viii Contents

10 SYSTRAN: it obviously works but how much

can it be improved?
Yorick Wilks 166
11 Current research in machine translation
Harold L.Somers 189

Bibliography 208
Glossary of terms 222
Index 228
Figures

5.1 Spanish-English unedited machine translation at PAHO 60

5.2 English-Spanish unedited machine translation at PAHO 62
7.1 The effectiveness of computer translation 103
8.1 Information provided for each term in EURODICAUTOM 135
8.2 Example of a completed record for the entry term
‘measles virus’ in the Surrey term bank with British
English as the source language (SL). Not all record
fields are necessarily completed. (The record format
is under revision.) 136

ix
Contributors

Dale A.Bostad is a technical adviser in the Directorate of

Translations at the Foreign Technology Division, US Air Force,
Dayton, Ohio. Originally a traditional translator from Russian into
English, he became involved in machine translation in 1976 and has
found the latter career considerably more challenging. Currently he
is studying Japanese, Arabic and Chinese.

John Chandioux is the founder and President of John Chandioux

Consultants Inc., which specializes in research and development in
the field of computational linguistics. He is the owner and designer
of the famed machine-translation system M ETEO and has
developed his own proprietary programming language known as
GramR. He is also the author of a number of spelling checkers
incorporated into French versions of word-processing programs,
including the initial dictionary sold with WordPerfect 5.0. His most
recent activities include the launching of a family of French-language
spelling and grammar checkers in Europe and North America under
the generic name GramR.

Annette Grimaila, originally a programmer-analyst in the finance

industry, has been associated with John Chandioux Consultants Inc.
for some four years. She was project leader and principal
programmer of the General TAO system.

Alex Gross served as a literary adviser to the Royal Shakespeare

Company during the 1960s, and his translations of Dürrenmatt and
Peter Weiss have been produced in London and elsewhere. He was
awarded a two-year fellowship as writer-in-residence by the Berliner
Künstler-Programm, and one of his plays has been produced in

x
 Contributors xi

several German cities. Having spent twelve years in Europe, he is

fluent in French, German, Italian and Spanish. He has published
works related to the translation of traditional Chinese medicine and
is planning further work in this field. Two more recent play
translations were commissioned and produced by UBU Repertory
Company in New York, one of them as part of the official American
celebration of the French Revolutionary Bicentennial in 1989.
Published play translations are The Investigation (Peter Weiss, 1986,
London: Calder & Boyars) and Enough is Enough (Protais Asseng,
1985, NYC: UBU Repertory Co.). His translation experience has
also encompassed journalistic, diplomatic and commerical texts, and
he has taught translation as part of NYU’s translation certificate
programme. Over the last few years a number of his articles on
computers, translation and linguistics have appeared in the United
Kingdom, the Netherlands and the United States. He is chairperson
of the Machine Translation Committee of the New York Circle of
Translators, is also an active member of the American Translators
Association and has been involved in the presentations and
publications of both groups.

Alan Melby is Associate Professor of Linguistics at Brigham Young

University (Provo, Utah, USA), where he also chairs the Translation
Research Group (BYU-TRG). He is also Director of Research and
Development at LinguaTech International (Orem, Utah), where he
designed MTX(tm), one of the first microcomputer-based
terminology management systems. He has done research in a
number of areas of natural language processing, including machine
translation, speech processing, translator tools and full-text electronic
dictionaries. During February 1990, he was Visiting Professor at the
Collège de France, one of the most prestigious academic institutions
in France, where he delivered a series of lectures on linguistic theory.
He believes that the difference between lexical and terminological
units is more fundamental than was previously suspected. His
favourite living philosophers are J.-M.Zemb and E.Levinas.

John Newton is a Senior Lecturer in the Department of Languages at

Manchester Polytechnic. His experience of the practical aspects of
machine translation and translation tools is long and varied. He was
the founding managing director of Weidner Translation (Europe)
Limited, and also served as operations director of another major
international translation company. More recently, he has concentrated
xii Contributors

on importing and distributing linguistic software via his own

consultancy company, Mentor Documentation Services Limited. He
has undertaken a number of international translation-related
consultancy assignments and has given talks and presentations on
machine translation throughout Europe. He has also been involved in
operational research for translation tools developers and has made
many presentations of such products to commercial, industrial,
institutional and academic audiences. At WTE, he played a leading
role in devising training programmes and in introducing the MicroCat
system to the UK higher education sector. He also supervised a
number of industrial machine translation installations. Mr Newton,
who has had a number of articles published on machine translation
and related topics, is a former member of the Institute of Linguists’
Council and was Chairman of the Institute’s Central Southern
Regional Society from 1987–91.

Jeanette Pugh worked at UMIST in the Department of Language

and Linguistics from 1985–91. Throughout this period, she was
closely involved with the E U ROTRA machine translation
programme, and played a major role in the planning and
coordination of the work of the British EUROTRA team. She was a
member of EUROTRA’s central management committee (liaison
group) and regularly attended its meetings at the European
Commission in Luxembourg. She was also appointed to
EUROTRA’s central planning committee which worked on the
formulation of proposals for the content of the 1991–2 EUROTRA
transition programme. Dr Pugh was a co-author of Multilingual
Aspects of Information Technology (1986, Gower Press), and recently
coauthored an article on machine translation in Europe which is to
appear in The Encyclopaedia of Language and Linguistics published by
Pergamon Press.

Klaus Schubert studied general linguistics, Nordic philology and

Slavic philology at the universities of Kiel (Germany) and Uppsala
(Sweden). He obtained a doctor’s degree from the University of Kiel
in 1982 where he also carried out sociolinguistic research for several
years. Since 1985, he has been with the software company Buro voor
Systeemontwikkeling (B SO) in Utrecht, the Netherlands. He
participated in the DLT machine translation project at B SO/
Research as a computational linguist and later became a linguistic
system designer and project manager. At present, he is attached to
 Contributors xiii

BSO/Language Technology in Baarn, the Netherlands, a department

which provides software services in the fields of natural language
processing, machine-aided translation and linguistic knowledge
engineering. Klaus Schubert’s major publications include
Anredeforschung (Braun, Kohz, Schubert, 1986, Tübingen: Narr),
Metataxis. Contrastive dependency syntax for machine translation (Schubert,
1987, Dordrecht/Providence: Foris), Interlinguistics. Aspects of the science
of planned languages (Schubert (ed.), 1989, Berlin/New York: Mouton
de Gruyter). He is also European editor of the scholarly journal
Language Problems and Language Planning (Amsterdam/Philadelphia:
Benjamins), and, in August 1992, will be moving to Flensburg,
Germany, to take up a position as a professor of computational
linguistics and technical translation at the Institut für Übersetzen of
the Fachhochschule.

Harold L.Somers is a Senior Lecturer in the Department of

Language and Linguistics and in the Centre for Computational
Linguistics at the University of Manchester Institute of Science and
Technology (UMIST). He is involved in teaching and research in
the field of machine translation (MT). Teaching activities include
lecturing to undergraduate students on the department’s BSc course
in computational linguistics, and to postgraduates on the MSc course
in MT, both courses unique and innovatory when they began in
1978 and 1987 respectively. In addition, he is frequently invited to
give guest lectures on MT and natural language processing (NLP) to
linguistics and computer science students elsewhere in the UK and
in Europe. His research interests are all aspects of MT, and more
narrowly syntax and semantics in NLP. He is the author of Case and
Valency in Computational Lingustics (1987, Edinburgh University Press),
and is co-author, with John Hutchins, of a textbook on machine
translation published by Academic Press. A Prolog and N LP
textbook aimed at linguistics students is also in preparation. He has
worked on several MT projects, including the EC’s EUROTRA
system, and various other MT projects at UMIST. In 1988–9 he
was a visiting research fellow at Toshiba Corporation’s R&D Centre
in Japan. Since his return to the UK, his main research interests have
been dialogue-based MT and example-based MT, in both cases
involving Japanese as well as European languages.

Patricia Thomas has made a career using languages, particularly

French, as translator and teacher, in addition to being involved in the
xiv Contributors

editing of learned scientific books and journals. After gaining an

MPhil at the University of Surrey in 1983, she followed terminology
training at Infoterm, the international terminology organization in
Vienna. She has been instrumental in creating a multilingual
terminology data bank in specialized subject domains at the
University of Surrey, for which some of the input resulted from her
supervision of MA students in translation studies who were
interested in pursuing computational methods of terminology for
their final dissertations. This led naturally to her becoming involved,
while at Surrey, in producing terminology for a project sponsored by
the European Commission’s ESPRIT II programme to develop a
multilingual translator’s workbench (project no. 2315). A number of
invited papers have been given at international meetings and she has
co-authored publications on terminology and term banks. She has
also participated in the work of international nomenclature and
taxonomy organizations in the biological sciences. She is a member
of ISO TC 37 Working Group (vocabulary of terminology), editor
for the International Institute for Terminology Research (IITR),
Vienna, regional editor for Southern England for Infoterm, Vienna,
and is a Fellow of the Institute of Linguists. She is currently doing
research for her PhD.

Muriel Vasconcellos holds a BS degree in languages and a Master’s

and PhD in linguistics from Georgetown University. She has
devoted her entire career to translation, including twenty years
specifically to machine translation. Since 1977, she has been
responsible for the development and implementation of MT at the
Pan American Health Organization, Regional Office for the
Americas of the World Health Organization.

Yorick Wilks is Director of the Computing Research Laboratory at

New Mexico State University, a centre of excellence for research in
artificial intelligence and its applications. He received his doctorate
from Cambridge University in 1968 for work in computer programs
that understand written English in terms of a theory later called
‘preference semantics’: the claim that language is to be understood
by means of a search for semantic ‘gists’, combined with a coherence
function over such structures that minimizes effort in the analyser.
This has continued as the focus of his work, and has applications in
the areas of machine translation, the use of English as a ‘front end’
for users of databases and the computation of belief structures. He
 Contributors xv

has published five books and numerous articles in this area of

artificial intelligence and, before going to New Mexico, was
Professor of Computer Science and Linguistics at the University of
Essex (UK).
Preface

Machine translation (MT) clearly arouses considerable interest

beyond the ranks of those directly concerned with developing
systems or operating them. When I was involved in supplying and
installing translation software, I was frequently asked to recommend
a book with a practical focus that would serve as a general
introduction to the field. This proved to be difficult, as most existing
publications were aimed at a specialist readership and assumed a
prior knowledge of the subject’s technical aspects that the general
reader was unlikely to possess. The idea of presenting a collection of
practical accounts from researchers, system developers, system
operators and translators, together with an appraisal of MT’s
position in the world today and an outline of the main avenues of
current research, grew out of this experience.
As opinions voiced on MT vary from the wildly optimistic to the
fiercely negative, we have attemped to place the matter in perspective
by describing the conditions under which MT systems are likely to
produce optimum results and by making some general statements on
their current limitations.
Any assessment of the computer’s contribution in the sphere of
natural-language translation should recognize the importance of so-
called ‘lower level’ computerized translation tools. Unlike MT
systems, which can only be gainfully employed under certain
conditions, computerized tools offer tangible benefits in almost every
area of written translation. Products in this category have therefore
been given due consideration in this volume.
Our aim of providing a broad practical introduction to the field
clearly demanded contributions from authors representing as many
strands of opinion and as many allied disciplines as was practicable

xvi
 Preface xvii

in a work this size. I therefore invited a number of leading authorities

in Europe and North America each to provide a chapter dealing with
their own specialism and was fortunate enough to receive their
enthusiastic support.
It would, of course, be impossible to deal with our subject
exhaustively in a work of any size, let alone within the confines of
this volume. We believe, however, that our policy of selecting
practical topics and providing sufficient space for an in-depth
treatment of each has yielded a collection which goes some way
towards meeting the needs of readers seeking information in an
accessible and (dare I say it?!) ‘user-friendly’ form.
I extend my warmest thanks to all the contributors and to
Routledge for turning this idea into a reality.
John Newton
Abbreviations and acronyms

AFNOR Association française de normalisation (French

standards organization)
ACH Association for Computing in the Humanities
ACL Association for Computational Linguistics
AI Artificial Intelligence
ALLC Association for Literary and Linguistic Computing
ALPAC Automatic Language Processing Advisory
Committee
ASCII American Standard Code for Information
Interchange
Aslib Association for Information Management
(Association of Special Libraries and Information
Bureaux)
ATA American Translators Association
ATR Automated Telephony Research
BSO Buro voor Systeemontwikkeling
BT British Telecom
CAT Computer-Assisted (-Aided) Translation
CCL Centre for Computational Linguistics (at UMIST)
CCRIT Centre Canadien de Recherche en Informatisation
du Travail
CD-ROM Compact Disc Read Only Memory
CEC Commission of the European Communities
CFE Caterpillar Fundamental English
CICC Center for International Cultural Cooperation
(Japan)
CL Computational Linguistics
CLL Centre for Computational Linguistics

xviii
 Abbreviations and acronyms xix

CMU Carnegie Mellon University

COLING (International Conference on) Computational
Linguistics
CPU Central Processing Unit
DARPA Defense Advanced Research Projects Agency
DIN Deutsches Institut für Normung
DLT Distributed Language Translation
DOS Disk Operating System
DTP Desk-Top Publishing
EC European Community
ECHO European Commission Host Organisation
EDR Electronic Dictionary Research (Project)
EEC European Economic Community
ENGSPAN English to Spanish MT system (at PAHO)
EURATOM European Atomic Energy Authority
EUSIDIC European Association of Scientific Information
Dissemination Centres
FAHQT Fully Automated High Quality Translation
FTD Foreign Technology Division (US Air Force)
GETA Groupe d’études pour la traduction automatique
GUI Graphical User Interface
HPSG Head-Driven Phrase Structure Grammar
HAMT Human-Assisted (-Aided) Machine Translation
HT Human Translation
IBM International Business Machines
ICCL International Conference on Computational
Linguistics (COLING)
ILSAM International Language for Service and
Maintenance
IPSS International Packet Switch Stream
ISSCO Institut Dalle Molle pour les études sémantiques et
cognitives
JEIDA Japanese Electronic Industry Development
Association
JICST Japan’s Information Centre of Science and Technology
Kb Kilobyte
KBMT Knowledge-Based Machine Translation
LCD Liquid Crystal Display
LFG Lexical Functional Grammar
LIDIA Large Internationalisation of Documents by
Interacting with their Authors
xx Abbreviations and acronyms

LSP Language for Special Purposes

MAHT Machine-Assisted (-Aided) Human Translation
MATER Magnetic Tape Exchange Format (for
Terminological/Lexicographical Records)
MicroMATER Magnetic Tape Exchange Format (for
microcomputers)
MIT Massachusetts Institute of Technology
MITI (Japanese) Ministry of International Trade and
Industry
MAT Machine-Assisted (-Aided) Translation
MT Machine Translation
NLP Natural Language Processing
NMSU New Mexico State University
NSF (US) National Science Foundation
NYU New York University
OCR Optical Character Reader, Optical Character
Recognition
PACE Perkins Approved Clear English
PAHO Pan American Health Organization
PC Personal Computer
PECOF Post-Editor’s Correction Feedback
PSS Packet Switch Stream
RAM Random Access Memory
SGML Standard Generalized Markup Language
SL Source Language
SPANAM Spanish to American (English) MT system (at
PAHO)
SUSY Saarbrücker ÜbersetzungsSYstem
TAO Traduction assistée par ordinateur
TAUM Traduction Automatique de l’Université de
Montréal
TEAM Terminology Evaluation and Management
TEI Text Encoding Initiative
TL Target Language
UMIST University of Manchester Institute of Science and
Technology
WYSIWYG What You See Is What You Get
Chapter 1

Introduction and overview

John Newton

A BRIEF HISTORICAL PERSPECTIVE

Translation, a sine qua non for the comprehension of a text by a reader
unfamiliar with the language in which it was first written, has
doubtless been practised in one form or another for several
millennia; yet, until modern times, translators possessed little in the
way of tools, apart from manual writing implements, erasers and
hard—copy reference sources.
Present-day translators are major beneficiaries of the
technological revolution in office practice which was born with the
advent of the typewriter in the nineteenth century and which led to
today’s proliferation of personal computers (PCs) and their
associated peripherals. Most professional translators now regard
word processing as the normal method of creating a target text, while
those who prefer to use dictation machines invariably resort to it for
transcription.
Modern technology has transposed some of the translator’s
traditional tools to a new and more flexible medium: more and more
dictionaries (monolingual, bilingual, multilingual) are becoming
available in machine-readable form. These utilities are accessible from
a text file, via windows, and allow selected translation equivalents to
be ‘pasted’ directly into the text at the cursor position; some also
permit the user to make new entries and modify existing entries.
Facsimile machines and electronic mail systems have brought
similarly significant benefits, eliminating delays in receiving and
transmitting texts, and making suitably equipped translators globally
and instantly accessible, irrespective of their base location.
Although of great benefit to translators, the systems and devices
mentioned so far were designed to meet a broad range of general

1
2 John Newton

needs outside of translation. The chapters which follow discuss and

describe systems that were conceived specifically to facilitate, or even
perform, translation.
The products that we are concerned with come under two
headings: machine translation systems, and translation tools.
Systems which perform a syntactic analysis of a source text and
then generate a target language rendering thereof which seeks to
preserve and reconstitute its semantic and stylistic elements are
described as ‘machine translation’ (MT) systems, while those
designed to facilitate human translation through providing a
terminology management system, instant access to on—line
dictionaries, and other utilities are referred to as ‘translation tools’.
Despite this distinction, it is important to emphasize the essentially
auxiliary role of MT in most of its applications.
A typical translation tools package has two main constituents: a
text analyser, which furnishes input for various wordlists and
bilingual glossaries; and a terminology database, which provides a
flexible framework for storage and retrieval. These systems are used
to create bilingual ‘document’ dictionaries containing all the lexical
items—except articles, conjunctions and prepositions—found in a
specified source text, along with their translation equivalents. They
also perform morphological reductions to generate lists of base forms
(i.e. for English, all verbal, plural and possessive inflections are
removed) to simplify dictionary updating. In common with the off-
the-shelf dictionary products, most translation tools packages offer
automatic look-up and ‘paste-in’.
Translation tools and machine-readable dictionaries have to be
integrated into whatever combination of hardware and software
the translator happens to be using and this often results in an
assemblage of disparate elements that do not form a congruous
whole. Recognition of this problem led some researchers to adopt a
holistic approach in the 1980s. The translator workstation or
translator’s workbench (see Chapter 8: Patricia Thomas; and
Chapter 9: Alan Melby) combines any or all of the functions and
facilities mentioned above in an ergonomically designed unit aimed
at creating greater efficiency, greater productivity, and greater
translator satisfaction and well-being.
Of course, if texts are to be machine processed, they have to be
available in a compatible magnetic format, or they need to be
entered via an optical character reader (OCR). Extensive use of
word processing, together with the availability of media
 Introduction and overview 3

conversion and budget-priced OCR systems, has made the bulk

of today’s commercially translated text accessible to MT and
translation tools.
Compared with translation tools, MT has a relatively long and
chequered history (see Chapter 2: Jeanette Pugh). Despite its
acknowledged success in a number of well-publicized restricted
domain or restricted input environments, only a small proportion of
today’s worldwide translation workload is processed using MT, and
the commercially available systems’ ability to handle general (i.e.
unrestricted) text will need to improve significantly before the
installed base shows any substantial growth.
The difficulties inherent in tackling the kaleidoscopic and
cognitive aspects of natural language constitute a daunting challenge
for system developers which has led to many different and ingenious
approaches. In Chapter 11, Harold Somers provides a detailed
outline of the main strands of current MT research, including some
radical new concepts, and in Chapter 6 Klaus Schubert gives us an
in-depth account of his experience with the BSO DLT system and
explains the rationale behind the choice of Esperanto as an
intermediate language.
At first sight, it may seem surprising that MT projects should
have existed several decades before anyone tackled the less
ambitious task of developing translation tools and translator
workstations. However, this appears less remarkable if one
considers that such products only became practically feasible and
commercially viable when the price of P Cs had dropped
sufficiently to encourage their widespread use among translators
(see Chapter 9).

MT VERSUS HUMAN TRANSLATION

What, then, is the status of MT? If we accept that translation
demands total sensitivity to the cognitive aspects of a source text, it
follows that a computer would need to understand language and
assimilate facts in the way that humans do in order to resolve
textual ambiguity and create a version that paid due regard to
semantic content and register. For example, an awareness of
context is essential for the correct interpretation of a sentence such
as visiting European dignitaries can be a nuisance. In translating this
sentence, a human translator would take into account the sentences
which preceded and followed, as well as the general context, the
4 John Newton

overall theme of the text and any relevant social, economic or

cultural factors. However, a computer’s inability to acquire,
comprehend and rationally apply real-world knowledge in this way
does not render MT useless as a production tool. Raw MT output
does not need to be perfect in order to be useful. With the
exception of a few specific applications (e.g. unedited raw
translation used for information only purposes), it is rarely regarded
as a finished product; like other raw materials, it is converted into
a finished product only through human agency (i.e. post-editing).
Direct comparisons between a system’s raw output and human
translation are therefore pointless; as MT is a production tool, its
capacity to increase or speed up production, within acceptable cost
parameters, is the only valid measure of its effectiveness. If its use
can be shown to increase productivity and/or reduce costs, it is
clearly advantageous; if it fails to do either, it is a white elephant.
In Chapter 10, Yorick Wilks presents for the first time a detailed
description of the methodology he used in 1979–80 when
conducting an evaluation of the SYSTRAN Russian to English MT
system for the US Air Force. After explaining the rationale behind
the test and enumerating his evaluation criteria, he gives a full
account of the implementation procedures and follows this with an
analysis and discussion of the results obtained. This document
should prove invaluable to anyone seeking a blueprint for MT
evaluation.
Contrary to a view sometimes expressed by people remote from
the translation process, a source text lexical item does not necessarily
have just one ‘correct’ target language equivalent. If this is true at the
word or term level, it applies even more at the level of the clause or
the sentence. The notion that there is always a unique and obvious
way to translate any combination of words from one natural
language to another is fallacious in the extreme. The number of
word combinations possible in any natural language is
unquantifiable and infinitely expandable, yet humans usually
manage to infer intended meanings from context. The headline
‘Navy bases are safe’, which appeared in a Hampshire (U K)
newspaper in February 1991, illustrates the importance of context.
Anyone unfamiliar with the local scene could have assumed that
terrorist threats had been made against naval bases, or that there had
been fears of radiation or some other form of contamination. The
headline was, in reality, aimed at dispelling rumours of closures and
consequent job losses. Anyone who had followed the controversy
 Introduction and overview 5

surrounding the bases would have grasped the intended meaning

immediately, and those without such prior knowledge would have
been fully enlightened by the first paragraph of the article, which
read: ‘The Royal Navy is set to expand its bases in the Solent…fears
that bases could be closed …were unfounded’.
Nevertheless, the human’s ability to infer meaning does not
always present translators with clear-cut solutions to translation
problems. Producing a translation which may need to convey
concepts, attitudes and physical realities alien to the target-language
culture demands considerable resourcefulness and creativity. Several
acceptable renderings of a sentence may be proposed but the
adoption of one in preference to the others must ultimately be based
on subjective criteria. No human translator can claim to render the
semantic content and stylistic features of any source text with total
fidelity, and it is unlikely that any two translators would ever
produce identical renderings of anything but the shortest and
simplest document. It is equally unlikely that a single human
translator would produce identical versions if asked to translate a
text of any substance on more than one occasion. These points are
made here solely to dispel any notion that translation is a
straightforward process involving clear choices.
Having stated that human translators usually score over MT
systems in the areas of interpretation and preservation of register, it
is important to stress that when it comes to spelling and
terminological consistency the computer invariably outperforms
the human. In these areas, it simply cannot make mistakes; all
spelling anomalies in its output are attributable to human error in
dictionary compilation or updating; likewise, it cannot deviate
from the translation equivalents that are entered (see Chapter 4 for
a description of dictionary entry in a batch-processing MT system).
Consumers of translation sometimes attempt to overcome the
problem of terminological inconsistency in human translation by
distributing glossaries—or document dictionaries, as described
above—to translators but this does not, in itself, ensure that the
prescribed terminology will be used exclusively and consistently
throughout a translation.
Computers can handle any number of variant spellings,
provided the dictionary entries and rules allow for them (e.g.
British and American variants ‘sceptical’ / ‘skeptical’; other
variants such as ‘time frame’ / ‘time-frame’ / ‘timeframe’ or
‘judgement’ / ‘judgment’), and render them with a single target
6 John Newton

language equivalent, if this is desired. Nevertheless, systems are

usually designed to deal only with standard syntax and grammar.
This means that they are unable to compensate for human errors
and ideolectal deviations in the form of non-standard syntax, usage
and punctuation, or inadvertently missed or repeated words.
Moreover, they obviously lack the capacity to comprehend what an
author intends to convey when he/she deliberately deviates from
the established conventions in a very personal way. In all of these
cases, a human translator will invariably deduce the author’s
intended meaning and produce a translation which is more
coherent than the source text. However, the ability to exercise
judgement in interpretation does not guarantee consistent use of
terminology, and it is not unknown for a human translator to
render recurrent sentences in different ways within a single text.
This does not imply that one rendering is ‘wrong’ and the others
‘right’, or even that one is necessarily ‘better’ than the others. Such
stylistic variation may be prized in some forms of writing but in
technical manuals and user instructions it can be disorientating and
is generally frowned upon. The problem of human inconsistency is
particularly acute when a text has to be divided between a number
of translators in order to meet a delivery deadline or on account of
its sheer volume. These cases highlight collective as well as
individual inconsistency and editing is generally needed to achieve
an acceptable level of homogeneity.
Spelling mistakes and typing errors are not at all unusual in
human-translated output, nor are serious errors of usage and
grammar (is anyone infallible in all of these areas?). Computers can
help here, of course, but many people who have spell checkers and
grammar checkers fail to use them (complacency, or pressure of
work?). When considering the merits or demerits of MT, we
should not therefore labour under the false notion that human
translation, or even edited human translation, is always perfect.
The very fact of being human creates a potential for error, and
instances of whole sentences, paragraphs or even pages being
‘skipped’ in translation are by no means uncommon, nor are
mistranslations. I have seen warnings on technical products
translated in a way that makes them say the opposite of what the
source text said. How can this happen? It is quite simply the
human factor at work. The world knowledge and life experience of
human translators can help them to grasp the precise meaning of a
sentence or passage but it can also cause them to make assumptions
 Introduction and overview 7

which are wrong. A translator who has been translating all day can
easily fail to spot a deviation from what appears to be a uniform
sequence of messages as the working day draws to a close. If
‘never’ occurs just once in a list of safety instructions beginning
‘always’, it is not difficult to imagine how an error could creep in.
In large organisations where translations are routinely proofread by
other translators or editors, errors of this kind would normally be
discovered and corrected but many translators have to work
without this safety net.
As stated above, MT systems cannot deviate from the translation
equivalents that figure in the dictionary or dictionaries specified for
any particular task. However, if the human-supplied dictionary input
is wrong, the system will consistently reproduce errors in raw
translation until the entry is corrected. Fortunately, most MT output
is post-edited, and a post-editor, like the proofreader of human
translation, can be expected to notice problems of this kind and take
corrective measures in the text and in the dictionary. Moreover, MT
does not carry the risk of lines, paragraphs or pages being
unintentionally ‘skipped’ in translation.

WHERE CAN MT BE USED TO ADVANTAGE?

The suitability of MT for any particular task depends on the nature
of the source text and the intended purpose of the translation;
several contributors to this book make this point (Chandioux, Gross,
Melby, Newton). Poetry and industrial parts lists represent two
extremes in this regard, for while it would be absurd to contemplate
using MT for the former, one could reasonably expect a perfect
result (requiring no editing) for the latter; provided, of course, that
all the relevant nomenclature was in the system’s dictionaries.
It is clear then that machine translation cannot be used
indiscriminately. Many source texts have to be excluded from MT
on the grounds that their text typology precludes anything but the
attention of a skilled and highly specialized human translator.
Literary works are obvious candidates for exclusion, as are all kinds
of advertising and promotional text. Indeed, for the latter category,
translation alone will not suffice; a complete adaptation is needed to
take account of the native aspirations, susceptibilities and prejudices
of the target market.
Notwithstanding the large volume of text that it would be
impractical or imprudent to submit to MT, there remains a
8 John Newton

substantial and growing body of technical and scientific text that can
be handled efficiently and profitably using computers if the requisite
operational conditions exist, as is exemplified by the applications
described in Chapters 3, 4 and 5 of this volume. It would,
nonetheless, be unrealistic to expect any system to perform well with
source texts drawn randomly from a wide range of subject areas and
text typologies.
I have stated above that an MT system’s raw output—‘raw
translation’—is not usually regarded as a finished product. This
reflects the fact that most users practise post-editing to some degree.
The techniques applied range from rapid post-editing, for
information-only purposes, to comprehensive ‘polishing’ aimed at
making the finished product indistinguishable from human
translation. Most MT systems have a synchronized split-screen
editing mode which enables the post-editor to view and scroll the
raw translation in tandem with the corresponding section of source
text. Quality control rests with the post-editor and must be assumed
to be as rigorous as that applied to human translation produced for
similar purposes.
Syntax errors in raw translation are only a problem if they are so
numerous that post-editing requires as much effort as normal
human translation (see Chapter 3: Annette Grimaila and John
Chandioux). It is always advisable to conduct tests to ascertain
whether MT is suited to a particular environment before incurring
the expense of purchasing a system, but if such tests are to be fair
to the prospective purchaser and to the system, they should involve
a substantial volume of truly typical text and the requisite level of
dictionary updating. Anything less than this is unlikely to yield an
accurate picture.

THE HUMAN FACTOR IN MT

If MT is to be used to maximum advantage, those responsible for
operating the system must be receptive to its introduction. It would
be unwise and unfair to impose MT on translators who had not been
consulted and given an opportunity to express their highly relevant
views on a proposal which has such momentous implications for
them. In common with other major computer applications, MT
systems demand a considerable learning effort from their operators;
it would therefore be unrealistic to expect translators to maintain
their normal workload while getting to grips with MT. Moreover, a
 Introduction and overview 9

system can only be fully exploited if users are trained in all aspects
of its operation and feel generally comfortable with it; thus, any
attempt to take shortcuts in the area of training would be likely to
have a very deleterious effect on the results obtained and on the
translators’ attitude to MT.
In Chapter 7, Alex Gross makes the point that MT demands
special skills of a very high order on the part of the operator.
Introducing it into a department which may lack the requisite
aptitude or commitment is therefore unlikely to produce optimum
results. While some translators take to MT like the proverbial duck
to water, others find the prospect of adapting to it extremely
daunting. My experience of training groups of translators to use MT
systems suggests that a good translator does not necessarily make a
good MT operator; although I did find that most translators—
including some who were nearing retirement—were able and willing
to make the necessary adjustment.
There should be no conflict of interest between human translators
and MT. The latter has proved beneficial to some translators
through opening up a new and challenging career, and the publicity
MT receives does, at least, bring discussion of translation into the
broader public domain. Furthermore, the presence of an MT system,
in a clearly subordinate role, serves to highlight the skills of the
human translator/post-editor. As one MT provider accurately
described it, ‘the system takes care of the donkey work, while the
human translator concentrates on fine-tuning the dictionaries and
polishing the raw output’. We should also recognize that some of the
tasks assigned to MT, e.g. information-only translation (as discussed
by Muriel Vasconcellos and Dale Bostad in Chapter 5), and
translation of time-sensitive restricted domain texts (as described in
Chapter 3), would probably not be performed at all if MT were not
available.

ATTITUDES TO MT
There are a number of (largely apocryphal) stories in circulation
about bizarre computer renderings of well-known phrases and
idioms; the most notorious of these is probably ‘out of sight, out of
mind’, allegedly rendered as ‘invisible idiot’. It calls to mind an
occasion when I was invited to demonstrate a French-English MT
system on television. It was clear from the outset that the presenter
wanted to maximize the entertainment value of the situation. After
10 John Newton

an introduction which made exaggerated claims for the system, he

gleefully produced a piece of paper bearing the words, Plongeur dans
un restaurant, il avait du mal à gagner sa vie and asked me to enter the
sentence and translate it. Given the fact that the only translation
for plongeur in my dictionary was diver, the desired comic effect was
achieved. Fortunately, I had time to create another entry for
plongeur, in a different dictionary, giving washer-up as its translation,
and I retranslated the sentence using a different dictionary look-up
sequence. This enabled me to demonstate that, with MT,
vocabulary and terminology are totally controllable. So stories
about ‘hydraulic rams’ being machine translated as ‘water goats’
should be treated with the contempt they deserve, because they are
evidence not of an MT system’s inability to handle vocabulary but
of a raconteur’s lack of familiarity with the workings of an MT
system.
Much criticism of MT is unjustly directed towards its inability
to perform tasks it was never designed to tackle (e.g. literary
translation). On the other hand, the extravagant accuracy claims
sometimes made for it by salesmen do nothing to enhance its
image among the better informed. Alex Gross (see Chapter 7)
points out the absurdity of attempting to quantify translation
accuracy in percentage terms, and also notes the ambivalence
displayed by some laymen who expect perfect results from MT,
but who are nevertheless among the first to chuckle at its alleged
bizarre renderings. In the same vein, I have met people who
refused to heed warnings about the unsuitability of MT for a
particular application, only to become indignant and disillusioned
when tests they had insisted on conducting proved this to be the
case.
Over the past two or three years, numerous cheap hand-held
‘translation’ devices have been promoted through mailings and off-
the-page advertisements. Although they are little more than
electronic dictionaries with very limited vocabularies, they have
helped to spread the notion that automatic translation is routine and
commonplace. Even the Collins English Dictionary (second edition
1986) encourages this assumption by defining ‘translator’ as ‘a
person or machine that translates speech or writing’.
Another misconception concerning MT stems from a false
analogy drawn between the unrestrained and flexible nature of
natural language and the exactitude and precision of mathematics.
Because computers can cope with the intricacies of a spreadsheet, it
 Introduction and overview 11

is assumed that they can be made to produce flawless on-demand

translations of any source text. If each word in a language had
one—and only one—clearly circumscribed value, and if the total
value of a group of words could be ascertained by adding together
their individual values, this analogy might have some merit. The
reality, however, is very different and those who propound this
theory would do well to ponder natural language a little more
intently.
Translators who have had direct involvement with MT tend to
have a realistic and positive attitude towards it; they know what it
can do and they are aware of its limitations. Some find working
with a system far more challenging and fulfilling than traditional
translation methods. However, as working environments which
lend themselves to MT are so rare, relatively few translators have
the opportunity to familiarize themselves with the workings of a
system. Some translators no doubt feel that MT is so remote from
their own activities, or that it accounts for so small a portion of
translated text, that it is irrelevant; however, most of those I have
spoken to expressed a keen desire for more information and it is
hoped that the present volume will go some way towards meeting
their needs.

CONCLUSIONS
Along with other writers, translators were quick to realize the
benefits of computers. Word processing is now used almost
universally among translators and translation tools and
computerized dictionaries are steadily gaining ground.
Commercially-available MT systems are often designed to
meet wide-ranging needs. The fact that potential users tend to
have very specific needs renders a large proportion of an off-the-
shelf system’s capability superfluous in most applications. It
would therefore seem desirable to extend the principle of subject
specialization, as practised by human translators, to MT. Bespoke
systems (as described in Chapter 3) undoubtedly maximize the
potential for success. At some future time, it may even be possible
to assemble systems from ‘bolt-on’ modules, selected or adapted
to cope with specified features of a particular text type; for
instance, in an English-French context it would be useful to be
able to choose how the negative imperative should be handled. If
this idea does become a reality one day, what is left out (for any
12 John Newton

single application) will be every bit as important as what is

included.
The fact that MT is already making a valuable contribution in
some areas is amply evidenced by the first-hand accounts provided
in the present volume. Nevertheless, it is important to bear in mind
that these successes are the fruit of a substantial amount of
preparation and the deployment of considerable human, physical
and financial resources. Moreover, putting in place the elements
needed for implementation does not in itself ensure success. Many
text types simply do not lend themselves to MT. Before embarking
on any new MT installation it is therefore imperative that tests be
conducted to determine its feasibility.
MT’s much-publicized propensity for producing the occasional
amusing rendering (e.g. engine views translated into French as le
moteur regarde or traffic jam as confiture de circulation) is irrelevant if
post-editing its output results in totally acceptable versions being
produced more quickly and/or at lower cost than with other
processing methods. In any event, repetition of such
mistranslations is easily avoided through dictionary updating, and
once the appropriate translation equivalents have been entered, the
system will use them unfailingly; it is worth restating that (given a
translation task of any magnitude) MT should always have the
edge over the human translator as far as terminological consistency
is concerned.
The emergence, in recent years, of international networks of
translation companies, and of translation companies specializing in
one or more (invariably high-tech) domains, is a strategic response
to changes in the pattern and nature of translation demand. These
developments, creating as they do larger pools of translators—either
on-site, or via on-line link-ups—tend to favour the spread of
translation tools and translator workstations. At the same time, the
standardization of formats for multilingual terminology databases
(see Chapter 8) should encourage translators to make greater use
of them.
It is, perhaps, regrettable that translation tools do not attract the
same level of media coverage as MT, for unlike the latter, they can be
used by any translator for almost any text type (subject to the
requisite language pairs or modules being available). Translator
workstations are similarly broad in their appeal and in the range of
their applications; when a standard specification emerges, it will no
doubt be warmly received.
 Introduction and overview 13

The diverse interests and activities of the contributors to this

volume give an indication of the potential that computers have to
offer in the field of natural language translation. There can be little
doubt that the balance of the 1990s will witness an increase in the
number of commercial MT installations, along with significant
growth in the use of translation tools.
Chapter 2

The story so far:

an evaluation of machine translation
in the world today
Jeanette Pugh

INTRODUCTION
This chapter presents an assessment of the position of machine
translation (MT) in the world today. In this assessment I address
questions such as: how was the current situation of MT arrived at?
What does MT mean today? How is MT perceived in different parts
of the world and by different sectors of society? What contribution
does MT have to make now and in the future?
It is assumed that the reader is familiar with the general
background to MT, and only a very brief historical survey of the
field is given. The emphasis throughout the chapter is on the
contemporary position and status of MT; hence, the origins and
development of the field to date are considered only to the extent
that they may shed light on modern attitudes and perspectives.
To circumscribe the domain under examination, a short overview
of the kinds of activities that make up ‘machine translation’ today
and how these relate to other associated disciplines and domains is
given. The purpose of this overview is to situate MT in a scientific
sense, and to give the reader a feel for the divergences in breadth and
depth of MT-related activities in different parts of the world
(Western Europe, Eastern Europe, the United States, Japan, etc.),
and among different sectors (academic research, the public sector,
commercial research and development, international initiatives, etc.).
In the main body of the chapter, I present an analysis of the
present position of machine translation and explore the reasons for
the divergences identified earlier. Obviously, a thorough, in-depth
analysis of all the elements which emerge would be beyond the scope
of this chapter. What I will aim at, then, is rather a broad review of
possible factors underlying such differences as those in national and

14
 Machine translation today 15

governmental attitudes, funding priorities, variations in levels of

‘intellectual’ commitment, etc. In addition, I will pay attention to
issues of public awareness: popular perceptions—and
misconceptions—about MT, and how these also vary.
My analysis will be performed not only from a geographical
viewpoint, looking at the different situations in different parts of the
world, but also from the point of view of different groups of
individuals within and across countries. Such groups include, for
example, MT and natural language processing (NLP) researchers,
end-users, customers, industry, etc. I will examine attitudes, usage
and beliefs about the actual and expected benefits of MT-related
activities and products.
I will conclude with a brief look to the future, speculating on
possible shifts that may occur in the position and standing of MT.
References to sources of further information are also provided.

HISTORICAL PERSPECTIVE
The origins and development of machine translation have been well
documented and I wil not give yet another detailed historical
account here (the interested reader is invited to consult Hutchins
1986 which is an excellent source). However, while the historical
facts are amply recorded, their significance today and, in particular,
the extent to which they have shaped modern attitudes to and within
MT, have been paid relatively little attention. In this part of the
chapter, I review the history of MT from a contemporary
perspective.
While the concept of mechanical translation can be traced back as
far as the seventeenth century, the enabling technology for the
present-day concept of ‘machine translation’ appeared less than fifty
years ago. The electronic digital computer made its first impact in
the Second World War. In the immediate post-War period, there was
a natural move to explore its potential capabilities, and the
translation of languages was soon identified as one of the obvious
candidates for exploitation.
These early explorations reached a watershed in 1949 with the
famous Weaver memorandum, which ‘in effect…launched machine
translation…in the United States and subsequently elsewhere’
(Hutchins 1986:28). The significance of the Weaver memorandum
is undeniable. Interestingly, it focused on the general strategies and
long-term objectives of machine translation rather than on the
16 Jeanette Pugh

concrete technical problems which had thus far emerged. Weaver

raised four points: the problem of multiple meaning and the
resolution of ambiguity, the logical basis of language, the
application of communication theory and cryptographic
techniques, and the possibilities of language universals. He thus
identified many of the fundamental problems of MT which remain
a challenge today.
Weaver’s memorandum sparked off a wave of interest not only in
the United States but also across the Atlantic in Britain. The first
MT conference was held at the Massachusetts Institute of
Technology (MIT) in 1952. Of the eighteen participants, seventeen
were U S and one was British. Four years later, the first real
international conference on MT was held, again in the United States,
but this time with US, British and Canadian delegates, and with
contributions from the Soviet Union which by then had also
established a research basis in the field. From 1956 onwards, activity
in MT flourished in many different parts of the world. In the United
States, it was given increasingly substantial financial support from
the government and from military and private sources. In 1966,
however, the bubble burst, with the publication of the Automatic
Language Processing Advisory Committee (ALPAC) report
(ALPAC 1966) which recommended the cessation of all MT
funding. The mood had changed from enthusiastic optimism to
fatalistic condemnation in less than a decade.
Viewed from today’s perspective, the English-speaking origins of
MT in the light of the subsequent fortunes of MT in the United
States are perhaps ironic. In a sense, then, the present apparent
resurgence of interest in American MT (see below) marks the return
of a ‘prodigal son’ to the fold.
The strong military influence in the emergence of MT is also
interesting in retrospect. A prime area for MT application remains
that of defence and related concerns today, and MT research is
conducted or supported by a number of military establishments, so
this aspect of MT has persisted. However, more important from a
scientific point of view is the original association of MT with
cryptography, which was very largely reponsible for the adoption of
over-simplistic techniques that ultimately proved flawed and led,
eventually, to the ALPAC condemnation.
While the ALPAC report was specifically directed towards an
assessment of the value of MT R&D in the United States, it had a
far-reaching negative impact and caused the virtual cessation of MT
 Machine translation today 17

funding in the United Kingdom and a severe slow-down in funding

in France and the Federal Republic of Germany, all of which had by
that time established MT research groups. The fortunes of MT
activities in Eastern Europe, where research groups existed in
Hungary, Czechoslovakia and the Soviet Union, were relatively
unaffected by ALPAC and flourished in the next decade, despite
technological disadvantages. In Western Europe, activities also
continued, albeit in a subdued atmosphere, and in 1967 the
Commission of the European Communities (CEC) started its first
tentative research on MT at its EURATOM establishment in Italy.
Considerable progress was made in the years up to the late 1970s,
which in retrospect can be seen as a period of reflection, learning and
of gathering strength.
The late 1970s saw a dynamic new impetus for MT in Western
Europe with the launch in 1978 of the ambitious EUROTRA
initiative. The Commission of the European Communities, after
initial abortive attempts to expand the imported American
SYSTRAN system to meet its multilingual needs, decided to start
its own R&D programme in MT. This decision proved revitalizing
for European MT and led to the expansion and consolidation of
existing MT research centres and to the establishment of new
centres in other countries, many of which had no previous
experience in the field. While MT in Europe cannot be uniquely
identified with EUROTRA (see below), its contribution to the
emergence of an integrated MT community with an established
communications infrastructure—which contrasts sharply with the
fragmentation of expertise and resources in the United States—
should not be ignored.
The progress of MT in the 1980s was marked by Japan’s arrival
on the scene. In the years which have followed, Japan has shown a
commitment to MT hitherto unparalleled in any country. During
that decade, MT activities continued to thrive in Europe, while in
the United States involvement was relatively much more limited.
The 1990s and beyond seem to herald a greater internationalization
of MT efforts, and there is reason to hope that the ghost of ALPAC
will finally be laid to rest.

THE SCOPE OF MT TODAY

Originally, the concept of machine translation was more exclusive—
and more uncompromising—than it is generally held to be today.
18 Jeanette Pugh

The goal of fully automated high quality translation (FAHQT) was

the aspiration of most MT researchers in the period up to the early
1980s. By that time, however, the elusiveness of this objective had
become all too apparent, and today FAHQT is more of a dream than
an ambition.
While the horizon may have receded, however, the breadth of the
goals of MT and the realism of those involved in it have expanded.
Computer technology—the engineering sine qua non of MT—which
was at first greeted as a liberating panacea, has over the years come
to be regarded by MT researchers as a tool which is not only
enabling but also constraining. The limitations of the available
technology have helped researchers to shape their ideas about the
process of translation, and forced them to adopt a disciplined,
scientific approach to linguistic problems, with the accent on formal
rigour and computability. In this way, the field of MT has acquired
enhanced respectability and has evolved as an independent area of
activity, although with close associations with other, related fields
such as computer science, artificial intelligence (AI), theoretical
linguistics and NLP.
The range of protagonists in MT has also expanded considerably
since its inception. It has extended beyond the domain of academic
research to include private industry (most notably software houses)
and national governments. In recent years, there has also been an
increasing internationalization of MT activities with established
international conferences and publications, and transnational project
collaboration.
The expansion of the concept of machine translation has been
mirrored by the appearance of a range of acronyms denoting
different levels of automation and human intervention. Thus, we
speak of MAHT (machine-assisted human translation), where the
major translating effort comes from a human user, as opposed to
HAMT (human-assisted machine translation), where the machine
performs the basic translation process and the human operator’s
intervention is restricted to pre- or post-editing texts. The trend in
recent years—which is continuing today—is a move away from large-
scale systems towards an increasing variety of automatic aids for
translation, including on-line dictionary facilities, multilingual
word-processing packages, interactive editing facilities, etc.
(Vasconcellos 1988: Section III). It is these sorts of aids which have
proved most susceptible to commercialization, although larger
systems can also be found on the market. From the human
 Machine translation today 19

translator’s perspective, MT has gradually become more of an ally

than a threat.
I thus take a very broad view of MT which, I feel, is in line with
the way in which the field has developed and is continuing to evolve.
In my survey of the state of MT activities in different parts of the
world, I focus on those projects where the degree of intended
automation of the translation process is greatest, but I have also tried
to make room for MAHT and HAMT as this is where, I believe, the
future thrust of expansion will occur. (For an overview of the
technical state of the art, see Chapter 11: Harold Somers.)

MT IN EUROPE
In Europe, the impact of the ALPAC report was initially dramatic,
but it took little more than a decade for its effects to disappear. As
we have seen, the late 1970s witnessed a veritable explosion of MT
activity in Europe, the most notable initiative being the launch by
the CEC of the EU ROTRA programme which has received
sustained high-level funding from both the European Commission
and the national authorities of all EC member states. It has
involved all twelve EC countries, working on all nine official EC
languages, with some twenty individual research sites and, at its
peak, about two hundred participants (Raw et al. 1989; Steiner
(ed.) 1991). At the time of writing, the EUROTRA programme has
taken a new and exciting turn, with the move to a two-year
transition programme in which both the scope of its activities and
the range of participants will be much more diversified. The
primary aim of this two-year programme is to prepare the
transition from EU ROTRA’s pre-operational prototype to an
industrialization of the system. This will involve the active
participation of European industry which, it is hoped, will invest
not only manpower but also financial resources. The EUROTRA
stage has thus become much wider and the range and roles of its
actors more varied (EEC 1990).
Although EUROTRA has dominated the European MT scene by
virtue of its sheer size, it is by no means the sole MT effort in
Europe. There are also a number of national MT programmes which
testify to significant public-sector commitment, as well as substantial
involvement by the private sector.
France, Germany and the United Kingdom have the longest
traditions in MT in Western Europe. The famous French centre,
20 Jeanette Pugh

Groupe d’études pour la traduction automatique (GETA), is one of

the oldest MT research groups in the world, and is renowned for its
development of the ARIANE system. Recently, France’s largest
technical documentation company, SITE, took on ARIANE and,
with public- and private-sector support, will collaborate with GETA
to move towards the eventual commercialization of the system.
Meanwhile, GETA continues to remain in the forefront of European
MT with new work on the development of interactive MT systems
or ‘dialogue-based MT’, where translation is performed via a natural
dialogue between the user and the system.
The largest university-based MT centre in Germany is at
Saarbrücken where the SUSY system was developed (Maas 1987),
with financial support from the Federal German government from
1972–86. MT activities at Saarbrücken have not been confined to
work on SUSY, however, and it has been involved in a number of
other projects, including EUROTRA, and some recent work on
‘text-oriented MT’. There are also active MT centres in Stuttgart,
Berlin and Bonn. The German government has an admirable
record of supplying financial backing to MT endeavours, to the
extent that one American observer was prompted to remark
recently that ‘German government funding for MT R&D alone, not
counting EEC sponsorship, is presently about ten times the entire
United States government R&D funding for MT’ (Carbonell
1990:12 0). Germany also has significant private-sector
involvement in MT R&D, the most notable example being
Siemens’ development of the METAL system (Thurmair 1990)
which was judged at the 1989 MT summit in Munich to be among
the most impressive demonstrated. There is also a substantial MT
user community in Germany, and a sizeable domestic market for
MT-related products.
The United Kingdom, consistent with its early influential role in
the field, continues to have a strong MT presence. While several of
its universities are active in NLP in general, the most specialized MT
centre is that in Manchester (U M I ST) at the Centre for
Computational Linguistics (CCL) which, in addition to being
involved in EUROTRA (together with the University of Essex),
participates in several MT projects. One of the earliest was NTRAN,
which was publicly funded from 1984–7 from the national Alvey
programme in information technology. NTRAN was a document-
processing and English-Japanese translation system which was
innovative in being designed for use by monolingual technical writers
 Machine translation today 21

(Whitelock et al. 1986). The prototype was successfully

demonstrated but implementation was stopped due to lack of further
public or industrial funding. Since 1987, CCL has been involved in
a number of other MT projects with British and, notably, Japanese
industrial backing.
In the UK private sector, there is a certain amount of activity in
MT-related R&D. British Telecom Research Laboratories are
carrying out their own research on automatic speech translation
(Steer and Stentiford 1989). Such industrial involvement in MT in
the UK is rare, however, and in general the level of commitment is
limited, and is certainly less than is to be found among its
continental European counterparts where large companies such as
Siemens, Philips and IB M-Europe are all making a significant
investment in this area. The extent to which this relative lack of
commitment can be attributed to the oft-cited British linguistic
chauvinism is a matter for debate, but as a possible contributory
factor it cannot be dismissed out of hand.
Elsewhere in Europe, industrial and public-sector interest in
MT is also in evidence, in addition to the multinational
investment in the EUROTRA programme, which I have already
mentioned. The distributed language translation (DLT) project is
a good example of a national MT effort with combined
sponsorship. A multilingual MT project, DLT is located in Baarn
at the software house Buro voor Systeemontwikkeling (B SO). It is
currently in a seven-year R&D period (1985–92), jointly funded
by BSO and the Netherlands Ministry of Economic Affairs (see
Chapter 6, this volume and Pappegaaij et al. 1986). Interestingly,
the Netherlands is also home to another major MT project with
private-sector backing in the form of the Rosetta project, which is
based at Philips Research Lab oratories at Eindhoven
(Landsbergen 1987).
In Switzerland, MT has also become an established activity. The
main MT centre is ISSCO, a research institute in Geneva which is
privately owned but which has had administrative links with the
university. There, work is under way on three main projects (King
1989). The first is in the area of text generation of avalanche
warnings; the second involves the development of a prototype MT
system; and the third concerns the evaluation of MT systems, which
is one of the current ‘hot topics’ in the field.
Industrial involvement in European MT has led in recent years to
a proliferation of MT and MT-related products. The main
22 Jeanette Pugh

commercial companies which have marketed such products in

Europe are Gachot, Logos, ALPNET, Globalink and the London
and Tel Aviv-based Tovna. The oldest commercialized system is
SYSTRAN, which has been available now for well over a decade.
SYSTRAN was originally based in La Jolla (California) but was
taken over in 1988 by the French company Gachot. Its main user is
the European Commission in Luxembourg to whom the SYSTRAN
system was sold under special licence. For general commercial
purposes, Gachot sells SYSTRAN output, rather than the system
itself, on diskettes, via modem or even over the French public
telecommunications system, Minitel (Trabulsi 1989).
Other available systems and translation aids include those
marketed by Logos, which has a European base in Germany (but
whose headquarters are in New Jersey in the United States) and
Globalink, which is again based in the USA. Until recently, another
American company, ALPNET, also marketed its products in Europe
from a base in Switzerland. It offered arguably the most sophisticated
system on the European market, with a suite of translation tools
ranging from on-line dictionary facilities to fully interactive
translation, although it covered a more restricted number of
language pairs than some of its competitors. Ironically, however,
some translators find its interactive translation facilities
(TransActive) cumbersome.
The commercial systems I have mentioned here are probably the
largest and best known but they do not account for the entire
European market in MT-related products. These cover a whole
gamut of applications ranging from multilingual word-processing
packages and sophisticated desk-top publishing (DTP) facilities to
multilingual dictionaries and term banks and multilingual OCR
software (for an illustrative catalogue of such products, see the
‘Multilingual Wordworker’s Resource Guide’ in LT/Electric Word, 13/
1989).
Some further observations are worth making here. First, it is
ironic that the most commercially successful systems in Europe all
have US connections. Second, while Western Europe represents a
huge multilingual community, the MT products available for the
most part deal with bilingual translation only, a situation which
may perhaps change as a result of the EU ROTRA transition
programme. Finally, although the quality of the available systems is
in many cases far from ideal, and is largely based on outdated
technology, the systems do sell. The market is substantial and is
 Machine translation today 23

continuing to grow. Indeed, demand probably exceeds supply—

which may account for the apparent customer satisfaction with
low-quality products.
Before moving on, let us look briefly at the public profile of MT
in Western Europe. Over recent years, some ‘popular’ journals (e.g.
Electric Word) have appeared, although generally they have been
rather too specialized to sustain broad commercial success.
Typically, such publications carry publicity for commercially
available products and ‘layman’ articles on aspects of MT and
related activities (multilingual word processing, speech analysis,
etc.). They are thus aimed at a narrow sector of the industrial
market. MT also receives attention from time to time in computing
publications (e.g. PC World, Computing, etc.), again targeting a
specialized audience. Occasionally, interest is shown by the general
media but while such exposure undoubtedly serves to heighten
public awareness, it is by no means clear that it improves public
understanding of the issues involved, and the time-worn recourse
by journalists to examples of ‘amusing’ mistranslations is still very
much apparent.
Finally, we should not neglect to mention Eastern Europe where
there is a thriving MT community, notably in Czechoslovakia,
Hungary, Bulgaria and the republics of the former Soviet Union.
Efforts have been fairly static since the mid 1980s but the opening up
of Europe seems likely to lead to a revitalization of activities.

CANADIAN MT—A SPECIAL CASE

Canada is a noteworthy example of a country with an enlightened
approach to machine translation. Canadian public-sector interest in
MT stems from its commitment to bilingualism, and Canada has the
claim to fame of being the first country in which an MT system
(METEO) was put to widespread public use. The METEO system
was originally developed at the University of Montreal and
translates meteorological bulletins from English to French
(Chandioux 1989). (See Chapter 3, this volume) While its
development was carried out with government sponsorship, in 1984
the system was taken into private ownership and a more efficient
reimplementation of it was developed. The Canadian government
has remained committed to the use of this high-quality system and
now pays for the service. Moreover, METEO is not the only
example of the Canadian government’s interest in MT. In addition,
24 Jeanette Pugh

it has sponsored several wide-ranging surveys of available systems

conducted by its large translation department.
MT work is also being carried out at the Centre Canadien de
Recherche en Informatisation du Travail (CCRIT) in Montreal
(Isabelle 1989), and at the University of Ottawa.
As a result, no doubt, of its particular linguistic circumstances,
Canada has always shown a keen awareness of the value of
goodquality translation. With its commitment to MT, it has also
exhibited a willingness to experiment and to develop an advanced
approach to its translation problems. That such commitment should
exist in the public sector is indeed enlightened, and stands in stark
contrast to the attitudes which have been displayed by its North
American neighbour.

THE STATE OF AMERICAN MT

In the United States, perhaps not surprisingly, the damning effects
of the ALPAC report have proved more difficult to shake off than
in Europe. ALPAC hit American MT very hard. The twenty-five
years which followed have been referred to as the ‘dark ages’ of
American MT and while this assessment is perhaps an unfair
reflection on the quite substantial research activities which have
survived in the USA despite ALPAC, it certainly does seem an
appropriate epithet when one compares the situation with the
progress of events in Europe and Japan. However, it is now widely
believed that the United States is at long last starting to emerge
from the ‘dark ages’, so that the 1990s could very well become the
Renaissance of American MT, leading perhaps—who knows?—to a
Golden Age in the next century.
At present, there are relatively few active MT research centres in
the United States. The largest is the Center for Machine
Translation at Carnegie Mellon University (CMU), whose work is
primarily supported by private funding, over half of which comes
from abroad (Nirenburg 1989). New Mexico State University
(NMSU) (Farwell and Wilks 1990) and the University of Texas at
Austin (Bennett and Slocum 1985) also have major MT groups,
and there are smaller efforts at New York University (NYU),
Brigham Young and Georgetown. The U S National Science
Foundation (NSF) provides some $350,000 per year to CMU,
NMSU and NYU, and it is hoped that a DARPA initiative will
boost this with another million. ‘The total amount, however, would
 Machine translation today 25

still be significantly inferior to the budget of a single well-sponsored

Japanese MT laboratory’ (Carbonell 1990:12 0). The Pan
American Health Organization (PAHO) has also developed MT
systems but the only large industrial company with a serious MT
effort in the USA is IBM.
The preceding overview illustrates the limited scale of MT
activity in the United States. It has been suggested that this is, at
least in part, attributable to the fact that Americans generally regard
translation as a low-status profession, an attitude which differs
sharply from that in Europe and Canada where ‘translation is a fact
of life rather than an oddity’ (Slocum 1984:546).
Despite such entrenched attitudes, there is evidence to suggest
that American MT is about to enjoy a new lease of life. Moreover,
the resurgence of interest is not confined to the private sector
looking to jump on the bandwagon of Japanese and European
successes but is also increasingly apparent within the U S
government. In December 1989, the US Department of Commerce
sponsored a national conference on Japanese-English machine
translation at the National Academy of Sciences (Valentine
1990:167–213), at which the keynote speaker was the foremost
Japanese MT expert, Makoto Nagao. It is widely believed that this
event did much to stimulate the US community. Subsequently, a
US task force organized a study team to visit Japan to review the
state of the art of Japanese MT and to prepare for specific
cooperative projects.
The reasons for the renewal of interest and activity in American
MT are many and diverse. In a recent hearing (itself a milestone in
recent American MT history) before the Subcommittee on Science,
Research and Technology of the Committee on Science, Space and
Technology in the US House of Representatives, the Assistant
Secretary for Technology Policy in the U S Department of
Commerce put forward her suggested reasons for the revival. A
major factor identified is economic: the massive fall in the cost of
computing, and the parallel rise in the costs of human translation,
especially for non-Roman-alphabet languages, notably Japanese.
This greater economic facilitation of MT R&D is reinforced by the
very considerable progress which has been made over the last
quarter of a century in computational linguistics (CL) and in
language theory in general. Moreover, the increasing number of
available translation software products and tools like large
automated dictionaries demonstrate the practical feasibility of MT,
26 Jeanette Pugh

reinforced by the emergence in Europe and Japan of a market for

MT and MT-related products. Furthermore, the central role of Japan
in recent technological advances makes the need for speedy access to
translated Japanese documentation crucial. Finally, machine
translation has emerged as an important test-bed for the larger fields
of computational linguistics, NLP and AI in which the US has
sustained a high level of research activity. From this point of view,
therefore, MT is not only an end in itself but an important means to
other objectives.
The apparent revival of interest in MT in the USA within
government, the private sector, the academic community and even
the Amerian public appears to stem from a mixture of motives
ranging from economic necessity through technolog ical
chauvinism to scientific aspiration. As I have indicated, the main
focus of the new interest in the United States appears to be on
translation from Japanese to English, which is scarcely surprising
given the economic motivations. The prioritization of this language
pair (and direction) seems bound to have important consequences
for MT not only at national level within the USA but also on an
international scale, leading to more specific collaboration between
the USA and Japan.
As a footnote, it is interesting to note that Washington DC was
chosen as the venue of the third MT summit in 1991. The previous
summits were held in Japan (1988) and Germany (1989), and the
decision to hold the third in the United States was itself illustrative of
the return to centre stage of American MT.

MT IN JAPAN
Although the future prospects for MT in the United States look
better now than at any other time in the post-ALPAC period, the
current situation still compares poorly with that in Japan. There,
MT enjoys a privileged status and is a highly valued, high-priority
activity in which an enormous investment of public and private
financial and human resources has been made. Every major
Japanese computer or electronics firm has invested considerable
effort in MT R&D, and many claim to have developed operational
systems.
Moreover, this investment is not confined to the private sector. A
major, long-term initiative in MT R&D was launched by the
Ministry of International Trade and Industry (M ITI) which
 Machine translation today 27

sponsored MT activity involving the Electrotechnical Laboratory,

the Science and Technology Agency’s Japan Information Centre of
Science and Technology (JICST) and Kyoto University (Nagao et
at. 1985). The ten-year electronic dictionary research (E DR)
project, launched in 1987, is conducted by MITI in connection
with the fifth generation computer project, and aims at the
development of a detailed dictionary with over 200,000 words and
multiple usages (Uchida and Kakizaki 1989). At Advanced
Telephony Research (ATR) laboratories, research is under way on
telephone dialogue translation (Iida 1989), with the support of the
Ministry of Post and Telecommunications. Recently, M ITI
initiated the Center for International Cultural Cooperation
(CICC) project to assist neighbouring countries to develop an MT
research programme for multilingual translation of Chinese, Thai,
Malaysian and Indonesian, as well as English and Japanese (Tsujii
1989).
The very solid commitment to MT R&D which thus exists in
Japan rests on basic attitudes which appear to differ fundamentally
from those in the United States. Undoubtedly, the quite different
situations of the two countries’ languages are significant in this
respect and account for greater importance being attached in Japan
to translation in general. Yet this does not provide a full
explanation of the contrast in attitudes. At both public- and private-
sector level, Japan demonstrates a strongly held belief in the long-
term benefits of MT, not only as a means to more efficient
translation but also as an essential part of the future development
of an information-based society. Of course, economic
considerations also come into play. A recent study by the Japanese
Electronic Industry Development Association (J E I DA 1989)
estimated that the annual market for translation in 1988 was some
800 million yen, with Japanese-English and English-Japanese
reportedly accounting for 90 per cent of this figure. However,
experts have noted that perceived commercial advantage is not
necessarily the primary motivation for large Japanese companies
(including Mitsubishi Electric, N EC, Fujitsu, Sharp, Toshiba,
Hitachi and Sanyo Electric, to name but a few) working on MT.
Rather, they regard machine translation as a learning tool which
will give them general insights into NLP which they consider will
be a key technology in the next century. As we have seen, this far-
sightedness is shared by the Japanese government which has
backed up its beliefs with massive, long-term financial investment.
28 Jeanette Pugh

Where MT activity has survived in the USA, it has been in the

main oriented towards theoretical research, and little attention has
been paid to development work. In contrast, MT workers in Japan
have by and large adopted a pragmatic, ‘problem-solving’ approach
and have concentrated on building working, eventually
commercializable, systems. An obvious consequence of this is that,
while there are many high-quality American research papers, few
operational systems have been produced, and, conversely, while
there are many working MT systems in Japan, the extent to which
they are the result of sound theoretical research is unclear.

MT IN THE FUTURE
In this chapter, I have sought to give the reader a ‘snapshot’ view
of the current status of MT in different parts of the world, focusing
on the contrast in attitudes between public and private sectors, and
on the varying policies of national governments. One fact which
clearly emerges is that MT is today a recognized international
scientific field with a worldwide community of researchers. Yet,
while its international status is now firmly established, the standing
which it enjoys at national level is by no means uniform across the
globe. Japan’s perception of MT as a key element in the future
development of an information-based society and its consequent
long-term commitment to MT activities lies at the extreme end of
the spectrum of national attitudes. In Europe, there is a solid
tradition of MT, and the EUROTRA programme has done much
to improve collaboration and to consolidate and expand expertise.
It is to be hoped—and expected—that future European research
programmes will ensure that ample place is given to MT and
related issues. As for the United States, the signs are very
encouraging and we can expect that, following the enlightened
example of Canada, the USA will come to play a leading role in the
future of MT.
There are also signs that involvement in MT will continue to
spread to include countries which have little experience in the field
so far. Recent developments in South Korea, for instance, indicate
that the private sector there is set to follow the example of its
Japanese counterpart with large-scale investment in MT R&D. It is
thus likely that a strong MT community will emerge in the Far East
which will set a challenge for the United States and Europe in an
 Machine translation today 29

area where the technological and economic stakes are likely to be

high, at least in the long term.
It also seems probable that the MT user community will grow in
the future. As the field has evolved, there has been increasing
emphasis on the development of viable MT-related products, and
an ever-widening market for these products has emerged. In an age
characterized by greater internationalization of commercial activity,
and with an expanding volume of documentation, companies large
and small are finding the need for efficient translation ever more
pressing. Human translators, whether working in-house or
freelance, have gradually come to abandon their initial trepidations
and, in general, no longer see their role threatened by MT. Indeed,
they are likely to be demanding an even greater variety of
automated translation aids. Small-scale MT systems or
computerized translation tools could become commonplace, as
Hutchins foresees:

It is reasonable to predict that in another twenty years MT and/

or MAT in various forms and packages will be normal and
accepted facilities in nearly every office and laboratory.
(Hutchins 1988:238)

Work on the basic research issues underlying MT will continue to

be given attention by academic researchers but it is to be hoped—
expected even—that there will eventually be a synergy of MT
efforts in the academic and industrial spheres, so that the
combined advantages of each can work to produce high-quality
products and systems which are efficient and which are based on
solid theoretical foundations. One would also hope that progress
will be made towards the development of a theory of machine
translation.
Similarly, we expect that the recent trend towards increasing
internationalization of MT efforts will continue with more and more
collaborative projects. The specific cooperative efforts of the United
States and Japan are a particularly exciting area and it will be
interesting to watch the progress in that direction.
Finally, there is little doubt that a major focus of future MT
activity will be the field of evaluation and standards. So far,
relatively little work has been done on the assessment of MT
systems but it has come to be recognized as an area vital to future
progress.
30 Jeanette Pugh

BIBLIOGRAPHY
Abbou, A. (ed.) (1989) Traduction Assistée par Ordinateur: Perspectives
technologiques, industrielles et économiques envisageables à l’horizon 1990: l’offre, la
demande, les marchés et les évolutions en cours, Actes du Séminaire international
(Paris, March 1988), Paris: DAICADIF.
ALPAC (1966) Language and Machines: Computers in Translation and Linguistics
(Report by the Automatic Language Processing Advisory Committee,
Division of Behavioral Sciences, National Research Council),
Washington, DC: National Academy of Sciences.
Bennett, W.S. and Slocum, J. (1985) ‘The LRC machine translation system’,
in Computational Linguistics 11:111–21; reprinted in J.Slocum (ed.) Machine
Translation Systems, Cambridge: Cambridge University Press, 1988, 111–
40.
Boitet, Ch. (1990) Towards personal MT: general design, dialogue
structure, potential role of speech’, in H.Karlgren (ed.) COLING-90:
Papers presented to the 13th International Conference on Computational Linguistics,
Helsinki: Yliopistopaino, vol. 3, 30–5.
Carbonell, J.G. (1990) ‘Machine translation technology: status and
recommendations’, in T.Valentine (ed.) (1990) ‘Status of machine
translation (MT) technology: Hearing before the Subcommittee on
Science, Research and Technology of the Committee on Science, Space
and Technology, US House of Representatives, 101st Congress, Second
Session, September 11, 1990’, [no. 153], Chairman: Rep. T.Valentine,
Washington: US Government Printing Office, 119–31.
Chandioux, J. (1989) ‘M ETEO: 100 million words later’, in ATA
conference proceedings, 449–53.
EEC (1990) ‘Council decision of 26 November 1990 adopting a specific
programme concerning the preparation of the development of an
operational E U ROTRA system’, Official Journal of the European
Communities, no. L 358/84, EEC/664/90, 21.12.90.
Farwell, D. and Wilks, Y. (1990) ‘Ultra: a multilingual machine translator’,
Research Report MCCS-90–202, Computing Research Laboratory, New
Mexico State University, Las Cruces, New Mexico.
Hammond, D.L. (ed.) (1989) Coming of Age and the proceedings of the Thirtieth
Annual Conference of the American Translators Association, October 11–15 1989,
Washington D.C., Medford, New Jersey: Learned Information.
Hutchins, W.J. (1986) Machine Translation: Past, Present, Future, Chichester:
Ellis Horwood.
——(1988) ‘Future perspectives in translation technologies’, in M.
Vasconcellos (ed.) Technology as Translation Strategy (American Translators
Association, Scholarly Monograph Series, vol. II, Binghamton, New
York: State University of New York Press, 223–40.
Iida, H. (1989) ‘Advanced dialogue translation techniques: plan-based,
memory-based and parallel approaches’, in ATR Symposium on Basic
Research for Telephone Interpretation, Kyoto, Japan, Proceedings 8/7–8/8.
Isabelle, P. (1989) ‘Bilan et perspectives de la traduction assistée par
ordinateur au Canada’, in A.Abbou (ed.) Traduction Assistée par Ordinateur:
Perspectives technologiques, industrielles et économiques envisageables à l’horizon
 Machine translation today 31

1990: l’off re, la demande, les marchés et les évolutions en cours, Actes du
séminaire international (Paris, March 1988), Paris: DAICADIF, 153–8.
JEIDA (1989) ‘A Japanese view of machine translation in light of the
considerations and recommendations reported by ALPAC, U SA’,
Tokyo: JEIDA.
King, M. (1989) ‘Perspectives, recherches, besoins, marchés et projets en
Suisse’, in A.Abbou (ed.) Traduction Assistée par Ordinateur: Perspectives
technologiques, industrielles et économiques envisageables à l’horizon 1990: l’offre, la
demande, les marchés et les évolutions en cours, Actes du séminaire international
(Paris, March 1988), Paris: DAICADIF, 177–9.
Landsbergen, J. (1987) ‘Isomorphic grammars and their use in the
ROSETTA translation system’, in M.King (ed.) Machine Translation Today:
the state of the art, Edinburgh: Edinburgh University Press, 351–72.
Maas, D. (1987) ‘The Saarbrücken automatic translation system (SUSY)’, in
Overcoming the language barrier, Commission of the European
Communities, Munich: Verlag Dokumentation, vol. 1:585–92.
Nagao, M., Tsujii, J. and Nakamura, J. (1985) ‘The Japanese government
project for machine translation’, Computational Linguistics 11:91–110.
Nirenburg, S. (1989) ‘Knowledge-based machine translation’, Machine
Translation, 4:5–24.
Pappegaaij, B.C., Sadler, V. and Witkam, A.P.M. (eds) (1986) Word Expert
Semantics: an Interlingual Knowledge-based Approach, (Distributed Language
Translation 1), Dordrecht: Foris.
Pogson, G. (1989) ‘The LT/Electric Word multilingual wordworker’s
resource guide’, LT/Electric Word 13, Amsterdam: Language Technology
BV.
Raw, A., van Eynde, F., ten Hacken, P., Hoekstra, H. and Vandecapelle, B.
(1989) ‘An introduction to the Eurotra machine translation system’,
Working Papers in Natural Language Processing 1, TAAL Technologie,
Utrecht and Katholieke Universiteit Leuven.
Slocum, J. (1984) ‘Machine translation: its history, current status, and future
prospects’, in 10th International Conference on Computational Linguistics,
Proceedings of COLING-84, Stanford, California, 546–61.
Steer, M.G. and Stentiford, F.W.M. (1989) ‘Speech language translation’, in
J.Peckham (ed.) Recent Developments and Applications of Natural Language
Processing, London: Kogan Page, 129–40.
Steiner, E. (ed.) (1991) ‘Special issue on Eurotra’, Machine Translation 6.2–3.
Thurmair, G. (1990) ‘Complex lexical transfer in METAL’, in Third
International Conference on Theoretical and Methodological Issues in Machine
Translation of Natural Languages, Austin, Texas, 91–107.
Trabulsi, S. (1989) ‘Le système SYSTRAN’, in A.Abbou (ed.) Traduction
Assistée par Ordinateur: Perspectives technologiques, industrielles et économiques
envisageables à l’horizon 1990: l’offre, la demande, les marchés et les évolutions en
cours, Actes du séminaire international (Paris, March 1988), Paris:
DAICADIF, 15–27.
Tsujii, J. (1989) ‘Machine translation with Japan’s neighboring countries’, in
M.Nagao (ed.) Machine Translation Summit, Tokyo: Ohmsha, 50–3.
Uchida, H. and Kakizaki, T. (1989) ‘Electronic dictionary project’, in M.
Nagao (ed.) Machine Translation Summit, Tokyo: Ohmsha, 83–7.
32 Jeanette Pugh

Valentine, T. (1990) ‘Status of machine translation (MT) technology:

Hearing before the Subcommittee on Science, Research and Technology
of the Committee on Science, Space and Technology, US House of
Representatives, 101st Congress, Second Session, September 11, 1990
[no. 153], Chairman: Rep. T.Valentine. Washington: US Government
Printing Office.
Vasconcellos, M. (ed.) (1988) Technology as Translation Strategy (American
Translators Association Scholarly Monog raph Series, vol. I I),
Binghamton, New York: State University of New York.
Whitelock, P.J., Wood, M.MacG., Chandler, B.J., Holden, N. and Horsfall,
H.J. (1986) ‘Strategies for interactive machine translation: the experience
and implications of the UMIST Japanese project’, in 11th International
Conference on Computational Linguistics, Proceedings of COLING-86, Bonn,
329–34.
Chapter 3

Made to measure solutions

Annette Grimaila in collaboration with
John Chandioux

Let us separate the machine from the translation and remember that
it is the machine that serves the translation and not the other way
round.
In all real-world applications of MT, the translator is not replaced.
In fact, he or she is the one person who must be consulted,
considered and helped by the application. If the machine output is of
such low quality or if its manipulation is so complex that the
translator wastes more time revising the results than he or she would
spend translating a source text, then the usefulness of the system is
seriously in doubt.
There is one world-renowned MT system which has been in
continuous use since the early 1980s: METEO and the Canadian
government’s use of it to translate public weather forecasts from
English to French (and from French to English since early 1989)
have been well publicized but its creator’s views on the subject have
seldom been sought out or clearly understood.

JOHN CHANDIOUX AND THE METEO SYSTEM

John Chandioux started out as an English-language specialist who
moved from France in 1973 to undertake further studies and
research with the Traduction Automatique de l’Université de
Montréal (TAUM) group. The late Professor Bernard Vauquois,
then director of the Groupe d’études pour la traduction automatique
(G ETA) of the University of Grenoble, France, had been
instrumental in setting up the TAUM group and over the years
continued to show an interest in its progress. TAUM’s most famous
project was a prototype of a mainframe machine-translation system
for the government of Canada which became known as TAUM-

33
34 Annette Grimaila with John Chandioux

METEO. What is less clear from the literature is that the prototype
developed in 1975–6 was never put into operation by the TAUM
group which thereafter concentrated its research on another project:
TAUM-AVIATION.
The Canadian government had practically shelved the
M ETEO project when Chandioux, who had by then left the
university, came to them with a proposal to resume development
of the prototype. What a shock when the first results showed that
only 40 percent of the sentences in the weather bulletins were
adequately translated! In fact, the prototype had been developed
on a non-representative corpus of texts which was too small and
did not include all of the weather regions for which the system
was destined. The Canadian weather is so diverse that it took a
full year of analysis, development and adjustments to ensure an
80 per cent success rate.
METEO-1 finally reached this goal in 1978. It ran on a Cyber-
7600 mainframe, required 1.6 megabytes of random access memory
and translated some 7,000 words per day.
Concurrently with his work on M ETEO, Chandioux had
undertaken the development of a programming language
specifically designed for linguistics. He wasn’t satisfied with what
could be done with traditional programming tools like PROLOG
or LISP, and even Q-SYSTEMS, with which METEO had been
developed, had had to be substantially rewritten in order to
eliminate a certain number of inconsistencies and unpredictable
bugs.
Chandioux had always been considered a visionary by the
computer experts at the University of Montreal. When he explained
what he wanted done, no one thought it was possible.
In 1976, Alain Colmerauer, a computer specialist who had
b een the first to implement P ROLO G as a programming
language, returned to Montreal to visit the TAUM group which
he had directed for some years and where he had written Q-
SYSTE MS. Discussions with Colmerauer, who considered the
project feasible, convinced Chandioux that the future of
computing and of his new language lay in the smaller machines
that would come to be known as microcomputers. In 1977,
Chandioux met with Professor Vauquois who, also agreeing with
Chandioux’s vision, took the time to detail what he would have
done differently with GETA’s machine translation projects were
he to start all over again.
 Made to measure solutions 35

As no one at the University was willing to get involved,

Chandioux set out to write the compiler and run-time engine
himself. By 1982, GramR was operating on a microcomputer and
being used in a prototype MT system to translate computer and
software manuals. This fourth-generation language is a deterministic
labelled-tree transducer, technically in the same family as Q-
SYSTEMS, but avoids combinatorial explosions thus making it an
effective development tool whereas Q-SYSTE MS, which is not
deterministic, is most useful as a research tool.
If GramR could translate computer jargon, why not weather
bulletins? After a year’s non-subsidized development and a three-
month comparative trial in 1983, METEO-2 beat METEO-1 on
performance, cost, reliability and user-friendliness. METEO-2 has
been in continuous use since 10 October 1984, rented by the
Canadian government’s translation bureau on a turnkey basis from
John Chandioux Consultants Inc. A software and hardware service
is provided twenty-four hours a day, seven days a week, with a
guarantee that full service is restored within four hours of a service
call. The principal component, like its fail-safe double, is a 68000
CPU microcomputer from Cromemco, under a Unix-like operating
system.
This workhorse is linked to a translator’s workstation as well as to
a Tandem communications node of the Canadian weather forecast
network. The Tandem directly feeds the weather bulletins to the
Cromemco from all of the eight weather centres: Gander, Halifax,
Montreal, Toronto, Winnipeg, Edmonton, Vancouver and the
Yukon. The Cromemco then sorts the English and French bulletins,
since those from the Province of Quebec are written in French, and
translates them into the target language. The translator’s
workstation then displays the bulletin’s code number to indicate that
it is ready for revision. This revision is mandated by law, due to the
nature of the information, with sea transport and the coastguard,
among other essential services, dependent on its accuracy. Control is
resumed by the Cromemco for transmission back to the Tandem at
9600 baud.
As at December 1990, less than 4 percent of the machine’s total
output required editing, turnaround time was less than 20 minutes
and the volume had risen from 7,000 words per day to 45,000. The
hardware is now operating near its physical limits and its successor,
METEO-3, is being developed to move the application to 80386-
based machines.
36 Annette Grimaila with John Chandioux

The secrets of success

What is METEO’s secret? How is it possible to reach and maintain
such a degree of accuracy in machine translation? There are three
principal reasons.
First, there is the nature of the problem itself. Machine translation
in its present state is far from capable of translating general texts.
Human language is much too ambiguous for a simple machine to
treat correctly and all attempts to date have been horrendously
expensive if not also totally laughable. Weather bulletins, even the
detailed ones prepared by meteorologists, are at least less ambiguous:
their subject is the weather, and only the weather. Some of the
remaining ambiguities can be circumvented by careful programming
but constant adjustments are required to keep up with non-standard
formulations in the source texts which are composed by human
meteorologists. Moreover, the useful life of weather bulletins
(approximately six hours) and their volume make them excellent
candidates for an automated system.
GramR, the language, is the second reason. Typical of second-
generation techniques which, as completely as possible, separate the
linguistics from the programming, this special tool is easily accessible
to a linguist or translator and allows him or her to compose a set of
transformational rules, compile these rules into an executable
program and test it on a chosen sample. Extremely reliable—its last
known bug was corrected in 1987—GramR lends itself to any
number of linguistic applications from MT to automatic
accentuation of French texts, to spelling and grammar checkers.
GramR’s appeal to linguists can be compared to the impact that
spreadsheet programs had on people who needed to perform
financial analysis in the early days of microcomputers.
The third, and not the least important, factor is the continuous
feedback and involvement of the translators who work with the
METEO system. They are constantly solicited for their input on the
system, for error detection in the output and for suggestions as to
possible improvements. Most of the staff assigned to the group have
been there for a number of years. Some even remember precomputer
days when weather bulletins came in over telex lines, were translated
on an ordinary typewriter and then handed back personally to a
communicator who retyped the bulletin in the outgoing telex
machine. Of course, the volume of text was much more limited in
those days. Today, 80 percent of their workload is machine
 Made to measure solutions 37

translated and requires them to assume their role of revisers with the
computer looking after the routines of receiving, translating and
transmitting. The remaining 20 percent of their workload takes up
half of their working hours even though terminals and other
microcomputers allow them quicker access to the bulletins and the
other tools which simplify their job.

Clearing up a misconception
It has often been said that the METEO system works because
weather bulletins use a simplified syntax, because the vocabulary is
limited or because the texts are so repetitive that there is no
challenge.
In reality, the syntax is not at all controlled at the input level: the
meteorologists are only required to respect certain semantic controls,
such as ‘strong winds’ being limited to velocities between ‘x’ and ‘y’
kilometers per hour. The recommended style resembles telegraphic
texts, with little if any punctuation or other syntactical points of
reference to help the translation process. The vocabulary, excluding
place names which are also translated, totals some 2,000 words.
Repetition? Yes, the texts can be repetitive but not predictably so.
The eight meteorological centres across Canada each reflect regional
differences and diverging styles as well as meteorological realities as
diverse as the country is vast.
The following is an extract of a weather bulletin submitted to the
METEO system on 31 January 1991:

METRO TORONTO.
TODAY…MAINLY CLOUDY AND COLD WITH
OCCASIONAL FLURRIES. BRISK WESTERLY WINDS
TO 50 KM/H. HIGH NEAR MINUS 7.
TONIGHT…VARIABLE CLOUDINESS. ISOLATED
FLURRIES. DIMINISHING WINDS. LOW NEAR MINUS
15.
FRIDAY…VARIABLE CLOUDINESS. HIGH NEAR
MINUS 6.
PROBABILITY OF PRECIPITATION IN PERCENT 60
TODAY. 30 TONIGHT. 20 FRIDAY.
WATERLOO-WELLINGTON-DUFFERIN
BARRIE-HURONIA
GREY-BRUCE
38 Annette Grimaila with John Chandioux

HURON-PERTH
…SNOWSQUALL WARNING IN EFFECT…
TODAY…SNOW AND LOCAL SNOWSQUALLS. BRISK
WESTERLY WINDS TO 50 KM/H CAUSING REDUCED
VISIBILITIES IN BLOWING AND DRIFTING SNOW.
ACCUMULATIONS OF 15 TO 25 CM EXCEPT LOCALLY
UP TO 35 CM IN HEAVIER SQUALLS. HIGH NEAR
MINUS 9.
TONIGHT…FLURRIES AND LOCAL SNOWSQUALLS
CONTINUING. NORTHWEST WINDS TO 40 KM/H.
LOW NEAR MINUS 16.
FRIDAY…FLURRIES AND SQUALLS TAPERING TO
SCATTERED FLURRIES. HIGH NEAR MINUS 7.
PROBABILITY OF PRECIPITATION IN PERCENT 90
TODAY. 90 TONIGHT. 70 FRIDAY.
METEO’s machine output, without any human revision, reads as
follows:
LE GRAND TORONTO.
AUJOURD HUI…GENERALEMENT NUAGEUX ET
FROID AVEC QUELQUES AVERSES DE NEIGE. VENTS
VIFS D’OUEST A 50 KM/H. MAXIMUM D ENVIRON
MOINS 7.
CETTE NUIT…CIEL VARIABLE. AVERSES DE NEIGE
EPARSES. AFFAIBLISSEMENT DES VENTS. MINIMUM
D ENVIRON MOINS 15.
VENDREDI…CIEL VARIABLE. MAXIMUM D
ENVIRON MOINS 6.
PROBABILITE DE PRECIPITATIONS EN
POURCENTAGE 60 AUJOURD HUI. 30 CETTE NUIT. 20
VENDREDI.
WATERLOO-WELLINGTON-DUFFERIN
BARRIE-HURONIE
GREY-BRUCE
HURON-PERTH
…AVERTISSEMENT DE BOURRASQUES DE NEIGE
EN VIGUEUR…
AUJOURD HUI…NEIGE ET BOURRASQUES DE
NEIGE PAR ENDROITS. VENTS VIFS D OUEST A 50
KM/H OCCASIONNANT UNE VISIBILITE REDUITE
DANS LA POUDRERIE HAUTE ET BASSE.
 Made to measure solutions 39

ACCUMULATIONS DE 15 A 25 CENTIMETRES MAIS

JUSQU A 35 CENTIMETRES PAR ENDROITS SOUS LES
BOURRASQUES PLUS FORTES. MAXIMUM D
ENVIRON MOINS 9.
CETTE NUIT…AVERSES DE NEIGE ET
BOURRASQUES DE NEIGE PAR ENDROITS. VENTS DU
NORD-OUEST A 40 KM/H. MINIMUM D ENVIRON
MOINS 16.
VENDREDI…AVERSES DE NEIGE ET BOURRASQUES
SE CHANGEANT PROGRESSIVEMENT EN AVERSES
DE
NEIGE EPARSES. MAXIMUM D ENVIRON MOINS 7.
PROBABILITE DE PRECIPITATIONS EN
POURCENTAGE 90 AUJOURD HUI. 90 CETTE NUIT. 70
VENDREDI.

GRAMR’S OTHER CHI LDREN

Has anything else been developed with GramR?
There are, at present, three other GramR MT systems, two in
operation and one still in development.

General TAO—the insurance industry

The first cannot really be called a machine-translation system as it
performs no analysis of the incoming text and basically does a
special kind of pattern matching.
Confederation Life Insurance Company’s Montreal office is the
location of the company’s linguistic services department. On
deciding to automate his department, Jean-Pierre Bernier carefully
analyzed the needs and working habits of his translators. This
resulted in his equipping his people with powerful Macintosh
workstations, a made-to-measure project-management program,
Microsoft Word as a text processor, a desk-top publishing program,
a terminology bank and a local area network under TOPS software.
A single MS-DOS machine was acquired to provide compatibility
with other company departments and reliance placed on Apple file
exchange to convert files.
Even with this environment to facilitate their work, however, the
translators were spending considerable time putting together
standard documents from an existing bank of pre-translated
40 Annette Grimaila with John Chandioux

paragraphs. Discussions between Jean-Pierre Bernier and John

Chandioux Consultants Inc. brought about the development of the
Système General de Traduction Assistée par Ordinateur (General
TAO).
The principal objective of General TAO was to reduce or
eliminate the routine cut-and-paste work required to translate these
standard documents while respecting the very rigid formatting
requirements of the lengthy texts (group insurance policies,
descriptive employee booklets, benefit administration agreements,
etc.). To complicate matters, the English texts are written on PCs
using WordPerfect 5.0/5.1.
The challenge in this development was twofold: finding a system
that would ‘speak’ to both PCs and Macintoshes, as well as a means
of dealing with the enormous volume of text involved.
The solution was to use the PC compatible as a server for the
system. Since the department’s computers were already linked into a
local-area network, it seemed much more efficient to practically
dedicate the little-used PC to the General and use TOPS to publish
the results on the network.
To deal with the volume and ensure that the required presentation
is respected, the input is first stripped of most of its formatting
information and an almost pure ASCII text is submitted to the
GramR kernel which paginates through both the text and the
‘overlays’ containing the matched patterns. These pseudo-code files
have reached a remarkable size: close to a megabyte for one single
type of document. The pattern matching can best be described as
‘fuzzy’ with a series of variables to deal with dates, periods, amounts,
etc. The overlays also contain mnemonic formatting codes which are
converted to Microsoft’s Rich Text Format on output, thus making
the results easily interpretable by the Macintosh version of Microsoft
Word.
The resulting document is perfectly formatted and, beause the
French portions have all been standardized, thus requiring absolutely
no revision, the translators need only ensure that any non-standard
clauses which remain in English and which are flagged with a special
character are correctly translated.
The following is an extract from a standard Confederation Life
insurance policy and its translation by the General TAO system:
 Made to measure solutions 41

DEFINITIONS
Earnings: gross monthly earnings, excluding bonus, commissions and
overtime. Earnings shall be determined where necessary on the basis
of 40 hours per week, 4.33 weeks per month and 12 months per year.
Renewal date: January 1st.

GENERAL PROVISIONS
If an employee suffers a specified loss as a direct result of a covered
accident, within 365 days after the date of such accident, the benefit
will be paid provided CONFED receives proof of claim.
The amount payable is based upon the amount specified in the
BENEFIT PLAN SUMMARY which is in effect at the time of loss and
calculated using the percentage for the loss set out in the following table:

Loss of life ------------------------------------------------------------------------------------- 100%

Loss of both hands ----------------------------------------------------------------------- 100%
Loss of both feet --------------------------------------------------------------------------- 100%
Loss of sight of both eyes ------------------------------------------------------------ 100%
Loss of one arm ------------------------------------------------------------------------------ 75%
Loss of one hand --------------------------------------------------------------------- 66 2/3%

ADMINISTRATIVE PROVISIONS
POLICY RENEWAL
At the end of each policy year, CONFED will renew this policy for
a further term of one policy year provided:

A the participation requirement of 100% of those employees

eligible for coverage is met;
B the minimum premium requirement of $100 for each month of
the policy year is met;
C the policyholder accepts any policy changes which CONFED
considers necessary for renewal purposes; and
D CONFED receives the first premium due for the new policy year.

DÉFINITIONS
Salaire: rémunération mensuelle brute, à l’exclusion des grati-
fications, commissions et heures supplémentaires. Le salaire est
calculé à raison de 40 heures par semaine, 4,33 semaines par mois et
de 12 mois par année.
*Renouvellement: le ler janvier, à minuit une minute.
42 Annette Grimaila with John Chandioux

CONDITIONS GÉNÉRALES
En cas de sinistre directement attribuable à un accident garanti,
survenu dans les 365 jours qui suivent l’accident, CONFED verse la
prestation après avoir reçu les pièces justificatives. La prestation
correspond à un pourcentage du capital stipulé *aux
CON DITION S PARTICU LIÈRE S au moment du sinistre,
conformément au tableau ci-dessous:

Décès .............................................................................................. 100%

Perte des deux mains ...................................................... ......... 100%
Perte des deux pieds ..................................................................... 100%
Perte de la vision des deux yeux ................................................ 100%
Perte d’un bras ................................................................................ 75%
Perte d’une main ...................................................................... 66 2/3%
GESTION ADMINISTRATIVE
RENOUVELLEMENT DU CONTRAT
A la fin de chaque année d’assurance, CONFED renouvelle le
contrat sous réserve des conditions suivantes:
A participation de tous les salariés admissibles;
B paiement d’une prime minimale de 100$ par mois;
C acceptation par le titulaire de toute modification jugée
nécessaire;
D paiement de la première prime de l’année suivante.

Here again the success of the development was assured by the assistance
of the department’s assistant manager, Paul Dupont, and the translators
who were directly involved in testing from the beginning and in
completing the grammars since October 1990. In December 1990, the
translator who is now in charge of the system, David Harris, confirmed
that he was able to produce a finished document from a General TAO
output in 30 minutes as compared to several hours before the system
was operational. Additional documents are being analyzed for possible
inclusion in the system and a French-English counterpart using the same
user interface is being considered.

DIGITRAD—the computer industry

In the case of Digital Equipment of Canada Limited, the translation
problem entailed dealing with an inventory control system which
contained some 165,000 entries, in abbreviated English to respect a
 Made to measure solutions 43

30-character field-length, and providing an equally short French

version for each of the hardware and software items contained
therein.
Francine Létourneau, translation manager in Montreal,
approached John Chandioux for a first-phase solution which would
deal with the software descriptions which represent some 70 percent
of the volume.
The challenge in this case was to respect the 30-character limit
and to decipher the English abbreviations which were produced by
inventory control officers located in the United States. There were
80,000 entries requiring translation at the start of the project with
constant updates coming in from the USA.
DIGITRAD runs in either batch or test mode. In the latter, the
translator enters the English text on a PC compatible displaying a
user-friendly interface quite similar to the General TAO system. The
description is translated into clear French to ensure accuracy and, by
successive calls to an abbreviation module via the return key, is
reduced to the required length. In batch mode, the process is
identical but without the translator’s intervention.
DIGITRAD is a true machine translation system. It analyzes an
input text written in a very specific sublanguage with its own
vocabulary and grammar. The different elements of the descriptions
can require translation, permutation and abbreviation.
Before DIGITRAD, the translator would access the database
and, using the update function, call up one description after
another and re-enter the full entry, translating where required.
Three or four translators could work on the database without
access to their colleagues’ files, thereby duplicating some
terminology work and creating different standards for the same
types of descriptions. One translator could translate some 2,000
words a day. In early January 1991, Jean-Claude Bergeron, who
edits the machine output, confirmed he had been able to translate
55,000 words in three days.
The following are sample entries from Digital’s database:

QA-0JQAA-WZ UWS USER/ADMIN V4.1 UPD DOC K

QA-GEJAA-H5 20/20 RISC MED & DOC
QL-B16A9-JT WPS-PLUS VMS ED CW:6000
QL-GHVA9-JN DECmsgQ VMS RTO CW:2400
QT-0JXAG-9M WIN/TCP DPMC 16MT9
QT-YNCAM-8M DPRT PRT SVCS BSMC 16MT9
44 Annette Grimaila with John Chandioux

Translated:
QA-0JQAA-WZ UWS USER/ADMIN V4.1 M/J
ENSEMBLE DOC
QA-GEJAA-H5 20/20 RISC SUPP ENREG DOC
QL-B16A9-JT WPS-PLUS VMS LIC E ENS CW:6000
QL-GHVA9-JN DECmsgQ VMS OPTION EXEC CW:2400
QT-0JXAG-9M WIN/TCP F/M DECSUPPORT 16MT9
QT-YNCAM-8M DPRT PRT SVCS F/M SER BASE 16MT9

Abbreviated:
QA-0JQAA-WZ UWS USER/ADMIN V4.1 M/J ENS DOC
QA-GEJAA-H5 20/20 RISC SUPP DOC
QL-B16A9-JT WPS-PLUS VMS LIC E ENS CW:6000
QL-GHVA9-JN DECmsgQ VMS OPT EXEC CW:2400
QT-0JXAG-9M WIN/TCP F/M DECSUPP 16MT9
QT-YNCAM-8M DPRT PRT SVCS F/M SER BASE 16MT9

Canadian National Railways

Canadian National’s system, which is still in development and has
not yet been baptized, resembles Digital Equipment’s in that it deals
with an inventory control system. The items, however, are not as
abbreviated and can contain up to six lines of text to describe the
parts and supplies required to run the railroad.
In this case, the challenge is in dealing with some 70,000 descriptions,
approximately half of which have already been translated. John
Chandioux is in the process of writing a GramR program to match the
English and French automatically and generate the translation rules by
computer. The system should be in operation by mid-1992.

A FEW FINAL WORDS

Will an MT system solve most translation problems? No. But it can
solve those problems which meet the following conditions:
specificity, volume and repetitiveness.
Will a special development be required? Probably, as no one at
present knows the exact proportions of all the possible language
problems and a made-to-measure solution has a greater chance of
consistently ensuring productivity gains in the 700 per cent to 800
per cent range.
In our view, a general translation system cannot work in the
 Made to measure solutions 45

present state of the art. We are too far from fully understanding the
nature of human language and its myriad linguistic representations.
A workable general translation system will eventually grow from
numerous specific translation projects, such as the ones we have
developed at John Chandioux Consultants. Through our research
and development activities we can slowly identify and solve
linguistic ambiguities and ensure a quality output that substantially
helps the translator to do what he or she is paid to do: translate
realities from one reference system to another.
Like any good tool, GramR has been found useful in related fields
for which it was not originally designed. Even though GramR was
originally developed to deal with English–French translation
problems, it is not specific to this language duo and can be used for
most European languages. It can even be used for other linguistic
applications.
John Chandioux Consultants Inc. has developed and launched
several general-use software products, all of which deal with the
French language: a spelling and grammar checker, a verb conjugator
and, most recently, a program that converts standard French spelling
to the ‘reformed’ spelling proposed by the Conseil Supérieur de la
Langue Frangaise in Paris on 6 December 1990. These are known as
the GramR family of products: ORTOGRAF+, the name used in
Canada for the spelling and grammar checker, CONJUGUONS!
and L’ORTHOGRAPHE MODERNE.
The development of a similar class of products for the English
language is under way and should prove interesting to people
‘bilingually’ involved.

NOTE
GramR, METEO and ORTOGRAF+ are registered trade marks
and are the property of John Chandioux.
Chapter 4

The Perkins experience

John Newton

BACKGROUND
One of the most successful machine translation applications known
to me is at Perkins Engines, Peterborough, England. It is an example
of what can be achieved when a system is introduced in a thoroughly
planned and methodical way into a restricted domain environment
to process controlled-language source texts.
Perkins has been a leading manufacturer of diesel engines since
1932 and is well established in worldwide export markets. Frequent
additions to the product range and modifications to existing
products have created a need for rapid production of high-quality
user documentation in five languages—English, French, German,
Spanish and Italian.
Until 1985, all translation had been done manually: some overseas
and some in the UK. The Technical Publications Manager, Peter Pym,
was keen to ensure that the four translated versions did not differ too
greatly in style or content from the English source texts or from each
other; close scrutiny of existing translations had revealed minor
semantic and stylistic divergences, as well as omissions and introduced
elements (i.e. elements not derived from the source text). As these
traits were particularly evident in translations produced or edited
overseas, greater control from Peterborough was clearly desirable but
the translations produced had to be acceptable to the overseas
subsidiaries, given that they and their customers were the consumers.
When Peter Pym decided to explore the possibility of using MT, he
already had a firm foundation on which to build: his department was
using a form of controlled English known as Perkins Approved Clear
English (PACE). PACE was initially based on the International
Language for Service and Maintenance (ILSAM), which in turn was

46
 The Perkins experience 47

based on Caterpillar Fundamental English (CFE). CFE comprised

around 800 words of basic English, plus whatever technical terms
were required to describe products. In 1990, the number of words in
PACE stood at approximately 2,500, of which around 10 per cent
were verbs. PACE is based on sound, commonsense principles: short
sentences, avoidance of gratuitous synonymy (e.g. right is the opposite
of left; its use in the sense of correct is therefore proscribed), avoidance
of ellipsis, and great emphasis on clarity of expression. Founded on the
principle ‘one word, one meaning’, the PACE dictionary lists and
defines or exemplifies every word that is approved for use in technical
publications, including articles, conjunctions, pronouns and
prepositions. In the case of homographs, it specifies the parts of speech
that can be used, e.g. seal is listed as both verb and noun, while stroke
is listed only as a noun. The technical authors also apply a set of rules
governing syntax and sentence patterns. This approach to writing
grew out of a desire to convey technical information and instructions
in as precise, clear and unambiguous a form as possible in the interests
of safety and efficiency.
The following examples provide an illustration of these rules in
practice:

Pre-PACE
The heavy duty oil bath cleaners are usually fitted with a
centrifugal pre-cleaner mounted on top of the main cleaner.
Using PACE
Heavy-duty air cleaners of the oil bath type are usually fitted with
a central pre-cleaner, which is mounted on top of the main
cleaner.
Pre-PACE
There are a few engines fitted with spur gears instead of helical
gears shown in this section.
Using PACE
Certain engines are fitted with spur gears, instead of helical gears
which are shown in this section.
(Pym 1990:86–7)

The texts produced in the technical publications department are

models of stylistic homogeneity and terminological consistency. As
was stated in the introduction to this volume, non-standard syntax
and other deviations from linguistic norms are among the most
48 John Newton

serious obstacles inhibiting the widespread use of MT. Perkins’ strict

adherence to the rules and principles of PACE produces texts that
are neutral in style and devoid of authors’ quirks.
Peter Pym was aware that his department’s controlled approach to
technical writing could facilitate the introduction of MT, and in
March 1984 he and his colleagues established their criteria for an
‘ideal’ system. After examining the (very few) systems that were
available, they concluded that Weidner’s MicroCat system matched
their requirements most closely and a decision was taken to organize
an operational trial using English-French. MicroCat is a PC-based
system which processes translations in batch mode, using the
transfer method, as described in Chapter 6.

DICTIONARY BUILDING AND UPDATING

The decision to conduct a test necessitated the creation of an
English/French version of the PACE dictionary within the MicroCat
system. Every word in PACE, including articles and prepositions,
had to be entered into the new dictionary as the total control of the
lexis that Peter Pym was seeking precluded the use of any other
dictionary. Moreover, limiting the parts of speech in homographic
entries (see below) to those that were allowed by PACE optimized
the system’s performance by reducing the number of available
variants. Entries for non-technical words were, in most cases, copied
from the system’s core dictionary and modified as necessary. The
core dictionary itself cannot be modified by the user but up to
twenty-nine additional dictionaries can be created and sequenced as
required for any particular task.
The existence of monolingual and bilingual dictionaries
containing all the words used in Perkins’ technical texts, along with
definitions, usage examples and part-of-speech designations, greatly
simplified the task of dictionary building. In an environment where
vocabulary and terminology were not controlled and finite it would
be necessary to use the system’s vocabulary search function to
identify lexical items that would need to be entered into the system’s
dictionaries.
Raw translation quality would be severely impaired in two ways
if a text submitted for translation contained a substantial number
of words that did not figure in the dictionaries specified for that
task; first, the fact that no translation equivalents would be
available for the ‘unfound’ elements would result in their being
 The Perkins experience 49

marked and left untranslated; second, and more important, the

absence of ‘flagged’ information concerning the word type and
other attributes of these elements would so limit the system’s
ability to perform its sentence-level syntactic analysis that it would
invariably lead to a seriously confused rendering of any sentence
thus affected.
During dictionary entry, the system elicits inflection rules and
a considerable amount of semantic and syntactic information
from the user. Correct entry of the inflection rules produces
automatic generation of inflected forms of target-language verbs,
nouns and adjectives, and recognition of English verbal adjectives
ending in -ed and -ing and adverbs ending in -ly. The system also
generates the appropriate forms of articles, forming contractions
as necessary.
Entering inflected forms could interfere with the translation of
identical forms functioning as other parts of speech; all source
language words are therefore entered into the dictionary in their
uninflected base form: English verbs in the infinitive and nouns in
the singular. In the same way, target words that can inflect in the
idiom must also be entered in their base form: verbs in the
infinitive, nouns, adjectives and articles in their masculine singular
form. For example, lubricating oil would need to be entered as
lubricate oil to avoid interferance with the various -ing forms
deriving from the entry for lubricate; accurate flagging of both
entries would ensure that lubricating oil was correctly recognized and
translated.
As dictionary updating plays such a key role in determining the
quality of MicroCat’s raw output, I felt that some readers would
appreciate a rather more detailed description of the procedures
involved; anyone less interested in this aspect of the subject should
skip the remainder of this section and go directly to ‘Adapting the
working methods to the system’ (page 53).

Multiple-word entry
The next step was to use the PACE English/French MicroCat
dictionary to produce raw translations of texts that had been
translated previously and then compare the machine’s output with
the human versions. This helped to identify recurrent structures
requiring multiple-word-entry solutions in the MicroCat dictionary;
for example, big end bearing, always ensure that.
50 John Newton

In addition to catering for single-word source entries (maximum

eighty characters) and homographs, MicroCat’s dictionary
structure permits the user to enter compound nouns and other
multiple-word forms which are referred to as ‘idioms’. In this
context, ‘idiom’ is used to describe any dictionary source item
which comprises more than one word; drive belt would therefore be
regarded as an idiom. Source idioms can contain up to eight words
(maximum 256 characters), and target idioms up to fourteen. In
addition, ‘idiom holes’ allow idioms to contain variable elements
which are translated independently of the idiom as a whole. Idiom
holes are of two kinds: ‘must match holes’ mark positions in idioms
that must be occupied by a word in order for the idiom to be
‘found’; ‘conditional match holes’ mark positions that can either be
occupied or left empty. This device allows a single entry to handle
an infinite number of permutations; for instance, be % % away,
given the translation être à % % (French), or estar a % % (Spanish),
can handle any combination which follows this pattern, such as the
church is 2 miles away from the school, the keyboards were 10 centimetres
away from the screens, the ladders will be several feet away from the wall.
This idiom has two ‘must match holes’, each represented by %.
The first hole is flagged as a quantifier and the second as a noun;
this means that any number or other quantifier will be accepted in
the first hole, and any noun (singular or plural) in the second. The
system would only apply the idiom entry if a combination of words
matched this profile exactly; a combination such as he is very far
away would therefore not result in the idiom being used. The
‘idiom holes’ feature allows the user to specify a single word type,
or a number of word types, for each hole. If a single word type is
chosen, various attributes can be selected to pinpoint qualifying
words still further, e.g. gender, number, person, tense, mood, etc.
The same entry procedure applies to ‘conditional match holes’,
represented by $.
In the case of the text used for the Perkins English/French trial (a
user’s handbook), a large number of multiple-word entries were
made after scrutiny of the machine output and the text was then
retranslated using the updated dictionary. A trial and error process
was thus begun in which consecutive dictionary updates and
retranslations finally resulted in greatly improved raw output. ‘Idiom
hole’ entries were used to provide solutions for frequently occurring
constructions which were consistent with regard to pattern (i.e. the
parts of speech of their component words and their relationships to
 The Perkins experience 51

each other) but variable as far as individual words in the ‘hole’

positions were concerned. It is easier and quicker to tackle less
frequent occurrences at the post-editing stage, and it is in any case
unwise to clutter the dictionary with a larger than necessary number
of complex entries as this can increase processing time and confuse
the system.

Homographic entries
For the purposes of the MicroCat system, a homograph is defined as
a word that can function as more than one part of speech, in its base
form, in an inflected form or in some combination of both. Because
English verbs invariably have inflected forms that function as non-
verbal parts of speech, e.g. adjectives ending in -ing and -ed, all verbs
are treated as homographs. Provision is made for various
homographic combinations, including: noun/adjective, noun/
adjective/adverb, adjective/adverb, verb/adverb, verb/noun/ adverb,
verb/noun/adjective and verb/noun/adjective/adverb, as well as verb
homographs which allow for the existence of a gerund or other -ing
noun, e.g. swim (swimming), build (building). Although the system
provides for up to nine parts of speech for each homographic entry,
it is unusual for more than five to be entered. The French
translations entered for the source homograph light in a general
context would probably be: lumière (noun), éclairage (-ing noun),
allumer (verb), léger (adjective), and qui allumer (the so-called -ing
adjective); the latter form is entered to cover cases where the -ing form
is used adjectivally, e.g. the boy lighting the candles, the man walking his
dog; however, walking stick would need a separate entry to avoid its
being translated as la canne qui marche.
The system’s ability to differentiate between the various parts of a
homograph, as exemplified by its rendering of the nonsense sentence
the light lady lighting the lighting is lighting light lights as la dame légère qui
allume l’éclairage allume des lumières légères, is impressive.

Using idioms to preserve target language distinctions

The impressive capacity for resolving homographs described
above does not overcome the problem of a source word or term
which requires different translations in different sub-contexts
within a given domain for a single part of speech; for example,
the Perkins PACE English/ French dictionary had no fewer than
52 John Newton

nine noun translations for gear. Multiple translation equivalents

deriving from target language usage distinctions demand special
attention, as a source language lexical item can be given only one
translation per part of speech. Scrutiny of similar texts translated
previously will usually reveal which equivalent occurs most
frequently and this should b e used for the main entry.
Occurrences of the same word requiring different translations can
often be accommodated through entering compound noun
idioms. For example, the main entry for gear could be given the
translation engrenage, while crankshaft gear could be entered as
pignon de vilebrequin, and worm gear as vis sans fin. Any cases not
resolved in this way would be addressed at the post-editing stage
using extended Alt Mode commands to effect the appropriate
substitutions.

Dictionary sequencing
In a situation where the prescribed vocabulary and terminology
were not contained within a single fully customized dictionary, a
hierarchical dictionary look-up sequence would be specified to
guide the system in its choice of translation equivalents. Up to
nine dictionaries can be sequenced by the user for any vocabulary
search or translation task, and up to twenty-six different
sequences can be resident at any time. In this mode, the system
searches for source-text words and idioms in each dictionary in
the specified sequence until it finds or finally fails to find the item
in question. In an automotive context, a look-up sequence might
contain a product-specific dictionary at the highest level, followed
by a general in-house dictionary, a general automotive dictionary
and the system’s core dictionary. In this sequence, terms coined
to describe features specific to the product would be found in the
first dictionary specified, terms relating to the company’s
products in general would be located in the second, general
automotive terms such as steering wheel and small end bush would be
found in the third and the fourth would provide general words
such as number and blue. Jumbling the sequence can produce
amusing but predictable results; e.g. sequencing the core
dictionary before the general automotive dictionary would result
in small end bush being rendered as le petit buisson de fin but this
would b e attributable to human error (or mischief ). The
MicroCat PACE dictionary’s total, yet exclusive, coverage of the
 The Perkins experience 53

requisite vocabulary and terminology made sequencing

unnecessary at Perkins.

ADAPTING THE WORKING METHODS TO THE SYSTEM

When it was considered that little further improvement could be
achieved through manipulation of the dictionary entries, it was
decided to look at the English source text to see whether any
general changes could be made which would improve the
quality of raw translation. The starting point for this exercise
was a markedup copy of the raw output obtained using the
repeatedly updated English/ French MicroCat PACE dictionary.
Sentences which had caused problems in translation were
rewritten in ways which produced better renderings from the
system but there was no compromise with regard to naturalness;
there would clearly have been no point in distorting the content
or ‘feel’ of the source text in order to accommodate the system’s
limitations. This activity, regarded as an extension of the PACE
principle, did not in any way result in a sterile or ‘unfriendly’
style but it did produce a more than threefold increase in the
number of verbs in the PACE dictionary: the total of eighty
verbs listed at the beginning of the exercise rose to around 250.
The technique used here was fundamentally different from pre-
editing as discussed elsewhere in this volume because the text
submitted for translation was the published source language
version. Moreover, the stylistic practices that evolved were
deemed to be generally applicable in the Perkins Technical
Publications context. Where pre-editing is practised, it is aimed
at generally improving a system’s rendering of an individual
source text, but the original, unedited version remains the
‘official’ source language document.
The amended source text was submitted for machine processing,
and this time the output was of an extremely high standard, requiring
very little post-editing; however, it must be said that some problems
were intractable and could only be solved through post-editing.
Although Perkins’ policy of writing in controlled language optimizes
the system’s performance, MicroCat’s analysis is limited to the
sentence and this means that links and agreements (anaphora) cannot
be carried beyond the sentence level; nevertheless, the quality overall
was such that Peter Pym placed an order for software to translate the
four language pairs that his department regularly handles.
54 John Newton

Implementation

Because the groundwork had been done with English/French, this

language pair was implemented first. However, as the text used for
the trial was unlikely to have contained every approved word and
idiom, a more thorough test of the customized dictionary was
needed. Several previously translated service publications were
therefore submitted to MicroCat for translation in order to expose
the system to a broader range of typical text and to highlight
dictionary entries still requiring attention. A series of retranslations
and dictionary updates then followed until the dictionary was
considered to be fully ‘tuned’ to the Perkins environment. When a
multiple-word entry was needed to overcome a particular problem,
model target-language idioms were gleaned from the existing
translations. It should be emphasized that this exercise was confined
to structures, as the vocabulary and terminology were already in the
PACE dictionary.
When the first text written under the modified PACE rules was
submitted for translation by MicroCat in 1986, the result was, if
anything, even better than expected. A copy of the raw translation,
processed overnight in batch mode, was sent to Perkins France for
editing and the changes recommended were then implemented in
Peterborough. After final proofreading by Perkins France, the text
was approved for publication. Painstaking preparation had created
conditions in which very short turnaround times were achievable
with minimal post-editing.
Since this first pioneering effort, every new service publication
produced by Perkins has been translated using MicroCat and,
whenever possible, source texts have been created in the system’s
own word processor. Other language pairs have been brought on
stream and UK-based post-editors have been commissioned to
‘polish’ the raw translations and so reduce turnaround time still
further. Additional time savings are being achieved in production
through source texts being fully coded for typesetting; this coding is
automatically reproduced in the raw translations and the post-edited
versions are then supplied to the printer on diskette.
The documentation’s usefulness to end-users and to those
responsible for marketing and servicing the products is optimized
through the Perkins company in each of the main target markets
remaining the authority for final approval and for all queries relating
to terminology and usage.
 The Perkins experience 55

Not surprisingly, it transpired that writing to accommodate the

system’s limitations resulted in even clearer communication and
reinforced the advantages of PACE to non-native speakers who are
obliged to use the English version if their mother tongue is not one
of the five in which the documents are published (it is undoubtedly
clearer for native speakers too).
Writing in controlled language produces texts which are more
easily translatable, whatever method is used; it also results in greater
homogeneity between different translated versions (i.e. versions in
different target languages and versions produced by different
translators into a single target language). As one might expect,
different translations of a source text written in controlled and
simplified language normally resemble each other more closely than
versions derived from texts having a less regulated structure;
comparing source texts written under varying degrees of constraint
with back-translations usually illustrates this point.
MicroCat’s vocabulary search function proved to be a useful tool
for monitoring adherence to PACE rules in newly written texts. As I
have stated above, if the PACE rules are strictly adhered to, a
vocabulary search should reveal no ‘unfound’ words; however, any
non-PACE words identified are examined in context to see if a new
PACE entry is needed, and any selected for inclusion (possibly
arising from the introduction of a new product or process) are first
defined and then given part-of-speech designations. The next task is
to identify and approve target-language equivalents in consultation
with the overseas subsidiaries and, when these procedures are
complete, entries are made in the PACE and MicroCat PACE
dictionaries. Hard copies are then distributed to authors and post-
editors. In addition, post-editors are encouraged to point out
recurrent structural problems in the hope that they might be solved
via the dictionary. This methodical, but openminded, approach
ensures that everyone concerned works within the same guidelines
and that any new or deviant vocabulary or terminology is identified
for approval, modification or rejection.

CONCLUSIONS
Not least of the benefits of Perkins’ bold decision to automate the
translation process was the opportunity it afforded for looking even
more closely at what was being written and how it was being
written. Introducing MT afforded Peter Pym a level of control that
56 John Newton

was previously unattainable and resulted in greater uniformity of

content between source texts and their various translated versions:

Using MicroCat, Perkins has been able to ensure consistent

terminology and to reduce translation time as well as translation
costs. Using the computerized databases, Perkins can control the
source and target text at all stages of publishing. Producing
translation using an MT system also ensures rigorous testing and
control of the source text.
(Pym 1990:92)

The pre-existing systematic approach to writing, based on continous

reappraisal, created ideal conditions for this project, as did the
enthusiastic co-operation of the technical authors and the personnel
in the overseas subsidiaries. The department’s relations with the
latter had already been strengthened through cooperation in
compiling the bilingual versions of the PACE dictionary. Likewise,
throughout the period of the MicroCat test (around six months),
Peter Pym had kept his colleagues in France fully briefed on
developments and had sought and acted upon their advice whenever
any queries had arisen concerning terminology or usage.
In addition, I had demonstrated the system at Perkins’ French
subsidiary before the trial began and personnel from Perkins
Engines France had attended meetings in Peterborough to help
decide how it should be conducted and what form their cooperation
would take; consulting all the parties involved in the project from the
outset ensured their support. Translators from Perkins’ parent
company, Massey-Ferguson, were also involved throughout, as was
Tony Hartley, a linguist from Bradford University, who advised
Perkins regarding modifications to PACE and who devised a set of
ten rules for simplified writing specifically designed to produce
optimum results from MicroCat in the Perkins context.
At first sight, it may appear rash to have installed MT in a
department which does not have any translators on its staff, but it
must be borne in mind that linguists and post-editors were, and still
are, consulted as external resources and Peter Pym has found this
arrangement to be efficient, flexible and cost-effective. Perkins’
disciplined approach to the MicroCat trial, which had demanded
and received all the preparation normally associated with full
implementation, ensured that the actual implementation of English
to French went very smoothly, as the system was delivered with the
 The Perkins experience 57

dictionary substantially tailored to the Perkins environment.

Another factor which maximized the efficiency of the Perkins
installation overall was the decision to introduce new language pairs
only when those already in use were fully operational; this enabled
the subsequent implementations to benefit from the lessons learned
from those which had preceded.
More than anything else, however, the success of the Perkins
application was made possible by the controlled and extremely
consistent nature of the source texts and by a willingness on the part
of all concerned to adapt the system to the working methods and the
working methods to the system.

REFERENCE
Pym, P.J. (1990) ‘Pre-editing and the use of simplified writing for MT: an
engineer’s experience of operating an MT system’, in P.Mayorcas (ed.)
Translating and the Computer 10: The Translation Environment 10 Years On,
London: Aslib, 80–96.
Chapter 5

Machine translation in a
high-volume translation
environment
Muriel Vasconcellos and Dale A.Bostad

The real test of machine translation is whether or not it is effective in

large-scale operations. These may be specialized applications, as in
the case of Canada’s METEO (see Chapter 3), or they may involve
the translation of a broad variety of text types. In the latter case, the
purpose of the translation will dictate the characteristics of the
installation, particularly the human post-editing component. The
purpose can run the gamut from publication for dissemination to
‘information only’. For the product that has a public and undergoes
scrutiny, a translator post-editor must be enlisted to eliminate all
problems from the output, interpret connotations appropriately and
make certain that the reader will understand precisely what the
author intended to say. On the other hand, MT, by automatically
generating a product where nothing existed before, has created a
new mode of work in which a less than ideal translation can be
provided for the consumer who merely needs to assimilate
information and is not concerned with disseminating it.
The first type of application is exemplified by the Spanish to
American (English) MT system (SPANAM) and English to Spanish
MT system (ENGSPAN) at the Pan American Health Organization
(PAHO) in Washington, DC, and the second by the SYSTRAN
operation at the US Air Force’s Foreign Technology Division in
Dayton, Ohio.

TRANSLATION FOR CLOSE SCRUTINY: PAHO

MT has been enlisted in the service of general-purpose practical
translation at PAHO 1 since January 1980. The first language
combination to be implemented was Spanish-English, using
SPANAM, the Organization’s batch MT system developed in-house

58
 A high-volume translation environment 59

(Vasconcellos and León 1988). E NG S PAN, a second, more

sophisticated in-house product with a strong English parser, 2
followed suit with English-Spanish in 1985 (León and Schwartz
1986).3 Later, SPANAM was entirely rebuilt using the conceptual
design of ENGSPAN. MT is now the principal mode of translation
for these two language combinations and is an integral part of the
Organization’s translation and terminology service.4
PAHO undertook to develop MT with two broad purposes in
mind: to meet its internal translation needs more efficiently, and to
disseminate health information in the Latin American and
Caribbean countries. Very few of the translations done at PAHO are
for information only. There are occasional applications of this kind
but the majority must be of a quality that will stand close scrutiny.
Some are for publication although few serve as the basis for
important decisions, covering diverse subjects ranging from the
commitment of resources to primary care for sick children.
SPANAM and ENGSPAN, each with a robust parser and large
deeply coded dictionaries (63,000 and 55,000 terms respectively),
are capable of handling a broad range of subjects and dealing
with free syntax in a variety of genres. Among the most
important subject areas are medicine, public health, sanitary
engineering, agriculture, computer science, management and law,
and there are a number of others as well. Cross-cutting this
spectrum is an even broader range of styles and discourse genres:
journal articles and abstracts, textbooks, manuals (both for
human health and for software), proposals for funding, reports of
missions, contracts and agreements, minutes of meetings,
business letters, diplomatic exchanges, certificates, product
specifications, supply lists, captions for displays and even
promotional materials and film scripts. This type of variety
actually presents more of a challenge than differences in subject
matter but SPANAM and ENGSPAN tackle them all. For most
applications the texts are post-edited by professional translators
working on-screen. The input texts are never pre-edited or
customized in any way. The only preprocessing is a quick review
by the clerical staff to ensure that the word-processing document
conforms to one of the standard formats used in the
Organization. Thus, from any perspective, it is safe to say that
MT at PAHO has been both designed and implemented as a
general-purpose product. Examples of SPANAM and ENGSPAN
output are shown in Figures 5.1 and 5.2.
Figure 5.1 Spanish-English unedited machine translation at PAHO.
Figure 5.2 English-Spanish unedited machine translation at PAHO.
64 Muriel Vasconcellos and Dale A.Bostad

The PAHO environment is distinguished by the fact that

production and system development are carried on side by side as
part of the same operation. As those who are working on a
particular translation are in the best position to suggest appropriate
target glosses, post-editors are encouraged to mark their copies of
machine output with suggestions for the dictionaries (new technical
terms being subject to review by the terminologist) and for system
improvement in general. The computational linguists, for their
part, conduct their research on production text and are constantly
monitoring the output to ensure that the two systems are
performing up to standard. Problems that are easy to solve are
dealt with from one day to the next. In this way, the post-editors
have the satisfaction of seeing their feedback incorporated. (More
difficult problems take their place on a ‘wish list,’ in order of their
priority). Some of the post-editors have learned how to update the
dictionaries themselves, and whether they do or not they still get to
see that they can have a personal hand in improving the quality of
the output.

Over 8 million words delivered

Getting under way

As of 1990, SPANAM and ENGSPAN had generated some 34,800
pages (8.7 million words) in the service of production translation,
with current machine output averaging around 6,000 pages (1.5
million words) a year. In 1989, MT was officially recognized as the
Organization’s primary mode of translation, as by that time it was
supporting at least 60 per cent of all regular production in the two
language directions. But the road to success had been bumpy at
times. The use of MT at PAHO went through several phases, and it
is of interest to take a look at this experience as it unfolded.
When SPANAM made its début, MT was operated as a separate
unit in PAHO, offering direct competition to the existing human
translation (HT) service. In view of the broad range of texts to be
translated, its managers realized from the beginning that efficient
and effective post-editing would be crucial to successful acceptance
of MT. With this in mind, a certified translator was hired on a full-
time basis to review and correct the output. This person also
updated the dictionaries based on problems that came up in the
 A high-volume translation environment 65

course of day-to-day production. The first major application was the

Organization’s biennial program budget, for which it was possible to
demonstrate savings of 61 per cent compared with human
translation, as well as a reduction in staff days of 45 per cent
(Vasconcellos 1984).
For the first five years, post-editing was done at no cost to the
requesting office. The choice of whether to use MT was left to the
end consumer. Many of the receiving units were pleased with the
service, and the good word spread. Cost was undoubtedly a factor in
the early popularity of MT, since the service was being offered for
free, whereas HT sometimes had to be farmed out and the resulting
cost charged to the client. Also, turnaround was faster. In addition,
MT was welcomed with enthusiasm because the delivered product
was machine readable, which was not true of HT at the time. Also,
MT with light post-editing was occasionally used for information
purposes only, and this service addressed a previously unmet need—
yet another factor that added to its successful reputation. Since no
one was forced to use MT, and since the service was free, there were
virtually no complaints.
By 1984, MT production from Spanish into English was reaching
100,000 words (400 pages) a month, and ENGSPAN was being
tried on an experimental application from English into Spanish. At
the end of that year, in the wake of a series of personnel changes
(including departure of the MT post-editor), the Organization’s
upper management decided to merge the MT and HT services and
delay the recruitment of new translators until a study could be done
on the overall allocation of human resources for translation. In the
interim, MT continued to be an option but post-editing had to be
done by contractors with the cost charged back to the requesting
offices. The contractors, most of them freelance professional
translators, were paid by the job at slightly more than half the then
prevailing rate for HT, and they came to the PAHO offices to use the
Organization’s word-processing equipment. As a result of these
changes, there was in fact a slight drop in MT production levels but
the monthly average never fell below 67,000 words (268 pages).
In the fall of 1987, still without any conclusive data in hand and
with several translator vacancies to be filled, PAHO’s management
decided to conduct an experiment that would yield concrete data
for decisions about the appropriate mix of the four possible
resources—namely, in-house MT, in-house HT, contract MT and
contract HT.
66 Muriel Vasconcellos and Dale A.Bostad

The experiment

A formal eleven-month experiment was undertaken from October

1987 through August 1988 to establish whether machine translation
in the PAHO environment was cost-effective, fast in turnaround and
capable of supporting a higher level of year-round average
production than human translation. The experiment also attempted
to establish whether MT output post-edited by a professional
translator was as serviceable as human translation.
Two temporary translators, one for each target language, were
recruited to serve as post-editors for the duration of the
experiment. MT ceased to be optional. All incoming work was
screened to determine whether or not it could be submitted to the
computer. A directive was issued instructing the requesting offices
to submit their texts for translation on diskette. In this way new
MT applications could be captured and use of the technology was
maximized.
Interim statistics for the first eight and a half months of the
experiment showed that MT, in-house and contract combined,
was used to process 79 percent of the jobs into English (839,635
words or 3,359 pages, corresponding to 262 jobs) and 60 percent
of those into Spanish (427,310 words or 1,709 pages, from 120
jobs). 5 MT was used less for Spanish because a large proportion
of the input texts were from published sources outside the
Organization and therefore not machine-readable. Optical
character recognition (OCR) was used whenever possible to scan
hard copy but it often turned out that cleaning up the scanner’s
mistakes was too timeconsuming to be worth the effort. 6 In total,
74 percent of the jobs into English were machine readable and, of
the rest, 5 percent could be scanned using the OCR but 21
percent could not. Into Spanish, on the other hand, only 54
percent were machine readable and, of the rest, 7 percent were
scannable and 39 percent were not.
Data for the full twelve months showed a significantly greater
volume translated by MT during the experiment than in previous
years. The texts represented a broad range of subject areas and
discourse types, as evidenced by the fact that more than 40 different
‘clients’ had been served.
The requesting offices, which usually did not know which mode
was used, were asked to provide feedback on the serviceability of
the product delivered to them. Responses were received for 30
 A high-volume translation environment 67

percent of the jobs processed. Both modes drew praise and

criticism but there was a slight preference for MT: 85.1 percent
indicated satisfaction with jobs processed using MT, compared
with 78.1 percent for HT. Interestingly, the lowest percentage of
satisfaction (60 percent of total response) was for Spanish HT.
Sometimes the client incorrectly assumed that the job had been
done using MT when in fact it was HT. For example, the following
comments were made about human translations: ‘machine
translation is not good enough. I practically had to rewrite it,’ and
‘the quality of the translation was more “rigid” than at other times;
it seemed more like a machine translation without having a
“human hand” go over it.’
The English in-house post-editor and some of the contractors
were able to achieve daily outputs of 7,000 words and higher in the
short term. In the long term, average daily in-house productivity for
English, calculated according to standard methods used in the
international organizations (i.e. net of leave, weekends and holidays)
and including periods without incoming translations during which the
translators worked on the dictionaries, was 2,546 words. This result
exceeded the United Nations target of 2,000 words a day by 27
percent and far surpassed the real average of 1,200 words a day
informally reported by many in-house services in the international
organizations. As MT does not involve transcription and certain
other types of general services support that were used for HT during
the period, it was considered that the goal of 30 per cent savings over
in-house HT had not only been achieved but in fact surpassed. For
contract translation, the cost differential between HT and MT was
self-evident, as MT post-editors were being paid 55 percent of the
HT rate.
As a result of the experiment, a new policy was announced in
February 1989 under which MT became the primary mode of
translation in PAHO. The policy entailed a restructuring of the
translation service so that as much incoming work as possible could
be channeled in the direction of MT.

Current situation
The use of MT has not been stabilized in PAHO. The new
technology continues to do the lion’s share of the work. The decision
to use MT, which rests entirely with the terminology and translation
service, is based on the following characteristics of the input text:
68 Muriel Vasconcellos and Dale A.Bostad

1 machine readability (or optical ‘scan-ability’);

2 complexity of format; and
3 linguistic characteristics (e.g. grammar, discourse genre, need for
between-the-lines interpretation, etc.).

These factors intersect with time-frame considerations, the

availability of post-editors and the special strengths of the individual
translators.
In-house translator positions have been established with
postediting built into the primary duty in the job description. At
the same time, contractors are still employed for much of the work
which, as always, comes in spurts. Compensation for contractors
has been brought in line with the three-tiered scale of the US
Department of State (general, semi-technical and technical). The
pay for MT post-editors is the same as for State Department
reviewers. A recent development is that most of the MT
postediting is now being done off-site by contractors working on
their home computers.

The translator and the MT output

The experience of the last three years, since the mechanism for client
feedback was formalized, has brought home increased awareness
that it is the translator’s skill, not the mode of work, that determines
the quality of the final translation product.
At the same time, not all the PAHO translators have chosen to
adapt to post-editing; some of the ones that we rely on have
preferred not to try it, and of those who have tried it, not all have
been equally successful. In total, about 30 post-editors have
worked with SPANAM or E NG S PAN and, of these, six into
Spanish and eight into English are now on the roster of contractors
regularly used. An effort is made to give these ‘regulars’ an HT
assignment every third or fourth time. There is a core group of
translators, all of them veterans, who consistently prefer the MT
mode.
Post-editing seems to be a special skill, somewhat related to
traditional editing. It involves learning how to preserve as much
of the machine’s output as possible and ‘zapping’ the text at
strategic points rather than redoing it from scratch. The post-
editor quickly develops a set of context-dependent techniques for
dealing with the patterns produced by the machine (Vasconcellos
 A high-volume translation environment 69

19 86, 19 87a, 19 87b; McElhaney and Vasconcellos 19 8 8;

Santangelo 1988).
There is a fine line to walk between allowing the output to stand
and meeting the high standard of quality that needs to be applied
to most of the translation jobs done at PAHO. In order to come up
to this standard, the post-editor must be certain of all technical
terminology, capture every nuance, provide accurate
interpretations for coherence (Vasconcellos 1989a), ensure
idiomatic equivalence, maintain a uniformly appropriate register,
build in cohesiveness where it is lacking, provide adequate ‘staging’
(Grimes 1975:323), and preserve the information structure
(Halliday 1967; Vasconcellos 1986).
Sometimes, in addition to using approved terminology, the
translation will call for special phraseology, as, for example, with
resolutions of the organization’s governing bodies. Here MT can
present a problem: the post-editor cannot settle for an equally clear
but unofficial version produced with the aid of the machine. In
such cases it can be seen that in-house translators and ‘regulars’ are
in a better position to make decisions about the MT output. To
help the contractors in particular, the documentalist, who is very
familiar with the workings of the Organization, reviews each
incoming job and locates all the background materials the
translator should have at hand.
When it is known that a translation will have only limited use—a
first draft, for example, or an internal trip report—MT offers the
possibility of delivering a less-than-ideal translation at a lower level
of effort—i.e. by doing less post-editing. PAHO is currently
attempting to implement an intermediate level of quality which
would be acceptable in such situations. The text is syntactically and
propositionally correct, but nuances may be sacrificed. So far, it has
been difficult for the translators to ‘lower their standards’, but
occasionally the press of time has been sufficient motivation for them
to cut the right corners. In-house translators are more appropriate for
such an approach, first because compensation is not at issue, but
mainly because they are in a better position to sense the minimum
level that will be acceptable. It is planned to experiment further with
translations of intermediate quality.
Unfortunately, it does not necessarily follow that when a
translation is of minor importance a raw or semi-polished MT
product can be used. It often happens that when the input text has
been prepared in a rush, maybe dictated but not read, or
70 Muriel Vasconcellos and Dale A.Bostad

transcribed from an oral presentation, it contains incomplete

fragments and grammatical lapses which throw the machine—and
the post-editor—into a tailspin. Not only must the output be
carefully post-edited, but the work of the post-editor is greatly
slowed down, to the point where sometimes the effort has to be
compensated at the full HT rate.
By far the most difficult challenge for the translators has been to
know when to leave PAHO-specific terminology untouched. Terms
that may look ‘funny’ to the uninitiated—such as fertility regulation,
where birth control is the more common expression, or harvesting and
its Spanish counterpart cosecha in reference to the culling of monkey
populations—are apt to get changed by posteditors.
Hypercorrections of this kind are the main source of complaints
from clients. Again, the advantage of in-house translators or
‘regulars’ who know the technical terminology is obvious. To help
deal with the problem, the entries in the SPANAM and ENGSPAN
dictionaries can be coded so that a word or phrase will be flagged
in the output, telling future posteditors that the term has been
researched and is reliable. This resource has been implemented to
a certain extent, but at the same time care needs to be taken not to
clutter up the output with so many flags that the text becomes
unreadable. And, of course, there is never a guarantee that a
flagged term will not require a different translation in another
context. Some of the other ways in which the post-editor can have
some control over the output are described in the next section.

The translator and the dictionary

PAHO post-editors can gain a certain amount of power over the
MT output by learning to manipulate the S PANAM and
ENGSPAN dictionaries, which offer a wide variety of approaches
for triggering expressions that are appropriate to the given
context. As a first step, they are expected to mark up their hard
side-by-side copy as they move along with each translation. The
following kinds of information can be recorded: glosses for words
that were not in the dictionary at all (which have averaged less
then 0.5 percent since 1984), different or additional glosses for
words that are in the dictionary, special translations for
multiword phrases, reliability flags, and notes about syntactic
errors the machine has made.
Perhaps the most immediately useful resource for the translator is
 A high-volume translation environment 71

the phrase-building capability. Whether or not the translator learns

to enter these units in the dictionary personally, it is still rewarding
to see them come up in future translations. If a particular term
happens to have required research, the translator’s efforts are
captured once and for all, avoiding the duplication that is so
common in translation services.
SPANAM and ENGSPAN can provide special translations in a
wide variety of situations. Some examples are: different subject areas
(English core for Spanish núcleo in an atomic energy text vs nucleus in
an ordinary text); singular and plural forms of the same source word
(toxicity for toxicidad and toxic effects for toxicidades); fixed phrases of two
to five words in length (in general for por lo general) or Pan American
Health Organization; long names of institutions and titles of
publications; multiword phrases that can be parsed in more than one
way if necessary; discontinuous phrasal verbs; glosses based on the
syntactic or semantic features of associated words (the English
source word raise with an inanimate object gives Spanish levantar;
with an animate object, criar; with a human object, educar; and
collocated with the word flag, izar). The post-editor can request these
translations or learn to actually enter them in the dictionary.
The post-editor also brings recurring structural problems to the
attention of the computational linguists.

Future applications
Thanks to ongoing dictionary work and system improvement,
ENGSPAN now produces raw output of sufficiently reliable quality
that consideration is being given to the translation of data bases and
other applications in which users can access MT directly. Of
particular interest are data bases that are available on compact disk
read only memory (CD-ROM). Several proposals have been made
and some of these may materialize into active projects.
The Consultative Group on International Agricultural Research
(CGIAR) has been collaborating with the PAHO MT project since
1986 and provided support for the installation of ENGSPAN at the
International Rice Research Institute in the Philippines and the
International Center for Tropical Agriculture (CIAT) in Colombia.
CGIAR is helping to form a donor consortium that will provide
PAHO with funds to adapt ENGSPAN to a microcomputer and
develop parallel systems from English into French and Portuguese,
as well as establish an MT center within the CGIAR network that
72 Muriel Vasconcellos and Dale A.Bostad

will support users and build specialized MT dictionaries in

agriculture.
The availability of PAHO’s MT systems on microcomputer will
enable the organization to outplace this technology in its member
countries, and this should contribute to the exchange and
dissemination of types of needed health information which have not
been translated in the past because of the cost and the limited
availability of translators.

TRANSLATION FOR INFORMATION: FOREIGN

TECHNOLOGY DIVISION, US AIR FORCE
Machine translation has undergone a long evolutionary
development at the Air Force’s Foreign Technology Division
(FTD), extending over more than twenty years. While for many
years the SYSTRAN Russian—English system was the sole MT
operation, with time additional language pairs were acquired:
French—English, German-English, Japanese-English and, recently,
Spanish-English. A large-scale machine translation environment
has been created, with some fifteen people inputting data, six or
seven people editing the data, and, since 1975, a group of four to
seven people constantly working on improving the dictionaries and
the parsing algorithms. At all times, development of the various
systems has gone hand in hand with use of the systems to translate
texts on a daily basis.
The SYSTRAN machine translation systems are used to translate
large amounts of information—roughly 50,000 to 60,000 pages of
source text per year. The translations are used by scientific and
technical analysts who need to keep abreast of foreign developments
in a wide range of technical fields and to prevent technological
surprise that could threaten the United States.
FTD got involved early with machine translation and has
consistently been a supporter of MT because it was enamored with
the prospects of what MT could do in the early years, and because
it was believed that machine translation was the only way to
translate massive amounts of material and get the translations back
to the requesters while the information was still useful. The early
intuition that machine translation could provide this service has
proven correct, and FTD has not veered from its solid commitment
to MT for over twenty years.
However, the way of producing machine-translation products has
 A high-volume translation environment 73

evolved over time at FTD, and it is of interest to briefly discuss how

FTD got started in the MT business and where it is now.
In the 1960s and the first years of the 1970s the Russian-English
system was used to produce a very rough hard-copy translation.
Post-editors edited the translation extensively—almost rewriting it—
and then the edited version was sent to a ‘recomposition’ pool where
it was retyped. Alternatively, some translators used the rough
machine translation as a ‘pony’ when they dictated the translation on
tape. The translator picked and selected those parts of the machine
translation that were valuable—certain technical terms and phrases—
and worked directly from the Russian text. The dictated translation
was then transcribed. Machine translation was really only used as an
aid to the translator. The end translation was a quite accurate
finished product.
Because this method was slow and productivity was low, in 1974
it was discarded and a very lightly edited hard-copy translation was
instituted. It was decided that this was the only way to cope with an
immense backlog that had built up. The degree of postediting was at
the discretion of the editor, but speed was the name of the game.
Changes were written in ink on large fanfold computer printout.
There were no stylistic changes. The end product could only be
called fairly crude and the format primitive—a low-quality utilitarian
product whose one saving grace was that it got the information
quickly to the requester and allowed an immense backlog to be
eliminated. This was FTD’s first effort at producing what we call
‘partially edited machine translation.’
In 1976 a dedicated input/editing minicomputer system using a
DEC PDP 11/05 and sixteen workstations was designed and
installed. This allowed the first on-line computer revision of
machine translation, and in 1977 the first version of EDITSYS was
written. EDITSYS is a software module called at the end of the
Russian-English analysis that identifies certain potential problem
areas in the output and brings these conditions to the attention of
the post-editor. When a given condition is met, the program
generates a full-page-width string of characters (a ‘flag’)
immediately in front of the condition. The post-editor must react to
the ‘highlighted’ condition and verify the machine translation
version or make a revision. The flag is removed during a later step
in the post-processing. Such conditions as ‘not-found’ words,
rearrangement, acronyms, problem words, mistyped words that
produce spurious match-ups, etc., are highlighted. Post-editing is
74 Muriel Vasconcellos and Dale A.Bostad

thus determined by a software program that tells the editors the

minimum number of conditions that have to be reviewed, although
the editors can—and some of them do—post-edit beyond the flagged
conditions. But sentences and blocks of text basically go through
unscrutinized by an editor. This type of partially edited text has
been our mainstay for the last fifteen years. The translation is then
printed on high-quality paper on a laser printer and has a much
better appearance than the previous product. This product has
been well received by FTD analysts.
As stated above, some editors do edit to a certain extent beyond
the flags. There are also other exceptions to the procedure just
described. Some short translations for dissemination require careful
sentence-by-sentence editing. Moreover, medical and biology texts, at
present, are closely edited for technical accuracy. Approximately 5
per cent of MT translations receive this type of post-editing. Finally,
some fifty to 100 translations per year, ranging from one paragraph
to two pages, are removed from the ‘first-in-first-out’ queue,
translated by MT, and then edited on hard copy by translators who
do not normally do post-editing. The idea is to get the jobs done
quickly and out the door; otherwise they tend to stagnate behind
larger jobs. Requesters expect fast turnaround on very short
translations; they understand that a 300-page book, for example,
takes longer.
In 1987, FTD developed a new MT application which we call
‘interactive machine translation.’ This system gives all users
individual access to MT at their own terminals. It is now available
to users via hookup with approximately 1,600 PCs within FTD.
This is raw machine translation without the mediation of
translators. The system is designed so that a user can rapidly
determine the significance of the material he wants translated and
weed out extraneous information. It is best used for rapid
translation of titles of books, tables of contents, captions of tables
and graphs, and individual sentences and paragraphs. However, it
can also be used effectively to translate complete short articles and
to get back a rapid translation instead of going through the
sometimes time-consuming operation of routing translations
through the formal bureaucracy. One very effective use of the
system is for ‘gisting’ a large book—that is, determining the
significant parts of the book and then routing this material through
the normal translation procedures. For example, if a user has a 350-
page book, the system might be used to determine that only
 A high-volume translation environment 75

Chapters 3, 7 and 12–15 are really pertinent to the research in

question. Obviously, the use of such a tool makes for tremendous
cost savings by eliminating the translation of irrelevant material.
FTD has conducted three extensive surveys of machine
translation over the last ten years to analyze the effectiveness and use
of MT and gain insights into how to improve the product. The two
most important insights coming out of the surveys are:

1 speed of translation is the most important consideration for FTD

analysis; and
2 the existing product, partially edited MT, is deemed satisfactory
in meeting most users’ translation requirements.

There has been some feedback spelling out deficiencies in certain

technical disciplines, and there was a small percentage of users who
found MT unsatisfactory as a translation product but overall the
acceptance rate has been very high.
The most recent survey, taken in December 1990, covered users
of both partially edited MT and interactive MT. Sixty surveys were
sent out; thirty-three were returned, for a return rate of 55 percent.
In addition, within this number, six users were personally
interviewed. According to the survey, 73 per cent of the respondents
felt that the current post-edited MT product was meeting or
exceeding their requirements for technical accuracy. An even higher
proportion—82 percent—stated that the post-edited product was
meeting or exceeding their requirements for readability! There was
high praise for the interactive system. However, because the analysts
themselves have to input the data, it was not deemed effective for
longer documents. More analysts would be willing to use raw MT if
the data could be input for them. They would foresake the current
post-editing if the translation directorate would input the documents
and send the raw MT product electronically to them directly after
translation. Very recently the utility of raw MT was emphasized by
the fact that 600 pages of French had to be translated in two weeks
and, due to the time constraint and the lack of post-editors from
French, the only translation that could be produced was raw MT.
The requester accepted the raw MT in order to get quick access to
the data and meet his deadline.
Recently an assessment of Russian partially edited MT was made
by an independent group of scientists conducting Air Force research
in a subfield of physics. For one particular portion of a book they
76 Muriel Vasconcellos and Dale A.Bostad

had access to both a human translation (done for another

government agency) and a machine translation. Their final report
contained the following unsolicited comments on the quality of the
two English translations:

While the [human] translation read somewhat more smoothly, it

seemed to use inappropriate or erroneous terminology more often
than the [machine] translation did. Consequently, we relied
primarily on the [machine] translation, using the [human]
translation mainly for reference.

FTD is now embarked on an ongoing project to improve the

efficiency of its MT operation. The areas being addressed include:
incorporation of software for individual users to modify dictionaries;
OCRs to scan Russian, French and German; expansion into other
language pairs; and continued refinement of specialist dictionaries.

NOTES
1 Regional Office of the World Health Organization for the Americas.
WHO is a specialized agency in the United Nations system.
2 Developed with partial assistance from the U S Agency for
International Development (AID) under Grant DEP-5443-G-SS-3048–
00. ENGSPAN is installed at AID and runs there on an IBM 3081
(OS/ VMS).
3 SPANAM and ENGSPAN are written in PL/1 and run on PAHO’s
IBM mainframe computer, currently an IBM 4381 (DOS/VSE/SP),
which is used for many other purposes.
4 The Organization’s working languages are Spanish and English. The
English-Spanish and Spanish-English combinations account for 90 per
cent of the translation workload. Portuguese and French, which
together make up the other 10 per cent, are also official languages of
the Organization but are handled by a separate service.
5 For a detailed review of the data from the experiment, see Vasconcellos
1989b.
6 The equipment on hand was already old at the time of the experiment.
Current OCR technology would undoubtedly do much better.

BIBLIOGRAPHY
Grimes, J.E. (1975) The Thread of Discourse, The Hague: Mouton. Halliday,
M.A.K. (1967) ‘Notes on transitivity and theme in English’, part 2,
Journal of Linguistics 3:199–244.
Hartmann, R.R.K. and Stork, F.C. (1976) Dictionary of Language and
Linguistics, New York: Wiley.
 A high-volume translation environment 77

Hutchins, W.J. (1988) ‘Future perspectives in translation technologies’, in

M.Vasconcellos (ed.) Technology as Translation Strategy (American
Translators Association Scholarly Monog raph Series, vol. I I),
Binghamton, New York: State University of New York.
León, M. and Schwartz, L.A. (1986) ‘Integrated development of English-
Spanish machine translation: from pilot to full operational capability:
technical report of Grant DPE-5543-G-SS-3048–00 from the US Agency
for International Development’, coordinated by M.Vasconcellos,
Washington, DC: Pan American Health Organization.
McElhaney, T. and Vasconcellos, M. (1988) ‘The translator and the
postediting experience’, in M.Vasconcellos (ed.) Technology as
TranslationStrategy (American Translators Association Scholarly
Monograph Series, vol. II), Binghamton, New York: State University of
New York.
Santangelo, S. (1988) ‘Making an MT system work: perspective of a
translator’, in M.Vasconcellos (ed.) Technology as Translation Strategy
(American Translators Association Scholarly Monograph Series, vol. II),
Binghamton, New York: State University of New York.
Vasconcellos, M. (1984) ‘Machine translation at the Pan American Health
Organization: a review of highlights and insights’, Newsletter, British
Computer Society Natural Language Translation Specialist Group.
——(1985) ‘Theme and focus: cross-language comparison vis translations
from extended discourse’, unpublished PhD dissertation, Georgetown
University, Washington, DC.
——(1986) ‘Functional considerations in the postediting of MT output:
dealing with V(S)O versus SVO’, Computers and Translation 1, 1:21–38.
——(1987a) ‘Postediting on-screen: machine translation from Spanish into
English’, in C.Picken (ed.) A Profession on the Move: Proceedings of Translating
and the Computer 8, London: Aslib.
——(1987b) ‘A comparison of MT postediting and traditional revision’, in
K.Kummer (ed.) Proceedings of the 28th Annual Conference of the American
Translators Association, Medford, New Jersey: Learned Information.
——(1989a) ‘Cohesion and coherence in the presentation of machine
translation products’, in Georgetown University Round Table on Languages and
Linguistics 1989, Washington, DC: Georgetown University Press.
——(1989b) ‘Long-term data for an MT policy’, Literary and Linguistic
Computing 4, 3:203–13.
Vasconcellos, M. and León, M. (1988) ‘SPANAM and ENGSPAN: machine
translation at the Pan American Health Organization’, in J. Slocum (ed.)
Machine Translation Systems, Cambridge: Cambridge University Press.
Chapter 6

Esperanto as an intermediate
language for machine translation
Klaus Schubert

PRESTIGE

According to an unconfirmed rumour, on the occasion of a

prestigious machine translation conference held in Japan in 1989, a
senior representative of a major Japanese computer manufacturer
reported that his company had sold X number of machine
translation systems and he proudly added that 10 per cent are actually
used. This cannot make sense unless 90 per cent of the customers
buy machine-translation systems like Van Gogh paintings: for
prestige.
True or untrue, the story may well confirm that a trade as uncertain
as machine translation, with its extremely long payback periods, is
indeed sensitive to prestige considerations. It could therefore seem
risky to include in a machine translation project a language like
Esperanto which has the unmerited but undeniable quality that the
mere mentioning of its name calls forth the most emotional rejections
from both laymen and linguists (Forster 1987; Piron 1987).
Despite all this, the idea of using Esperanto in machine translation
is almost as old as the attempts to make computers translate. In
recent years, the idea has become a reality, and I shall here address
the question, ‘what is so special about Esperanto that a prestige-
sensitive software company chose to adopt it?’.
The first section gives a short history of how Esperanto became
involved in machine translation. The second section prepares the
discussion about Esperanto as an intermediate language,
explaining why an intermediate language is needed at all. The third
section points out that the choice of an intermediate language is
conditioned by the basic design of the machine translation system.

78
 Esperanto as an intermediate language 79

The fourth section discusses the virtues and drawbacks of

alternative intermediate representations, and the fifth section shows
under which circumstances the best choice is Esperanto. The final
section describes in greater detail those properties of Esperanto that
make it particularly suitable for functioning within a machine
translation system.

ESPERANTO’S ROAD INTO MACHINE TRANSLATION

Esperanto became associated with computational linguistics (which
in the early decades was almost exclusively machine translation) in
three stages (Schubert 1989a:26–9). These may be labelled:

1 ‘the idea’;
2 ‘Esperanto on equal terms’; and
3 ‘Esperanto for its specificity’.

The first stage, ‘the idea’, had its origin in the very early years of
machine translation in the late 1940s and early 1950s. After the first
wishful attempts, it was soon understood that natural language is
more intricate than the decoding tasks the first computers had
performed well for military and intelligence applications. When
natural languages turned out to be too difficult, it was suggested that
something more consistent be tried, such as, for instance, Esperanto.
Yehoshua Bar-Hillel put forward this suggestion in his famous state-
of-the-art report of 1951 (Bar-Hillel 1951; Hutchins 1986:33–4).
This first stage of ideas about Esperanto in computational
language processing was preceded by a period when various scholars
used Esperanto as a kind of universal syntactic representation. Most
interesting in this respect are Lucien Tesnière’s dependency syntax
(Tesnière 1959/1982:64) and Petr Smirnov-Trojanskij’s method of
machine translation developed prior to the computer age (Denisov
1965:80ff.). Both systems were developed in the 1930s.
The second stage, ‘Esperanto on equal terms’, begins when
Esperanto is actually used in computational linguistics. First, a series
of studies appear which merely investigate the feasibility of the idea
(Sjögren 1970; Kelly 1978; Dietze 1986), then smaller programs are
written of which only a minority may have been published (Ben-Avi
1977), and, finally, larger implementations are realized. Such
implementations are carried out, for example, within the SUSY
machine-translation project in Saarbrücken (Maas 1982, 1985,
80 Klaus Schubert

1987:40), at Chubu University in Aichi, Japan (Katumori and

Hukuda 1984; Makino et al. 1986) and at the Chinese Academy of
Social Sciences in Peking (Li Wei 1986) and in other places. In the
age of the personal computer, much work was done by individuals
(Kat 1985; Mohai 1986; Briem 1990). Esperanto was also addressed
in speech processing (Sherwood 1978, 1982, 1985; Gordos 1985;
Koutny et al. 1988), in computer-aided instruction (Sherwood and
Sherwood 1982; Janot-Giorgetti 1985) and even in combined
systems including both machine-translation and speech-processing
functions (Sato and Kasuya 1987; Sato 1989:165–76). I call this
stage ‘Esperanto on equal terms’, as in these projects and systems
Esperanto is treated basically as any other, ethnic, language. There is
only a difference of degree in that Esperanto turns out to be easier to
handle than ethnic languages, because of its great formal
consistency.
The third stage, which I term ‘Esperanto for its specificity’, begins
with the DLT machine-translation project. Distributed language
translation (DLT) is the name of a long-term research and
development effort by the Dutch software company Buro voor
Systeemontwikkeling (BSO), in Utrecht. Based on an invention by
Toon Witkam of BSO in 1979, DLT took shape in 1984–7, when a
first prototype version was implemented (feasibility study: Witkam
1983; prototype architecture: Schubert 1986; syntax: Schubert 1987;
prototype semantics: Papegaaij et al. 1986; latest news: Sadler
forthcoming), followed by an improved one in 1988.
In the DLT system, Esperanto functions as the intermediate
language. The original idea was to include in international
datacommunications networks a facility that would allow each
subscriber to read and to contribute messages in their own language.
Potential applications are not confined to public or corporate
electronic mail; they include document management, information
retrieval and other functions where variable volumes of text are
stored, accessed and updated in several languages. The transmission
form in such a network would then be the DLT system’s
intermediate language, i.e. Esperanto. Translation from the source
language into Esperanto takes place in the sender’s office, before the
text enters the network, and translation from Esperanto into the
target language is carried out in the computer where the receiver
reads the message. (Hence the term distributed in the system’s name.)
It is not the aim of this chapter to give a full account of the
linguistic and computational developments that were required to
 Esperanto as an intermediate language 81

realize this idea to the extent to which it has been accomplished at

the time of writing. Here, I shall only address the question, ‘which
properties of Esperanto make it suitable for this application?’.

WHEN IS AN INTERMEDIATE LANGUAGE NEEDED?

Before answering the question, ‘under what circumstances should
the intermediate language of a machine translation system be
Esperanto?’, it may be useful to explain why an intermediate
language is needed at all.
Translating is an intricate intellectual process, and attempting to
have it done by a computer does not by any means make the task
easier. Machine translation necesarily involves a series of steps which
use linked but distinct modules. Representations are therefore
needed for partially processed versions of the text. As soon as the
design of the system includes more than two languages,
combinatorial problems arise. The so-called transfer systems link
each pair of a source and a target language directly via a series of
intermediate steps specific to that language pair. It is well known that
systems with this architecture need for n languages n (n-1)2 transfer
modules with as many language-pair-specific representations. A
transfer system for the nine official languages of the European
Community, for instance, would thus need no fewer than seventy-
two translation modules. The transfer representations are usually
made up of elements from the two languages involved, plus some
formal symbols. The more languages such a system contains, the
more expensive it becomes to add yet another language: 2 n
interfaces to and from the n languages already included (eighteen
new modules for adding a tenth language).
A remedy for this combinatorial explosion is a true intermediate
representation. The n (n-1)2 formula is based on the assumption that
every source language is linked directly with every target language.
If these direct links can be given up in favour of a single, central
representation, the combinatorial problem is removed. Then a text
from whatever source language is first translated into the
intermediate representation and from there into the target
language(s) chosen. This set-up requires only 2 n modules, and the
addition of a new language requires only two modules (thus eighteen
instead of seventy-two modules for the EC example, two more for a
tenth language).
Thus, if a system is designed to be multilingual, an intermediate
82 Klaus Schubert

representation is required to keep the complexify of the architecture to

a minimum and, at the same time, to make the system easily
extensible.

CONDITIONS FOR THE INTERMEDIATE LANGUAGE

In principle, most of what was said in the previous section is basic
knowledge for anyone acquainted with machine translation. It is not
always fully acknowledged, however, that the reasons which
motivate the inclusion of an intermediate representation in the
architecture of a system inherently imply a number of conditions for
the nature of that representation.
First, an intermediate representation can only help to avoid
combinatorial explosion if it is the sole link between any source and
target language. The system’s architecture can only shed the
combinatorial problems (and their cost) if each language is
connected to the intermediate representation only. No direct links to
other languages can be permitted. To achieve this goal, it does not
suffice merely to allow the data to flow technically through the one
intermediate representation. It is essential that the source and target
language modules be truly independent. Second, therefore, a text
represented in the intermediate form must not contain any features
that are specific to a given source or target language. The first
requirement addresses, so to speak, the external combinatorial
explosion. The second requirement seeks to avoid a similar
explosion that would otherwise occur within the intermediate
representation.
The second condition is less obvious. A transfer system without
a central intermediate representation needs its n (n-1) 2 transfer
modules, as, for instance, translating from English requires one set
of rules and dictionaries when the target language is German, and
a different one when the target language is Italian, and translating
into Portuguese requires one set of rules and dictionaries when the
source language is Danish, and a different one when the source
language is Greek. In a system with an intermediate representation,
this combinatorial problem must necessarily be removed as well,
otherwise the system cannot profit from its less explosive
architecture. If this condition is not observed, there may be 2 n
modules but there will nevertheless have to be n (n-1)2 transfer rule
systems and dictionaries. This is what I call the internal explosion.
In small ‘toy’ systems it can be circumvented, but in real
 Esperanto as an intermediate language 83

production-scale machine translation systems the internal

explosion is disastrous and should be avoided in the basic design of
the system architecture.
As a consequence of all this, an essential requirement for the
intermediate representation is that it should be fully independent of
the source and target languages. At this stage of the reasoning, this
does not necessarily mean that the words (or other symbols for
semantic elements) in the intermediate representation could not be
taken from a language which also functions as a source and target
language in the system. The requirement only means that a text,
represented in intermediate form, must not bear any particular traces
of its original language that would be needed to translate it in any
specific form into the target languages.
At this stage of the argument, one might well build an
intermediate representation using English words. In the case of the
DLT system, other considerations make this impossible and make
Esperanto preferable.

ALTERNATIVE INTERMEDIATE REPRESENTATIONS

When an intermediate representation is designed, several basic
decisions have to be made. The first decision is between a language
and an artificial symbol system. If one opts for a language, the next
decision is between an ethnic language and a planned language. And
finally, in the group of ethnic languages there are several thousands
to choose from, in the group of planned languages several hundreds
(Dulicenko 1989:51).
As is often the case in language-related system design, the
decisions are in reality not as clear-cut as they appear to be in theory.
The most commonly adopted solution is a more or less strongly
artificial symbol system with words, morphemes or other elements
from an ethnic language among the symbols. Such a representation
is on the artificial side of the scale, but it is not totally devoid of
linguistic elements. The opposite solution is a text in a plain human
language, ethnic or planned, normally with a few enrichments or
precisions for which artificial, non-linguistic, symbols are applied.
The DLT solution belongs to the latter kind.
The independence requirement, formulated above, does not at
first sight appear to have much bearing on the question, ‘what kind
of elements should the intermediate representation be made up of?’
Nonetheless, it has a corollary which is so obvious that its
84 Klaus Schubert

consequences are often overlooked. As the intermediate

representation by virtue of the independence requirement must not
contain any source or target language-specific features, it should
itself be self-contained and fully capable of expressing the complete
content of the text being translated.
From one of the classics of linguistic theory, Louis Hjelmslev, we
learn an interesting consequence of this expressiveness
requirement. Hjelmslev has investigated the relationship between
artificial symbol systems and human language. Interestingly
enough, he arrives at a criterion which establishes a direct link to
the issue discussed here: translatability. Hjelmslev shows that
meaning expressed in an artificial symbol system can always be
translated into a human language, whereas a text in a human
language cannot always be translated into a given artificial symbol
system. In other words, according to Hjelmslev, artificial symbol
systems are inherently less expressive than human languages
(Hjelmslev 1963:101; Schubert 1988:137). More precisely, artificial
symbol systems are so much less expressive than human languages
that it is impossible to translate the full content of an ordinary text
into an artificial system. Thus, the difference in expressiveness
between an artificial symbol system and a human language is not
merely one of degree but one of quality.
If Hjelmslev is right in this finding, the attempts to use artificial
symbol systems as intermediate representations in machine
translation are in obvious contradiction to the independence
requirement and to its corollary, the expressiveness requirement.
As a consequence of Hjelmslev’s diagnosis, artificial symbol
systems cannot fulfil the function of an intermediate representation
in machine translation. A single scholar’s opinion may not be
accepted as a proof, but, to say the least, the experience of the
machine translation community with intermediate representations so
far certainly does not suggest the contrary! A possible solution, then,
is to adopt a human language for the intermediate function, that is,
to opt for an intermediate language instead of an intermediate
representation.
The experience in both literary and informative translation shows
that translating through a bridge language is without any doubt
possible but normally leads to a lower-quality result. A high-quality
machine translation system therefore needs an ‘upgraded’ version of
a human language. Such a solution has been tried out with
considerable success in the DLT project. The discussion in the next
 Esperanto as an intermediate language 85

sections shows that, contrary to what critics sometimes assume, the

solution found for the DLT system is not being put forward as a
universal and generally applicable remedy for all machine
translation systems of whatever type, but that it is a carefully chosen,
made-to-measure instrument for a specific, well-described purpose. It
is in this context that the choice of Esperanto should be understood
and appreciated.

ESPERANTO AS A CONSEQUENCE OF SYSTEM DESIGN

Belonging to a new generation which is not yet on the market, the
DLT machine translation system was given a clearly more
ambitious design than existing systems. From the very first plan
drawn up by its inventor, Toon Witkam, DLT’s design has been
ambitious in a fundamental way: it is aimed at the monolingual
layman user. This is noteworthy, as the DLT designers subscribe to
the generally agreed finding of the 1950s that fully automatic high-
quality translation is impossible. The basic parameters of the DLT
architecture therefore include—in addition to the intermediate
language, the primary functions of which are to avoid
combinatorial explosion and to maintain the extensibility of the
system—a system-initiated disambiguation dialogue. The dialogue
takes place during the first half of the translation process, i.e., when
the text is being translated from the source language into the
intermediate language. In the dialogue, the user is asked to decide
such cases of structural or semantic ambiguity in the source text
that cannot with sufficient confidence be resolved by the system
itself. The dialogue is carried out entirely in the source language
and its wording is arranged in such a way that no knowledge of
grammar, of linguistic terminology or of the target language(s) is
needed, let alone of Esperanto.
Along with the requirement for expressiveness formulated above,
the goal of high-quality output places a severe condition on the
intermediate language. Because of the distributed architecture, a
disambiguation dialogue can reasonably be included only in the
source-to-intermediate part of the overall translation process, so that
there is interaction once for each text, but not once for each text and
each target language. A logical consequence is that the intermediate-
to-target parts have to be fully automatic, and obviously they should
deliver translation of a high quality. Existing machine translation
systems require post-editing, an activity to be carried out by a skilled
86 Klaus Schubert

translator or reviser for each of the target languages involved. By

contrast, the DLT design aims at ready-to-use high-quality output
which does not need post-editing.
In the DLT setting, the intermediate language should thus allow
for fully automatic high-quality translation, a thing which is known
to be impossible between ‘normal’ human languages. As I have
shown above, the combinatorial considerations suggest the use of
some intermediate language, no matter which. The additional
requirement of fully automatic translation from the intermediate
language places a severe restriction on the designer’s freedom of
choice. Given this requirement, the intermediate language has to be
much more precise than the source and target languages, as it has to
render the results of the disambiguation process. It is through the
additional information gathered in the disambiguation steps, plus the
inherent clarity of the intermediate language itself, that fully
automatic high-quality translation can be achieved in the second half
of the process.
This is the essence of the reason for choosing Esperanto as the
intermediate language: the expressiveness requirement makes it
necessary that the language into which texts from all source
languages are translated be a human language rather than an
artificial symbol system. The requirement of fully automatic high-
quality translation from the intermediate language into any one of a
number of target languages makes it necessary for the intermediate
language to be fundamentally clearer and more precise than
ordinary languages. In the DLT setting, Esperanto has been found
to offer an excellent way out of this seemingly contradictory
situation.

WHAT MAKES ESPERANTO SO WELL-SUITED FOR

MACHINE TRANSLATION?
In the previous sections I have discussed the external aspects of the
situation in which Esperanto functions as an intermediate language
in the DLT system. The essence of the argument so far is the
insight that Esperanto combines the expressiveness of an ethnic
language with the extreme clarity of an artificial symbol system. I
have not yet offered much proof of this. In this section, therefore, I
shall go more deeply into the grammatical (and even
sociolinguistic) properties that make Esperanto especially suitable
for this purpose.
 Esperanto as an intermediate language 87

In order to make this discussion interesting for those readers who

are not especially interested in the particulars of the DLT machine
translation system, I have chosen as a starting point a rather
fundamental view of machine translation, or, in fact, of natural-
language processing in general.

The essential problem of natural-language processing

Natural-language processing, and in particular its oldest endeavour,
machine translation, sometimes seems to be a neverending struggle
towards a goal that ultimately cannot be reached. In my view, the
essence of the problem lies in the fact that language is infinite.
Language is an open-ended system. There is always more than
what can be covered by explicit rules or enumerated in dictionaries
or in lists of exceptions to the rules. There are no sharp
borderlines, especially between grammatical and ungrammatical
utterances, between existing and non-existing words and
expressions, between allowed and forbidden combinations of
words or meanings, etc. All this makes language vague, but this
vagueness is an extremely fruitful prerequisite for language to cope
with the infinite multitude and diversity of situations that people
need to communicate about. A symbol system which is not infinite
in this vast and multi-layered sense is insufficient. Hjelmslev
understood this.
As a consequence of this infiniteness of language it is often
impossible to cover every phenomenon encountered in a text to be
translated in pre-formulated rules that range over finite or
enumerable sets of elements. A significant stream of research
interest in general and computational linguistics has, therefore, for
quite some time been focusing on corpora as the basic source of
knowledge about language. In current work on natural-language
processing at the BSO software company, this approach has been
rigidly adhered to. Based on the experience gained with the DLT
prototypes. Victor Sadler has developed a method of using corpora
not only to inspire grammarians and lexicographers (who would
then write explicit rules and entries), but directly as a knowledge
source in machine translation (Sadler 1989: Part I I; Schubert
1990:219). It may be noteworthy that a corpus in this sense is a set
of examples. A technique with direct use of corpora thus implies
that the knowledge needed for the automated translation process is
derived from examples, and translating on the basis of examples
88 Klaus Schubert

means that the basic principle used is analogy. If properly applied,

this principle caters very smoothly for the infiniteness of language:
instead of trying to describe the precise limits of linguistic
phenomena, the analogy-based technique uses typical examples of
constructions, translation equivalents, etc., that are known to be
correct and extrapolates from those to new texts previously unknown
to the system.
The details of this approach are far beyond the scope of this
chapter. For the discussion here it is interesting that, as a
consequence of a technique which takes into account the infiniteness
of language, the capacity to accommodate extrapolation processes in
a comfortable and expedient way becomes a touchstone for an
intermediate language in machine translation.

Removing syntactic ambiguity

The DLT feasibility study devoted extensive attention to
modifications in the grammar of Esperanto that were felt necessary
for its function in DLT (Witkam 1983:IV-1-IV-111). During the
development of the prototypes, however, it was found that most of
these modifications could be abandoned, so that DLT’s
intermediate language became more and more similar to common
Esperanto as it is spoken and written by humans. (This obviously
facilitated the transition to corpus-based techniques later on, for
which it is advantageous for texts in undoctored language to be
incorporated in a straightforward manner.) At the end of the
prototype phase, the difference between common Esperanto and
DLT’s version can be summarized under a single condition; DLT’s
Esperanto is syntactically unambiguous. The term syntax is meant in
a broad sense which covers all grammatical properties on the
formal side of the linguistic sign (Schubert 1987:14–16).
Essentially, this is achieved by the insertion of two kinds of
delimiter. As Witkam (1983:IV-93) had suggested, extra spaces
between words are used as markers for the end of a syntagma (a
phrase). This resolves most of the notorious attachment problems
that are extensively discussed in syntax and parsing theory. (Ni
acet’as libr‘o‘n de Goethe ‘We buy a-book by Goethe’: normal single
space before de, thus de Goethe belongs to libr‘o‘n and translates as
‘by’; Ni acet‘as libr‘o‘n #de Goethe ‘We buy a-book from Goethe’;
extra space [shown as #] before de, indicating a syntagma
boundary so that de Goethe cannot belong to the [closed] syntagma
 Esperanto as an intermediate language 89

libr‘o‘n and accordingly belongs to acet‘as and translates as buy from.)

As Esperanto is totally agglutinative, DLT also introduced overt
markers of morpheme boundaries as shown above. In this way,
ambiguity in the analysis of words is excluded. (Common
Esperanto sendata becomes either sen‘dat‘a ‘without’ + ‘date’ +
adjective marker = ‘undated’ or send‘at‘a ‘send’ + passive
progressive participle + adjective marker = ‘being sent’.)
It may be surprising that these two minor artificial elements
should suffice to make a human language unambiguous, albeit only
at the word- and sentence-syntactic level. Obviously, this is possible
in Esperanto only because the language has a highly unambiguous
structure anyway. A well-known feature is the word class system: all
content words bear an identifying morpheme (e.g. a for adjectives,
see above), and the function words are a finite list. With the
exception of a handful of function words, no word class ambiguity
exists. This may not seem much but in view of the enormous
problems such a simple thing as word class recognition brings about
in parsing and corpus analysis efforts for English, this clear structure
of Esperanto is a major asset.

Productive word-formation
One of the central and most efficient aspects of the infiniteness of
language is the possibility of combining existing linguistic signs to
express new meanings. In all languages, the formation of syntagmata
and sentences offers an opportunity for the expression of new
meanings, but many languages also have a very productive
combinatorial capacity within words. Productive word formation is a
major instrument of infinite expressiveness. In English, this
possibility is marginal, in Esperanto, as in all agglutinative
languages, it is an essential feature.
Esperantology has brought about a highly elaborated theory of
word formation (Schubert 1989b). On the basis of a grammatical
foundation which was designed as an open, extensible system and
under the influence of the pragmatic factors to be discussed below,
Esperanto has, in its hundred-year history, developed a highly
compositional system of word-formation. As a consequence, the
language provides a powerful mechanism for forming new words
from existing roots, while at the same time, the derivation of the
meaning of these words can be automated to a high degree. (This
derivation mechanism cannot be fully exhaustive for theoretical
90 Klaus Schubert

reasons. A language in which this were possible would have lost its
infiniteness in a crucial field.)
The ability to isolate the component elements of complex words
using automated processes is extremely important for a good and
effective functioning of word-based knowledge banks such as those
used in DLT, where the basic knowledge source is made up of
structured corpora. A speciality of Esperanto is, in this respect, its
unusually high degree of compositionality. This means that the
meaning of an extremely high percentage of complex words can be
inferred in a straightforward manner from the individual meanings
of their component elements. This is the case with Esperanto, as its
design objective was ease of learning and inferability is an obvious
advantage in language learning. As it turns out, the same
characteristic pays off in machine translation as well. The fact that
Esperanto not only started with a perspicuous word-formation
system but maintained this system intact throughout a hundred
years of uncontrolled use can be attributed to the pragmatic factors
addressed below.

Why semantic ambiguity cannot be removed

It is not sufficient to remove ambiguity at the syntactic level. More
intricate and more challenging to machine translation is ambiguity at
the semantic level. What about Esperanto in this regard?
The sceptics rail against Esperanto with an endless store of
arguments and prejudice. The optimists, on the other hand,
sometimes smile complacently and say, ‘Oh, I understand, with this
Esperanto of yours, you have found a totally unambiguous language
for the computer!’ The sceptics have been proved wrong by the
experience gained in DLT practice, and elsewhere. But the optimists
are mistaken, too. At the semantic level, Esperanto is not
unambiguous, because there is no such thing as semantic
unambiguity. Slogans like ‘one word—one meaning’ are nonsensical,
as they presuppose that there is an objective, cross-linguistically
valid, dissection of meaning in discrete portions. There is, however,
plenty of evidence that no such portions exist. This is amply borne
out by the unsuccessful attempts at defining basic sets of such
portions of meaning, normally called semantic atoms or semantic
primitives. The art, then, is not the dissection of meaning with a
view to achieving basic, universally valid elements that cannot be
subdivided further and constructing therefrom an intermediate
 Esperanto as an intermediate language 91

representation; it is the handling of meaning in whatever shape it

takes in different, fully expressive languages.
In addition, it should be borne in mind that the main problem in
machine translation is not the ‘semantic ambiguity’ that some
scholars attempt to resolve at the monolingual level. The main
problem is the ambiguity at the translation level, that is, the semantic
ambiguity in lexical or structural transfer. In other words, the main
hurdle is the ambiguity between languages, rather than the ambiguity
within languages. To put it simply: if your English-Chinese
dictionary offers you ten translations for machine, which one do you
choose, what are your criteria for choosing it and how do you
automate this choice?
The DLT approach addresses this problem by aligning parallel
corpora of translated texts and taking them as the basic knowledge
source. DLT’s ‘bilingual knowledge banks’ are built from these
corpora, one side of which is in Esperanto. The bilingual
knowledge banks are based on the principle of extrapolation. For
extrapolation, obviously, a regular language structure is a major
advantage.

A word about pragmatics

Compositionality in the semantic part of a language’s grammar is
expedient for machine translation. Many readers will agree with
that. In Esperanto, as in other languages, Compositionality is not
total. So, what is so special about Esperanto?
The answer is twofold: first, Esperanto has developed from a
deliberately constructed grammatical system, initiated only a
century ago, and has thus been able to preserve its
compositionality to a very considerable extent. This gives it a
quantitative advantage compared to ethnic languages. Second,
however, being a language spoken in an international language
community, Esperanto has a qualitative advantage over other
languages. As Esperanto’s main and almost exclusive function is
international communication, it has always maintained its special
suitability for communication between people with radically
different linguistic backg rounds and preconceptions. This
condition has favoured a development on the basis of the intrinsic
regularities of the language itself, rather than through imitating
other languages and adopting loan patterns. Because of this
pragmatic factor, Esperanto has always developed with a natural
92 Klaus Schubert

tendency towards consistency. This is an important asset for its

function in machine translation.

Two prejudices
Finally, a few words about two frequently heard prejudices
concerning Esperanto that are closely related to the subject of this
discussion. First a practical, then a theoretical one.

Prejudice 1
‘I understand that some simple corespondence among pen pals and
stamp collectors is possible in Esperanto. But, of course, you can’t
say everything in that language, can you?’ Of course you can. Like
any other language, Esperanto can express everything which has
already been expressed in it. This sounds contradictory but it holds
for all languages, English as well as Yoruba: if you have to prove that
something can be said in a language, you will only succeed if this has
been done (and recorded) before. So the problem resides not only in
existing language use but in the language’s capacity to develop.
‘Western’ languages like English, German or Russian have very rich
vocabularies, but terminologists still work hard to extend them, as
unfortunately, they experience that you cannot really say everything
in English, German or Russian. The same holds true for Esperanto.
The language has the capacity to develop, and it has developed
whenever it was necessary for it to do so. Unlike ethnic languages,
however, Esperanto develops in the sociolinguistically unique setting
of international communication.

Prejudice 2
‘You argue that artificial symbol systems are insufficient. But
Esperanto is an artificial language par excellence, isn’t it?’ It is not.
What is often overlooked in this sort of discussion is the fact that
Esperanto has already entered its second century. When the first
textbook of Esperanto was published in 1887, Esperanto was
artificial, and it was not yet a language. But since then, Esperanto
has, in a slow and unnoticed development, become a language
spoken by people. According to Detlev Blanke (Blanke 1985:105ff
and Tabelle 2), Esperanto is the only project of a planned language
which has totally accomplished the transition from an artificial
 Esperanto as an intermediate language 93

system to a real language. So, the argument should now be based on

a serious study of Esperanto as it exists today, not on remarks
quoted by journalists from outdated encyclopaedias.

CONCLUSION
The experience of the DLT machine translation project so far has
shown that Esperanto fulfils a specific requirement in language
technology: it can be used to good advantage as an intermediate
language in machine translation, when fully automatic high-quality
translation from the intermediate language into the target
language(s) is aimed at.

BIBLIOGRAPHY
Bar-Hillel, Y. (1951) ‘The state of machine translation in 1951’, American
Documentation 2:229–37.
Ben-Avi, S. (1977) ‘An investigation into the use of Esperanto as an
intermediate language in a machine translation project’, PhD Thesis,
Manchester.
Blanke, D. (1985) Internationale Plansprachen, Berlin: Akademic-Verlag.
Briem, S. (1990) ‘Maskinoversaettelse fra esperanto til islandsk’, in J.Pind
and E.Rögnvaldsson (eds) Papers from the Seventh Scandinavian Conference of
Computational Linguistics (Reykjavík 198 9), Reykjavik: Institute of
Lexicography/Institute of Linguistics, 138–45.
Denisov, P.N. (1965) Principy modelirovanija jazyka (na materiale vspomogatelaych
jazykov dlja avtomaticeskogo poiska i perevoda), Moscow: Moskovskij
Universitet.
Dietze, J. (1986) ‘Projekt der rechnergestützten Erarbeitung eines
Wörterbuchs von Esperantowurzeln’, Wissenschaftliche Zeitschrift der
Universität Halle, 35 [G, 5]: 90–1.
Dulicenko, A.D. (1989) ‘Ethnic language and planned language, and its
place in language science’, in K.Schubert (ed.) Interlinguistics. Aspects of
the science of planned languages, Berlin/New York: Mouton de Gruyter,
47–61.
Forster, P.G. (1987) ‘Some social sources of resistance to Esperanto’, in José-
Luis Melena Jiménez et al. (eds) Serta gratvlatoria in honorem Juan Régulo,
vol. 2: Esperantismo, La Laguna: Universidad, 203–11.
Gordos, G. (1985) ‘Parol-sintezo kun limigita vortaro’, in I.Koutny (ed.)
Perkomputila tekstoprilaboro, Budapest: Scienca Eldona Centro, 11–29.
Hjelmslev, L. (1963) Sproget, 2nd edn, Copenhagen: Berlingske forlag.
Hutchins, W.J. (1986) Machine Translation: Past, Present, Future, Chichester:
Ellis Horwood.
Janot-Giorgetti, M.T. (1985) ‘Parol-rekono kun limigita vortaro, gia apliko
en la lernado de parolataj lingvoj’, in I.Koutny (ed.) Perkomputila
tekstoprilaboro, Budapest: Scienca Eldona Centro, 57–68.
94 Klaus Schubert

Kat, J.O. de (1985) ‘Traduko el la internacia lingvo al naciaj’, in I. Koutny

(ed.) Perkomputila tekstoprilaboro, Budapest: Scienca Eldona Centro, 259–
66.
Katumori, H. and Hukuda, M. (1984) ‘Esuperanto o tyûkai-gengo to suru
kikai-hon’yaku no kokoromi’, WGNL meeting 1984–7–26.
Kelly, I.D.K. (1978) ‘How Esperanto can aid machine translation’, Computer
Weekly 19 January 1978.
Koutny, I., Olaszy, G. and Kisfaludy, K. (1988) ‘Esperanto speech synthesis
and its application in language learning’, in T.Szende (ed.) From Phonology
to Applied Phonetics, Magyar Fonetikai Füzutek// Human Papers in
Phonetics, 19, Budapest: Linguistics Institute of the Hungarian Academy
of Sciences, 47–54.
Li Wei (1986) ‘Automata tradukado el la internacia en la cinan kaj anglan
lingvojn (Esperanta-angla/cina masintraduka sistemo)’, Grundlagenstudien
aus Kybernetik und Geisteswissenschaft/Humankybernetik 27:147–52.
Maas, H.D. (1982) ‘Automata tradukado en kaj el Esperanto’, in H.F.
Yashovardhan and B.Frank-Böhringer (eds) Lingvokibernetiko kaj aliaj
internacilingvaj aktoj de la IX-a Internacia Kongreso de Kibernetiko//
Sprachkybernetik und andere internationalsprachige Akten vom IX. Internationalen
Kybernetikerkongreß, Tübingen: Narr, 75–81.
——(1985) ‘Pri kelkaj strategioj por fraz-analizo’, in I.Koutny (ed.)
Perkomputila tekstoprilaboro, Budapest: Scienca Eldona Centro, 175–205.
——(1987) ‘The MT system SUSY’, in M.King (ed.) Machine Translation Today:
the State of the Art, Edinburgh: Edinburgh University Press, 209–46, 392–
435.
Makino, S., Hirata, M. and Katumori, H. (1986) ‘Esuperanto o tyûkaigengo
to suru kikai hon’yaku’, WGNL meeting 1986–3–28.
Mohai, L. (1986) ‘Komputila tradukprogramo el Esperanto hungaren’, in
P.Broczkó, I.Koutny and A.Lukács (eds) Language Cybernetics, Educational
Cybernetics//Lingvokibernetiko, pedagogia kibernetiko//Cybernétique de la langue,
cybernétique de l’éducation, Budapest: Neumann Society, 47–51.
Papegaaij, B.C., Sadler, V. and Witkam, A.P.M. (eds) (1986) Word Expert
Semantics: an Interlingual Knowledge-based Approach (Distributed Language
Translation 1), Dordrecht/Riverton: Foris.
Piron, C. (1987) ‘Esperanto: l’immagine e la realità’, in A.Chiti-Batelli (ed.)
La comunicazione internazionale tra politica e glottodidattica, Milan: Marzorati,
68–116.
Sadler, V. (1989) Working with Analogical Semantics. Disambiguation Techniques in
DLT, Dordrecht/Providence: Foris.
——(forthcoming) ‘Machine translation project reaches watershed’, Language
Problems and Language Planning 15.
Sato, S. (1989) ‘Phonetic form generation in a semantics-to-speech system of
Japanese’, Sendai: Dissertation, Tohoku University.
Sato, S. and Kasuya, H. (1987) ‘Automatic translation/speech synthesis of
Japanese from written Esperanto incorporating a linguistic knowledge
base editor’, in J.Laver and M.A.Jack (eds) European Conference on Speech
Technology (Edinburgh 1987), vol. 2, Edinburgh: CEP, 414–17.
Schubert, K. (1986) ‘Linguistic and extra-linguistic knowledge’, Computers
and Translation 1:125–52.
 Esperanto as an intermediate language 95

——(1987) Metataxis. Contrastive Dependency Syntax for Machine Translation,

Dordrecht/Providence: Foris.
——(1988) ‘Ausdruckskraft und Regelmäßigkeit. Was Esperanto für
automatische Übersetzung geeignet macht’, Language Problems and
Language Planning 12:130–47.
——(1989a) ‘Interlinguistics—its aims, its achievements, and its place in
language science’, in K.Schubert (ed.) Interlinguistics. Aspects of the Science of
Planned Languages, Berlin/New York: Mouton de Gruyter, 7–44.
——(1989b) ‘An unplanned development in planned languages. A study of
word grammar’, in K.Schubert (ed.) Interlinguistics. Aspects of the Science of
Planned Languages, Berlin/New York: Mouton de Gruyter, 249–74.
——(1990) ‘Kunskap om världen eller kunskap om texten? En metod for
korpusstödd maskinöversättning’, in J.Pind and E.Rögnvaldsson (eds)
Papers from the Seventh Scandinavian Conference of Computational Linguistics
(Reykjav?ik 1989), Reykjavik: Institute of Lexicography/Institute of
Linguistics, 218–28.
Sherwood, B.A. (1978) ‘Fast text-to-speech algorithms for Esperanto,
Spanish, Italian, Russian and English’, International Journal of Man-Machine
Studies 10:669–92.
——(1982) ‘Raporto pri sintezo de Esperanta parolado’, in I.Koutny (ed.)
Homa lingvo kaj komputilo, Budapest: NJSzT//Societo Neumann, 64–75.
——(1985) ‘Komputila tradukado de esperanta teksto’, in I.Koutny (ed.)
Perkomputila tekstoprilaboro, Budapest: Scienca Eldona Centro, 153–60.
Sherwood, J.N. and Sherwood, B.A. (1982) ‘Computer voices and ears
furnish novel teaching options’, Speech Technology [1982]: 46–51.
Sjögren, S. (1970) En syntax för datamaskinell analys av esperanto, FOA P rapport
C 8264–11(64), Stockholm: Försvarets forskningsanstalt.
Tesnière, L. (1959) Elements de syntaxe structural, 2nd edn, 4th print. 1982,
Paris: Klincksieck.
Witkam, A.P.M. (1983) Distributed Language Translation. Feasibility Study of a
Multilingual Facility for Videotex Information Networks, Utrecht: BSO.
Chapter 7

Limitations of computers as
translation tools
Alex Gross*

As should be more than evident from other contributions to this

volume, the field of computer translation is alive and well—if anything,
it is now entering what may prove to be its truly golden era. But there
would be no need to point this out if certain problems from an earlier
time had not raised lingering doubts about the overall feasibility of the
field. Just as other authors have stressed the positive side of various
systems and approaches, this chapter will attempt to deal with some of
these doubts and questions, both as they may apply here and now to
those planning to work with computer translation systems and also in
a larger sense as they may be connected to some faulty notions about
language held by the general public and perhaps some system
developers as well. Explaining such doubts and limitations forthrightly
can only help all concerned by making clear what is likely—and what
is less likely—to work for each individual user. It can also clarify what
the underlying principles and problems in this field have been and to
some extent still are.
To begin with, the notion of computer translation is not new.
Shortly after the Second World War, at a time when no one dreamt
that word processors, spreadsheets, or drawing programs would be
widely available, some of the computer’s prime movers, Turing,
Weaver and Booth among them, were already beginning to think
about translation.1 They saw this application mainly as a natural
outgrowth of their wartime code-breaking work, which had helped
to defeat the enemy, and it never occurred to them to doubt that
computer translation was a useful and realizable goal.
The growing need to translate large bodies of technical
information, heightened by an apparent shortage of translators, was
one factor in their quest. But perhaps just as influential was a
coupling of linguistic and cultural idealism, the belief that removing

96
 Limitations of computers 97

‘language barriers’ was a good thing, something that would promote

international understanding and ensure world peace. Two related
notions were surely that deep down all human beings must be
basically similar and that piercing the superstratum of language
divisions could only be beneficial by helping people to break through
their superficial differences.2 Underlying this idealism was a further
assumption that languages were essentially some kind of code that
could be cracked, that words in one tongue could readily be replaced
by words saying the same thing in another. Just as the key to
breaking the Axis code had been found, so some sort of linguistic
key capable of unlocking the mysteries of language would soon be
discovered. All these assumptions would be sorely tested in the
decades ahead.

SOM E BASIC TERMS

Some of the most frequently used terms in this field, although also
defined elsewhere in this book, will help the reader in dealing with
the subject. It will quickly become evident that, merely by
providing these definitions, I will also have touched upon some of
the field’s major problems and limitations, which can then be
explained in greater detail. For example, a distinction is frequently
made between machine translation (usually systems that produce
rough text for a human translator to revise) and computer-assisted
translation devices (usually but not invariably software designed to
help translators do their work in an enhanced manner). These are
often abbreviated as MT and CAT respectively. So far, both
approaches require the assistance or active collaboration to one
extent or another of a live, human translator. Under machine
translation one finds a further distinction between batch,
interactive and interlingual approaches. A batch method has rules
and definitions which help it ‘decide’ on the best translation for
each word as it goes along. It prints or displays the entire text thus
created with no help from the translator (who need not even be
present but who nonetheless may often end up revising it). An
interactive system pauses to consult with the translator on various
words or asks for further clarification. This distinction is blurred
by the fact that some systems can operate in either batch or
interactive mode. The so-called interlingual approach operates on
the theory that one can devise an intermediate ‘language’—in at
least one case a form of Esperanto—that can encode sufficient
98 Alex Gross

linguistic information to serve as a universal intermediate stage—or

pivot point—enabling translation back and forth between numerous
pairs of languages, despite linguistic or cultural differences. Some
skepticism has been voiced about this approach, and to date no
viable interlingual system has been unveiled.
Batch and interactive systems are sometimes also referred to as
transfer methods to differentiate them from interlingual theories,
because they concentrate on a trade or transfer of meaning based on
an analysis of one language pair alone. I have tried to make these
distinctions as clear as possible, and they do apply to a fair extent to
the emerging PC-based scene. At the higher end on mini-and
mainframe computers, there is, however, a certain degree of overlap
between these categories, frequently making it difficult to say where
CAT ends and MT begins.
Another distinction is between pre-editing (limiting the extent of
vocabulary beforehand so as to help the computer) and postediting
(cleaning up its errors afterwards). Usually only one is necessary,
although this will depend on how perfect a translation is sought by a
specific client. ‘Pre-editing’ is also used to mean simply checking the
text to be translated beforehand so as to add new words and
expressions to the system’s dictionary. The work devoted to this type
of pre-editing can save time in post-editing later. A more extreme
form of pre-editing is known as controlled language, the severely
limited vocabulary of which is used by a few companies to make MT
as foolproof as possible.
Advocates of MT often point out that many texts do not require
perfect translations, which leads us to our next distinction, between
output intended for information-only skimming by experts able to
visualize the context and discount errors, and ‘full-dress’
translations, for those unable to do either. One term that keeps
cropping up is FAHQT for fully automatic high-quality translation,
which most in the field now concede is not possible (although the
idea keeps creeping in again through the back door in claims made
for some MT products and even some research projects).3 Closer to
current reality would be such descriptions as FALQT (fully
automatic low-quality translation) and PAMQT (partly automatic
medium-quality translation). Together, these terms cover much of
the spectrum offered by these systems.
Also often encountered in the literature are percentage claims
purportedly grading the efficiency of computer translation systems.
Thus, one language pair may be described as ‘90 percent accurate’
 Limitations of computers 99

or ‘95 percent accurate’ or occasionally only ‘80 percent accurate.’

The highest claim I have seen so far is ‘98 percent accurate.’ Such
ratings may have more to do with what one author has termed
spreading ‘innumeracy’ than with any meaningful standards of
measurement.4 On a shallow level of criticism, even if we accepted a
claim of 98 percent accuracy at face value (and even if it could be
substantiated), this would still mean that every standard double-
spaced typed page would contain five errors—potentially deep
substantive errors, as computers, barring a glitch, never make simple
mistakes in spelling or punctuation.
It is for the reader to decide whether such an error level is
tolerable in texts that may shape the cars we drive, the medicines and
chemicals we take and use, and the peace treaties that bind our
nations. As for 95 percent accuracy, this would mean one error on
every other line of a typical page, while with 90 percent accuracy we
are down to one error in every line. Translators who have had to post-
edit such texts tend to agree that with percentage claims of 90
percent or below it is easier to have a human translator start all over
again from the original text.
On a deeper level, claims of 98 percent accuracy may be even more
misleading—does such a claim, in fact, mean that the computer has
mastered 98 percent of perfectly written Engish or rather 98 percent of
minimally acceptable English? Is it possible that 98 percent of the
latter could turn out to be 49 percent of the former? There is a great
difference between the two, and so far these questions have not been
addressed. Thus, we can see how our brief summary of terms has
already given us a bird’s-eye view of our subject.

PRACTICAL LIMITATIONS
There are six important variables in any decision to use a computer
for translation: speed, subject matter, desired level of accuracy,
consistency of translation, volume and expense. These six
determinants can in some cases be merged harmoniously together in
a single task but they will at least as frequently tend to clash. Let us
take a brief look at each:

Speed
This is an area where the computer simply excels—one mainframe
system boasts 700 pages of raw output per night (while translators
100 Alex Gross

are sleeping), and other systems are equally prodigious. How raw
the output actually is—and how much post-editing will be required,
another factor of speed—will depend on how well the computer has
been primed to deal with the technical vocabulary of the text being
translated. Which brings us to our second category:

Subject matter
Here, too, the computer has an enormous advantage, provided a
great deal of work has already gone into codifying the vocabulary of
the technical field and entering it into the computer’s dictionary.
Thus, translations of aeronautical material from Russian to English
can be not only speedy but can perhaps even graze the ‘98 percent
accurate’ target, because intensive work over several decades has
gone into building up this vocabulary. If you are translating from a
field the computer vocabulary of which has not yet been developed,
you may have to devote some time to bringing its dictionaries up to
a more advanced level. Closely related to this factor is the third
category:

Desired level of accuracy

I have already mentioned the former in referring to the difference
between full-dress translations and work needed on an information-
only basis. If the latter is sufficient, only minimal post-editing—or
none at all—may be required, and considerable cash savings can be
the result. If a full-dress translation is required, however, then much
post-editing may be in order and there may turn out to be—
depending once again on the quality of the dictionaries—no
appreciable savings.

Consistency of vocabulary
Here the computer rules supreme, always assuming that correct
prerequisite dictionary building has been done. Before computer
translation was readily available, large commercial jobs with a
deadline would inevitably be farmed out in pieces to numerous
translators with perhaps something resembling a technical glossary
distributed among them. Sometimes the task of ‘standardizing’ the
final version could be placed in the hands of a single person of
dubious technical attainments. Even without the added problem of a
 Limitations of computers 101

highly technical vocabulary, it should be obvious that no two

translators can be absolutely depended upon to translate the same
text in precisely the same way. The computer can fully exorcize this
demon and ensure that a specific technical term has only one
translation, provided that the correct translation has been placed in
its dictionary (and provided, of course, that only one term with only
one translation is used for this process or entity).

Volume
From the foregoing, it should be obvious that some translation
tasks are best left to human beings. Any work of high or even
medium literary value is likely to fall into this category. But
volume, along with subject matter and accuracy, can also play a
role. Many years ago a friend of mine considered moving to
Australia, where he heard that sheep farming was quite profitable
on either a very small or a very large scale. Then he learned that
a very small scale meant from 10,000 to 20,000 head of sheep, a
very large one meant over 100,000. Anything else was a poor
prospect, and so he ended up staying at home. The numbers are
different for translation, of course, and vary from task to task and
system to system but the principle is related. In general, there will
be—all other factors being almost equal—a point at which the
physical size of a translation will play a role in reaching a
decision. Would-be users should carefully consider how all the
factors I have touched upon may affect their own needs and
intentions. Thus, the size and scope of a job can also determine
whether or not you may be better off using a computer alone,
some computer—human combination or having human
translators handle it for you from the start. One author proposes
8,000 pages per year in a single technical specialty with a fairly
standardized vocabulary as minimum requirements for
translating text on a mainframe system. 5

Expense
Given the computer’s enormous speed and its virtually foolproof
vocabulary safeguards, one would expect it to be a clear winner in
this area. But for all the reasons I have already mentioned, this is
by no means true in all cases. The last word is far from having been
written here, and one of the oldest French companies in this field
102 Alex Gross

has just recently got round to ordering exhaustive tests comparing

the expenses of computer and human translation, taking all factors
into account.6
As we can see quite plainly, a number of complications and
limitations are already evident. Speed, wordage, expense, subject
matter and accuracy/consistency of vocabulary may quickly
become mutually clashing vectors affecting your plans. If you can
make allowances for all of them, then computer translation can be
of great use to you. If the decision-making process involved seems
prolonged and tortuous, it perhaps merely reflects the true state of
the art not only of computer translation but of our overall
knowledge of how language really works. At least some of the
apparent confusion about this field may be caused by a gap
between what many people believe a computer should be able to
do in this area and what it actually can do at present. What many
still believe (and have, as we shall see, continued to believe over
several decades, despite ample evidence to the contrary) is that a
computer should function as a simple black box: you enter a text in
language A on one side, and it slides out written perfectly in
language B on the other. Or, better still, you read it aloud, and it
prints or even speaks it aloud in any other language you might
desire.
This has not happened and, barring extremely unlikely
developments, will not happen in the near future, assuming our goal
is an unerringly correct and fluent translation. If we are willing to
compromise on that goal and accept less-than-perfect translations, or
wish to translate texts within a very limited subject area or otherwise
restrict the vocabulary we use, then extremely useful results are
possible. Figure 7.1 provides a rough indication of the suitability or
otherwise of various text types for MT. Some hidden expenses may
also be encountered; these can involve retraining translators to
cooperate with mainframe and minicomputers and setting up
electronic dictionaries to contain the precise vocabulary used by a
company or institution. Less expensive systems running on a PC
with built-in glossaries also require a considerable degree of
customizing to work most efficiently, as such smaller systems are far
more limited in both vocabulary and semantic resolving power than
their mainframe counterparts.
Furthermore, not all translators are, at present, prepared to make
the adjustments in their work habits needed for such systems to
work at their maximum efficiency. And even those able to handle
 Limitations of computers 103

If machine translation techniques were 100% effective in all cases, the

above graph would be a solid block of black. The extent of white space
shown reflects the current situation. The areas where these
techniques can be quite effective or marginally effective are probably
limited to the three bars on the upper left, though even here some
post-editing by a translator may be needed for the computer output to
be read by non-specialists. It is important to stress, however, that
much technical and industrial translation does fall into this area of the
graph where computer techniques can be of use. It should be clear
from the graph that wherever language is more general or literary or is
more dependent on context or real-world understanding, computer
techniques are less effective. One author recommends their use
where the annual volume of restricted vocabulary technical text for a
single language direction (e.g. French to English) is 5,000 pages or
more (Newman 1988). Statistics for this chart are based on a face-
value acceptance of accuracy claims published by MT vendors, these
figures should therefore be treated with some caution.

Figure 7.1 The effectiveness of computer translation.

Prepared by Alex and llene Gross
104 Alex Gross

the transition may not be temperamentally suited to making such

systems function at their most powerful level. All attempts to
introduce computer translation systems into the work routine
depend on some degree of adjustment by all concerned, and in many
cases such adjustment is not easy. Savings in time or money are
usually only achieved at the end of such periods. Sometimes
everyone in a company, from executives down to stock clerks, will be
obliged to change their accustomed vocabularies to some extent to
accommodate the new system.5 Such a process can on occasion
actually lead, however, to enhanced communication within a
company.

DEEPER LIMITATIONS
This section explains how changing standards in the study of
linguistics may be related to the limitations in machine translation
we see today and perhaps prefigure certain lines of development in
this field. Those interested only in the practical side should turn
immediately to p. 115.
Some practical limitations of MT and even of CAT should
already be clear enough. Less evident are the limitations in some of
the linguistic theories which have sired much of the work in this
field. On the whole, Westerners are not accustomed to believing that
problems may be insoluble and, after four decades of labor, readers
might suppose that more progress had been made in this field than
appears to be the case. To provide several examples at once, I can
remember standing for some time by the display booth of a
prominent European computer translation firm during a science
conference at MIT and listening to the comments of passers-by. I
found it dismaying to overhear the same attitudes voiced over and
over again by quite sane and reasonable representatives from
government, business and education. Most of what I heard could be
summed up as:

1 language can’t really be that complex as we all speak it;

2 language, like nature, is an alien environment which must be
conquered and tamed;
3 there has to be some simple way to cut through all the nonsense
about linguistics, syntax and semantics, and achieve instant high-
quality translation;
4 why wasn’t it all done yesterday?
 Limitations of computers 105

To understand the reasons behind these comments and why they

were phrased in this particular way—and also to understand the
deeper reasons behind the limitations of computer translation—it
may be helpful to go back to the year 1944, when the first stirrings
of current activity were little evident and another school of
linguistics ruled all but supreme. In that year Leonard Bloomfield—
one of the three deans of US linguistics along with Edward Sapir and
Benjamin Lee Whorf7—was struggling to explain a problem that
greatly perturbed him.
Bloomfield was concerned with what he called ‘secondary
responses to language.’ By these he meant the things people say
and seem to believe about language, often in an uninformed way.
He called such opinions about language ‘secondary’ to differentiate
them from the use of language in communication, which he saw as
‘primary.’ People delivering such statements, he observed, are often
remarkably alert and enthusiastic: their eyes grow bright, they tend
to repeat these opinions over and over again to anyone who will
hear and they simply will not listen—even those who, like the ones
I met at M IT, are highly trained and familiar with scientific
procedures—to informed points of view differing from their own.
They are overcome by how obvious or interesting their own ideas
seem to be.8
I would add here that what Bloomfield seems to be describing is
a set of symptoms clinically similar to some forms of hysteria. As he
put it:

It is only in recent years that I have learned to observe these

secondary…responses in anything like a systematic manner, and I
confess that I cannot explain them—that is, correlate them with
anything else. The explanation will doubtless be a matter of
psychology and sociology.
(Hockett 1970:420)

If it is indeed hysteria, as Bloomfield seems to suggest, I wonder if it

might not be triggered because some people, when their ideas about
language are questioned or merely held up for discussion, feel
themselves under attack at the very frontier of their knowledge
about reality. For many people, language is so close to what they
believe that they are no longer able to tell the difference between
reality and the language they use to describe it. It is an unsettling
experience for them, one they cannot totally handle, somewhat like
106 Alex Gross

tottering on the edge of their recognized universe. The relationship

between one’s language habits and one’s grasp of reality has not
been adequately explored, perhaps because society does not yet train
a sufficient number of bilingual, multilingual or linguistically
oriented people qualified to undertake such investigations.9
Bloomfield went even further, to define ‘tertiary responses to
language’ as innately hostile, angry or contemptuous comments
from those whose secondary responses are questioned in any serious
way. They would be simply rote answers or rote repetitions of
people’s ‘secondary’ statements whenever they were challenged on
them, as though they were not capable of reasoning any further
about them. Here he seemed to be going even further in identifying
these responses with irrational or quasi-hysterical behavior.
What was it that Bloomfield found so worrisome about such
opinions on language? Essentially he—along with Whorf and Sapir—
had spent all his life building what most people regarded as the
‘science of linguistics.’ It was a study which required extended
fieldwork and painstaking analysis of both exotic and familiar
languages before one was permitted to make any large
generalizations even about a single language, much less about
languages in general. Closely allied to the anthropology of Boas and
Malinowski, it insisted on careful and thoughtful observations and a
non-judgmental view of different cultures and their languages. It was
based on extremely high standards of training and scholarship and
could not immediately be embraced by society at large. In some
ways he and his colleagues had gone off on their own paths, and not
everyone was able to follow them. Whorf and Sapir had, in fact,
both died only a few years earlier, and Bloomfield himself would be
gone five years later. Here are a few of the ‘secondary’ statements
that deeply pained Bloomfield and his generation of linguists:

Language A is more _______ than language B

(‘logical’, ‘profound’, ‘poetic’, ‘efficient’, etc., fill in the blank
yourself).
The structure of language C proves that it is a universal language,
and everyone should learn it as a basis for studying other
languages.
Language D and language E are so closely related that all their
speakers can always easily understand each other.
Language F is extremely primitive and can only have a few
hundred words in it.
 Limitations of computers 107

Language G is demonstrably ‘better’ than languages H, J and L.

The word for ‘___________’ (choose almost any word) in language M
proves scientifically that it is a worse—better, more ‘primitive’ or
‘evolved,’ etc.—language than language N.
Any language is easy to master, once you learn the basic structure
all languages are built on.
(Bloomfield 1944:413–21)

All of these statements are almost always demonstrably false upon

closer knowledge of language and linguistics, yet such opinions
are still quite commonly heard. In this same piece Bloomfield also
voiced his sadness over continual claims that ‘pure Elizabethan
English’ was spoken in this or that region of the American South
(a social and historical impossibility—at best such dialects contain
a few archaic phrases) or boasts that the Sequoyan Indian
language was so perfect and easy to learn that all citizens of the
State of Oklahoma should study it in school (Hockett 1970:414–
16). What he found particularly disturbing was that this sort of
linguistic folklore never seemed to die out, never yielded to
scientific knowledge, simply went on and on seemingly
repropagating itself with a life of its own. Traces of it could even
be found in the work of other scholars writing about language
and linguistics.
Bloomfield’s views were very much a reflection of his time.
They stressed a relativistic view of language and culture and the
notion that languages spoken by small indigenous groups of people
had a significance comparable to that of languages spoken by much
larger populations. They willingly embraced the notion that
language, like reality itself, is a complex matrix of factors and
tended to reject simplistic generalizations of any sort about either
language or culture. Moreover, Bloomfield certainly saw his
approach as being a crucial minimum stage for building any kind
of true linguistic science.
Less than ten years after his death these ideas were replaced, also
in the name of science, by a set of different notions, which
Bloomfield would almost certainly have dismissed as ‘secondary
responses to language.’ These new observations, which shared a
certain philosophical groundwork with computational linguistics,
constitute the credo of the Chomskian approach, now accepted as
the dominant scientific view. They include the following notions:
108 Alex Gross

1 All languages are related by a ‘universal grammar’.

2 It is possible to delineate the meaning of any sentence in any
language through knowledge of its deep structure and thereby
replicate it in another language.
3 A diagram of any sentence will reveal this deep structure.
4 Any surface-level sentence in any language can easily be related to
its deep structure, and this in turn can be related to universal
grammar in a relatively straightforward manner through a set of
rules.
5 These and related statements are sufficient to describe not only
the structure of language but the entire linguistic process of
development and acculturation of infants and young children
everywhere and can thus serve as a guide to all aspects of human
language, including speech, foreign-language training and
translation.
6 The similarity of these deep- and surface-level diagrams to the
structure of computer languages, along with the purported
similarity of the human mind to a computer, may be profoundly
significant.10

These ideas are clearly not ones Bloomfield could have approved
of. They are not relativistic or cautious but universalist and all-
embracing; they do not emphasize the study of individual
languages and cultures but leap ahead into stunning
generalizations. As such, he would have considered them examples
of ‘secondary responses’ to language. In many ways they reflect the
USA of the late 1950s, a nation proud of its own new-found
dominance and convinced that its values must be more substantial
than those of ‘lesser’ peoples. Such ideas also coincide nicely with
a seemingly perennial need academia feels for theories offering a
seemingly scientific approach, suggestive diagrams, learned jargon
and a grandiose vision.
We all know that science progresses by odd fits and starts and
that the supreme doctrines of one period may become the
abandoned follies of a later one. But the turnabout we have
described is surely among the most extreme on record. It should
also be stressed that the outlook of Bloomfield, Whorf and Sapir
has never truly been disproved or rejected and still has followers
today.11 Moreover, there is little proof that these newer ideas, while
they may have been useful in describing the way children learn to
speak, have ever helped a single teacher to teach languages better
 Limitations of computers 109

or a single translator to translate more effectively. Nor has anyone

ever succeeded in truly defining ‘deep structure’ or ‘universal
grammar.’
No one can, of course, place the whole responsibility for
machine translation today on Noam Chomsky’s theories about
language—certainly his disciples and followers 12 have also played a
role, as has the overall welcome this entire complex of ideas has
received. Furthermore, their advent has certainly also coincided
with the reemergence of many other ‘secondary responses’,
including most of the comments I mentioned overhearing at MIT.
Much of the literature on machine translation has owed—and
continues to owe—a fair amount to this general approach to
linguistic theory. Overall understanding of language has certainly
not flourished in recent times, and the old wives’ tale of a single
magical language providing the key to the understanding of all
other tongues now flourishes again as a tribute both to Esperanto
and the Indian Aymara language of Peru.13 Disappointment with
computer-translation projects has also been widespread throughout
this time, and at one point even Chomsky seemingly washed his
hands of the matter, stating that, ‘as for machine translation and
related enterprises, they seemed to me pointless as well as probably
quite hopeless’ (Chomsky 1975:40).
Even such lofty notions as those favored by Turing and Weaver,
that removing ‘language barriers’ would necessarily be a good thing,
or that different languages prevent people from realizing that they
are ‘really all the same deep down,’ could turn out to be ‘secondary
responses’. It may also be that language barriers and differences have
their uses and virtues, and that enhanced linguistic skills may better
promote world peace than a campaign to destroy such differences.
But popular reseeding of such notions is, as Bloomfield foresaw,
quite insidious, and most of these ideas are still very much with us,
right along with the proof that they may be unattainable. This is
scarcely to claim that the end is near for computers as translation
tools, though it may mean that further progress along certain lines of
enquiry is unlikely.
There are probably two compelling sets of reasons why
computers can never claim the upper hand over language in all its
complexity, one rooted in the cultural side of language, the other in
considerations related to mathematics. Even if the computer were
suddenly able to communicate meaning flawlessly, it would still fall
short of what humans do with language in a number of ways. This
110 Alex Gross

(as linguists have long been aware) is because communication of

meaning is only one among many functions of language. Others
are:

1 demonstrating one’s class status to the person one is speaking or

writing to;
2 simply venting one’s emotions, with no real communication
intended;
3 establishing non-hostile intent with strangers, or simply passing
time with them;
4 telling jokes;
5 engaging in non-communication by intentional or accidental
ambiguity, sometimes also called ‘telling lies’;
6 two or more of the above (including communication) at once.

Under these circumstances it becomes very difficult to explain how a

computer can be programmed merely to recognize and distinguish
these functions in language A, much less make all the adjustments
necessary to translate them into language B. As we have seen,
computers have problems simply with the communications side, not
to mention all these other undeniable aspects of language. This
would be hard enough with written texts but with spoken or ‘live’
language, the problems become all but insurmountable.
Closely related here is a growing awareness among writers and
editors that it is virtually impossible to separate the formulation of
even the simplest sentence in any language from the audience to
whom it is addressed. Said another way, when the audience changes,
the sentence changes. Phrased even more extremely, there is no such
thing as a ‘neutral’ or ‘typical’ or ‘standard’ sentence—even the most
seemingly innocuous examples will be seen on closer examination to
be directed towards one audience or another, whether by age,
education, class, profession, size of vocabulary, etc. While those
within the target audience for any given sentence will assume its
meaning is obvious to all, those on its fringes must often make a
conscious effort to absorb it, and those outside its bounds may
understand nothing at all. This is such an everyday occurrence that
it is easy to forget how common it really is. And this, too, adds a
further set of perplexities for translators to unravel, as they must
duplicate not only the ‘meaning’ but also the specialized ‘angling’ to
an analogous audience in the new language. Perhaps the most ironic
proof of this phenomenon lies in the nature of the ‘model’ sentences
 Limitations of computers 111

chosen by transformational and computational linguists to prove

their points. Such sentences rarely reflect general usage—they are
often simply the kinds of sentences used by such specialists to
impress other specialists in the same field.
Further proof is provided here by those forms of translation often
described as ‘impossible’, even when performed by humans—stage
plays, song lyrics, advertising, newspaper headlines, titles of books
or other original works and poetry. Here it is generally conceded
that some degree of adaptation may be merged with translation.
Theatre dialogue, in particular, demands a special level of ‘fidelity.’
Sentences must be pronounceable by actors as well as literally
correct, and the emotional impact of the play must be recreated as
fully as possible. A joke in language A must also become a joke in
language B, even if it isn’t. A constantly maintained dramatic build-
up must seek its relief or ‘punch lines’ at the right moments. This
may seem far from the concerns of a publications manager anxious
to translate product documentation quickly and correctly. But in a
real sense all use of words is dependent on building towards specific
points and delivering ‘punch lines’ about how a product or process
works. The difference is one of degree, not of quality. It is difficult to
imagine how computers can begin to cope with this aspect of
translation.
Cross-cultural concerns add further levels of complexity, and no
miraculous ‘universal structure’ 14 exists for handling them.
Languages are simply not orderly restructurings of each other’s ideas
and processes, and a story I have told elsewhere15 may perhaps best
illustrate this. It relates to a real episode in my life when my wife and
I were living in Italy. At that time she did most of the shopping to
help her learn Italian, and she repeatedly came home complaining
that she couldn’t find certain cuts of meat at the butcher’s. I told her
that if she concentrated on speaking better Italian, she would
certainly find them. But she still couldn’t locate the cuts of meat she
wanted. Finally, I was forced to abandon my male presumption of
bella figura and go with her to the market, where I patiently explained
in Italian what it was we were looking for to one butcher after
another. But even together we were still not successful. What we
wanted actually turned out not to exist.
The Italians cut their meat differently from the way Americans
do. There are not only different names for the cuts but actually
different cuts as well. Their whole system is built around it—they feed
and breed their cattle differently so as to produce these cuts. So one
112 Alex Gross

might argue that the Italian steer itself is different; technically and
anatomically, it might just qualify as a different subspecies.
This notion of ‘cutting the animal differently’ or of ‘slicing reality
differently’ can turn out to be a factor in many translation problems.
It is altogether possible for whole sets of distinctions, indeed whole
ranges of psychological or even tangible realities, to vanish when
going from one language to another. Those which do not vanish
may still be mangled beyond recognition. It is this factor which poses
one of the greatest challenges even for experienced translators. It
may also place an insurmountable stumbling block in the path of
computer-translation projects, which are based on the assumption
that simple conversions of obvious meanings between languages are
readily possible.
Another cross-cultural example concerns a well-known wager AI
pioneer Marvin Minsky has made with his MIT students. Minsky
has challenged them to create a progam or device that can
unfailingly tell the difference, as humans supposedly can, between a
cat and a dog. Minsky has made many intriguing remarks on the
relation between language and reality16 but he shows in this instance
that he has unwittingly been manipulated by languageimposed
categories. The difference between a cat and a dog is by no means
obvious, and even ‘scientific’ Linnaean taxonomy may not provide
the last word. The Tzeltal Indians of Mexico’s Chiapas State in fact
classify some of our ‘cats’ in the ‘dog’ category, rabbits and squirrels
as ‘monkey,’ and a more doglike tapir as a ‘cat,’ thus proving in this
case that whole systems of animals can be sliced differently.
Qualified linguistic anthropologists have concluded that the Tzeltal
system of naming animals—making allowance for the fact that they
know only the creatures of their region—is ultimately just as useful
and informative as Linnaean latinisms and even includes
information that the latter may omit.17 Comparable examples from
the other cultures are on record.18
An especially dramatic cross-cultural example suggests that at
least part of the raging battle as to whether acupuncture and the
several other branches of Chinese medicine can qualify as
‘scientific’ springs from the linguistic shortcomings of Western
observers. The relationships concerning illness the Chinese
observe and measure are not the ones we observe, their
measurements and distinctions are not the same as ours, their
interpretation of such distinctions is quite different from ours, the
diagnosis suggested by these procedures is not the same and the
 Limitations of computers 113

treatment and interpretation of a patient’s progress can also diverge

radically from our own. Yet the whole process is perfectly logical
and consistent in its own terms and is grounded in an empirical
procedure. 15 The vocabulary is fiendishly difficult to explain to
non-specialists in this highly developed branch of the Chinese
language. No one knows how many other such instances of large
and small discontitiuities between languages and their meanings
may exist, even among more closely related tongues like French
and English, and no one can judge how great an effect such
discontinuities may have on larger relationships betwen the two
societies or even on ordinary conversations between their all-too-
human representatives.
Just as the idea that the earth might be round went against the
grain for the contemporaries of Columbus, so the notion that whole
ranges of knowledge and experience may be inexpressible as one
moves from one language to another seems equally outrageous to
many today. Such a notion, that language A cannot easily and
perfectly replicate what is said in language B, simply goes against
what most people regard as ‘common sense.’ But is such insistence
truly commensensical or merely another instance of Bloomfield’s
‘secondary responses’? Something like this question lies at the root of
the long-continuing and never fully resolved debate among linguists
concerning the so-called Whorf-Sapir hypothesis.7
Mathematical evidence suggesting that computers can never fully
overtake language is quite persuasive. It is also, in part, fairly simple
and lies in a not terribly intricate consideration of the theory of sets.
No subset can be larger than the set of which it is a part. Yet all
mathematics—and, in fact, all science and technology, as members of
a linguistics school known as Glossematics19 have argued—can be
satisfactorily identified as a subcategory, and possibly a subset, of
language. According to this reasoning, no set of its components can
ever be great enough to serve as a representation of the superset they
belong to, namely language. Allowing for the difficulties involved in
determining the members of such sets, this argument, by analogy
alone, would tend to place language and translation outside the
limits of solvable problems and consign them to the realm of the
intractable and undecidable.20
The theory of sets has further light to shed. Let us imagine all the
words of language A as comprising a single set, within which each
word is assigned a number. Now let us imagine all the words of
language B as comprising a single set, with numbers once again
114 Alex Gross

assigned to each word. We’ll call them set A and set B. If each
numbered word within set A meant exactly the same thing as each
word with the same number in set B, translation would be no
problem at all, and no professional translators would be needed.
Absolutely anyone able to read would be able to translate any text
between these two languages by looking up the numbers for the
words in the first language and then substituting the words with the
same numbers in the second language. It would not even be
necessary to know either language. And computer translation in
such a case would be incredibly easy, a mere exercise in ‘search and
replace,’ immediately putting all the people searching through books
of words and numbers out of business.
But the sad reality of the matter—and the real truth behind
machine translation efforts—is that word # 152 in language A does
not mean exactly what word #152 in language B means. In fact, you
may have to choose between words 152, 157, 478 and 1,027 to obtain
a valid translation. It may further turn out that word 152 in language
B can be translated back into language A not only as 152 but also
149, 462 and 876. In fact, word #152 in language B may turn out to
have no relation to word #152 in language A at all. This is because
forty-seven words with lower numbers in language B had meanings
that spilled over into further numbered listings. It could still be
argued that all these difficulties could be sorted out by complex trees
of search and ‘goto’ commands. But such altogether typical examples
are only the beginning of the problems faced by computational
linguists, as words are rarely used singly or in a vacuum but are
strung together in thick, clammy strings of beads according to
different rules for different languages. Each bead one uses influences
the number, shape and size of subsequent beads, so that each new
word in a language A sentence compounds the problems of
translation into language B by an extremely non-trivial factor, with a
possible final total exceeding by several orders of magnitude the
problems confronted by those who program computers for the game
of chess.
There are, of course, some real technical experts, the linguistic
equivalents of chess grandmasters, who can easily determine most of
the time what the words mean in language A and how to render
them most correctly in language B. These experts are called
translators, though thus far no one has attributed to them the power
or standing of chess masters. Another large irony: so far, the only
people who have proved capable of manipulating the extremely
 Limitations of computers 115

complex systems originally aimed at replacing translators have, in

fact, been translators.

TRANSLATORS AND MT DEVELOPERS: MUTUAL

CRITICI SM S
None of the preceding necessarily makes the outlook for machine
translation or computer-aided translation all that gloomy or
unpromising. This is because most developers in this field long ago
accepted the limitations of having to produce systems that can
perform specific tasks under specific conditions. What prospective
users must determine, as I have sought to explain, is whether those
conditions are also their conditions. Although there have been a few
complaints of misrepresentation, this is a situation most MT and
CAT developers are prepared to live with. What they are not ready
to deal with (and here let’s consider their viewpoint) is the
persistence of certain old wives’ tales about the flaws of computer
translation.
The most famous of these, they will point out with some ire, are
the ones about the expressions ‘the spirit is willing but the flesh is
weak’ or ‘out of sight, out of mind’ being run through the
computer and coming out as ‘the vodka is good but the meat is
rotten’ and ‘invisible idiot’ respectively. There is no evidence for
either anecdote, they will protest, and they may well be right.
Similar stories circulate about ‘hydraulic rams’ becoming ‘water
goats’ or the headline ‘Company posts sizeable growth’ turning
into ‘Guests mail large tumour’. Yet such resentment may be
somewhat misplaced. The point is not whether such and such a
specific mistranslation ever occurred but simply that the general
public—the same public equally prepared to believe that ‘all
languages share a universal structure’—is also ready to believe that
such mistranslations are likely to occur. In any case, these are at
worst only slightly edited versions of fairly typical MT errors—for
instance, I recently watched a highly regarded PC-based system
render a ‘dead key’ on a keyboard (touche morte) as ‘dead touch.’ I
should stress that there are perfectly valid logical and human
reasons why such errors occur, and that they are at least as often
connected to human as to computer error. There are also perfectly
reasonable human ways of dealing with the computer to avoid
many of these errors.
The point is that the public is really quite ambivalent—even
116 Alex Gross

fickle—not just about computer translation but about computers in

general, indeed about much of technology. Lacking Roman
gladiators to cheer, they will gladly applaud at the announcement
that computers have now vanquished all translation problems but
just as readily turn thumbs down on hearing tales of blatant
mistranslations. This whole ambivalence is perhaps best
demonstrated by a recent popular film where an early model of a
fully robotized policeman is brought into a posh boardroom to be
approved by captains of industry. The board chairman instructs an
impeccably clad flunky to test the robot by pointing a pistol at it.
Immediately the robot intones, ‘if you do not drop your weapon
within twenty seconds, I will take punitive measures.’ Naturally the
flunky drops his gun, only to hear, ‘if you do not drop your weapon
within ten seconds, I will take punitive measures.’ Some minutes
later they manage to usher the robot out and clean up what is left of
the flunky. Such attitudes towards all computerized products are
widespread and coexist with the knowledge of how useful computers
can be. Developers of computer translation systems should not feel
that they are being singled out for criticism.
These same developers are also quite ready to voice their own
criticisms of human translators, some of them justified. Humans
who translate, they will claim, are too inconsistent, too slow or too
idealistic and perfectionist in their goals. It is, of course, perfectly
correct that translators are often inconsistent in the words they
choose to translate a given expression. Sometimes this is inadvertent,
sometimes it is a matter of conscious choice. In many Western
languages we have been taught not to repeat the same word too
often: thus, if we say the ‘European problem’ in one sentence, we are
encouraged to say the ‘European question’ or ‘issue’ elsewhere. This
troubles some MT people, although computers could be
programmed easily enough to emulate this mannerism. We also have
many fairly similar ways of saying quite close to the same thing, and
this also impresses some MT people as a fault, mainly because it is
difficult to program for.
This whole question could lead to a prolonged and somewhat
technical discussion of ‘disambiguation,’ or how and when to
determine which of several meanings a word or phrase may have—or,
for that matter, of how a computer can determine when several
different ways of saying something may add up to much the same
thing. Although the computer can handle the latter more readily
than the former, it is perhaps best to assume that authors of texts will
 Limitations of computers 117

avoid these two extreme shoals of ‘polysemy’ and ‘polygraphy’ (or

perhaps ‘polyepeia’) and seek out the smoother sailing of more
standardized usage.
Perhaps the most impressive experiments on how imperfect
translation can become were carried out by the French several
decades ago. A group of competent French and English translators
and writers gathered together and translated various brief literary
passages back and forth between the two languages a number of
times. The final results of such a process bore almost no resemblance
to the original, much like the game played by children sitting in a
circle, each one whispering words just heard to the neighbour on the
right (Vinay and Darbelnet 1977:195–6). Here, too, the final result
bears little resemblance to the original words.
The criticisms of slowness and perfectionism/idealism are
related to some extent. While the giant computers used by the CIA
and NSA can, of course, spew out raw translation at a prodigious
rate, this is our old friend fully automatic low-quality output and
must be edited to be clear to any but an expert in that specialty.
There is, at present, no evidence suggesting that a computer can
turn out high-quality text at a rate faster than a human—indeed,
humans may in some cases be faster than a computer, if FAHQT is
the goal. The claim is heard in some MT circles that human
translators can only handle 200 to 500 words per hour, which is
often true, but some fully trained translators can do far better. I
know of many translators who can handle from 800 to 1,000 words
per hour (something I can manage under certain circumstances
with certain texts) and have personally witnessed one such
translator use a dictating machine to produce between 3,000 and
4,000 words per hour (which, of course, then had to be fed to
typists).
Human ignorance—not just about computers but about how
languages really work—creeps in here again. Many translators report
that their non-translating colleagues believe it should be perfectly
possible for a translator simply to look at a document in language A
and ‘just type it out’ in flawless language B as quickly as though it
were the first language. If human beings could to this, then there
might be some hope of computers doing it. Here again we have an
example of Bloomfield’s secondary responses to language: the
absolute certainty that any text in one language is exactly the same
in another, give or take some minimal word juggling. There will be
no general clarity about computer translation until there is also a
118 Alex Gross

greatly enhanced general clarity about what languages are and how
they work.
In all this the translator is rarely perceived as a real person with
specific professional problems, as a writer who happens to
specialize in foreign languages. When MT systems are introduced,
the impetus is most often to retrain and/or totally reorganize the
work habits of translators or replace them with younger staff whose
work habits have not yet been formed, a practice likely to have
mixed results in terms of staff morale and competence. Another
problem, in common with word processing, is that no two MT
systems are entirely alike, and a translator trained on one system
cannot fully apply experience gained on it to another. Furthermore,
very little effort is made to persuade translators to become a factor
in their own self-improvement. Of any three translators trained on
a given system, only one at best will work to use the system to its
fullest extent and maximize what it has to offer. Doing so requires
a high degree of self-motivation and a willingness to improvise
glossary entries and macros that can speed up work. Employees
clever enough to do such things are also likely to be upwardly
mobile, which may mean soon starting the training process all over
again, possibly with someone less able. Such training also forces
translators to recognize that they are virtually wedded to creating a
system that will improve and grow over time. This is a great deal to
ask in either the U SA’s fast-food job market or Europe’s
increasingly mobile work environment. Some may feel it is a bit
like singling out translators and asking them to willingly declare
their life-long serfdom to a machine.

AND THE FUTURE?

Computer translation developers prefer to ignore many of the
limitations I have suggested, and they may yet turn out to be right.
What MT proponents never stop emphasizing is the threefold
increase in computer capacity awaiting us in the not so distant
future: increasing computer power, rapidly dwindling size and
plummeting prices. Here they are undoubtedly correct, and they
are also probably correct in pointing out the vast increase in
computer power that advanced multi-processing and parallel
processing can bring. Equally impressive are potential
improvements in the field of artificial intelligence, allowing for the
construction of far larger rule-based systems likely to be able to
 Limitations of computers 119

make complicated choices between words and expressions. 21

Neural nets, 22 along with their Hidden Markov Model cousins, 23
also loom on the horizon with their much publicized ability to
improvise decisions in the face of incomplete or inaccurate data.
Beyond that stretches the prospect of nanotechnology, 24 an
approach that will so miniaturize computer pathways as to single
out individual atoms to perform tasks now requiring an entire
circuit. All but the last are already with us, either now in use or
under study by computer companies or university research
projects. We also keep hearing advanced warnings of the imminent
Japanese wave, ready to take over at any moment and overwhelm
us with all manner of ‘voice writers’, telephone translators and
simultaneous computer interpreters.
How much of this is simply more of the same old computer
hype, with a generous helping of Bloomfield’s secondary responses
thrown in? Perhaps the case of the ‘voice writer’ can help us to
decide. This device, while not strictly a translation tool, has always
been the audio version of the translator’s black box: you say things
into the computer, and it immediately and flawlessly transcribes
your words into live on-screen sentences. In most people’s minds, it
would take just one small adjustment to turn this into a translating
device as well.
In any case, the voice writer has never materialized (and perhaps
never will) but the quest for it has now produced a new generation
of what might best be described as speaker-assisted speech-
processing systems. Though no voice writers, these systems are quite
useful and miraculous enough in their own way. As you speak into
them at a reasonable pace, they place on the screen their best guess
for each word you say, along with a menu showing the next best
guesses for that word. If the system makes a mistake, you can simply
tell it to choose another number on the menu. If none of the words
shown is yours, you still have the option of spelling it out or keying
it in. This ingenious but relatively humble device, I predict, will soon
take its place as a useful tool for some translators. This is because it
is user controlled rather than user supplanting and can help those
translators who already use dictation as their means of transcribing
text. Those who lose jobs because of it will not be translators but
typists and secretaries.
Whenever one discovers such a remarkable breakthrough as these
voice systems, one is forced to wonder if just such a breakthrough
may be in store for translation itself, whether all one’s reasons to the
120 Alex Gross

contrary may not be simply so much rationalization against the

inevitable. After due consideration, however, it still seems to me that
such a breakthrough is unlikely for two further reasons beyond those
already given. First, the very nature of this voice device shows that
translators cannot be replaced, simply because it is the speaker who
must constantly be on hand to determine whether the computer has
chosen the correct word, in this case in the speaker’s native language.
How much more necessary does it then become to have someone
authoritative nearby, in this case a translator, to ensure that the
computer chooses correctly amid all the additional choices imposed
where two languages are concerned? And second, really a more
generalized way of expressing my first point, whenever the suspicion
arises that a translation of a word, paragraph or book may be
substandard, there is only one arbiter who can decide whether this is
or is not the case: another translator.
There are no databases, no foreign language matching
programs, no knowledge-engineered expert systems sufficiently
supple and grounded in real world knowledge to take on this job.
Writers who have tried out any of the so-called ‘style-checking’
and ‘grammar-checking’ programs for their own languages have
some idea of how much useless wheel-spinning such programs
can generate for a single tongue and so can perhaps imagine what
an equivalent program for ‘translation checking’ would be like.
Perhaps such a program could work with a severely limited
vocabulary but there would be little point to it, as it would only
be measuring the accuracy of those texts computers could already
translate. Based on current standards, such programs would at
best produce verbose quantities of speculations which might
exonerate a translation from error but could not be trusted to
separate good from bad translators except in the most extreme
cases. It could end up proclaiming as many false negatives as false
positives and become enshrined as the linguistic equivalent of the
lie detector. And, if a computer cannot reliably check the fidelity
of an existing translation, how can it create a faithful translation
in the first place?
Which brings me almost to my final point: no matter what
gargantuan stores of raw computer power may lie before us, no
matter how many memory chips or AI rules or neural nets or
Hidden Markov Models or self-programming atoms we may lay end
to end in vast arrays or stack up in whatever conceivable architecture
the human mind may devise, our ultimate problem remains:
 Limitations of computers 121

1 to represent, adequately and accurately, the vast interconnections

between the words of a single language on the one hand and
reality on the other;
2 to perform the equivalent task with a second language; and
3 to map out completely and correctly, all the interconnections
between them.

This is ultimately a linguistic problem and not an electronic one at

all, and most people who take linguistics seriously have been racking
their brains over it for years without coming anywhere near a
solution.
Computers with limitless power will be able to do many things
today’s computers cannot do. They can provide terminologists
with virtually complete lists of all possible terms to use, they can
branch out into an encyclopedia of all related terms, they can
provide spot logic checking of their own reasoning processes, they
can even list the rules which guide them and cite the names of
those who devised the rules and the full text of the rules
themselves, along with extended scholarly citations proving why
they are good rules. But they cannot reliably make the correct
choice between competing terms in the great majority of cases. In
programming terms, there is no shortage of ways to input various
aspects of language nor of theories on how this should be done;
what is lacking is a coherent notion of what must be output and to
whom, of what should be the ideal ‘front end’ for a computer
translation system. Phrased more impressionistically, all these
looming new approaches to computing may promise endless
universes of artificial spiders’ webs in which to embed knowledge
about language but will the real live spiders of language—words,
meaning, trust, conflict, emotion—actually be willing to come and
live in them?
And yet, Bloomfieldian responses are heard again: there must
be some way around all these difficulties. Throughout the world,
industry must go on producing and selling; no sooner is one
model of a machine on the market than its successor is on the
way, urgently requiring translations of owners’ manuals, repair
manuals and factory manuals into a growing numb er of
languages. This is the driving engine behind computer translation
that will not stop, the belief that there must be a way to bypass,
accelerate or outwit the translation stage. If only enough studies
were made, enough money spent, perhaps a full-scale program
122 Alex Gross

like those intended to conquer space, to conquer the electron,

DNA, cancer, the oceans, volcanoes and earthquakes. Surely the
conquest of something as seemingly puny as language cannot be
beyond us. But at least one computational linguist has taken a
radically opposite stance:

A Manhattan project could produce an atomic bomb, and the

heroic efforts of the Sixties could put a man on the moon, but
even an all-out effort on the scale of these would probably not
solve the translation problem.
(Kay 1982:74)

He goes on to argue that its solution will have to be reached

incrementally if at all and specifies his own reasons for thinking this
can perhaps one day happen in at least some sense:

The only hope for a thoroughgoing solution seems to lie with

technology. But this is not to say that there is only one solution,
namely machine translation, in the classic sense of a fully
automatic procedure that carries a text from one language to
another with human intervention only in the final revision. There
is in fact a continuum of ways in which technology could be
brought to bear, with fully automatic translation at one extreme,
and word-processing equipment and dicating machines at the
other.
(Kay 1982:74)

The real truth may be far more sobering. As Bloomfield and his
contemporaries foresaw, language may be no puny afterthought of
culture, no mere envelope of experience but a major functioning
part of knowledge, culture and reality, their processes so
interpenetrating and mutually generating as to be inseparable. In a
sense humans may live in not one but two jungles, the first being
the tangible and allegedly real one with all its trials and travails.
But the second jungle is language itself, perhaps just as difficult to
deal with in its way as the first.
At this point I would like to make it abundantly clear that I am no
enemy either of computers or of computer translation. I spend
endless hours at the keyboard, am addicted to downloading all
manner of strange software from bulletin boards and have even
ventured into producing some software of my own. As I also love
 Limitations of computers 123

translation, it is natural that one of my main interests would lie at the

intersection of these two fields. Perhaps I risk hyperbole but it seems
to me that computer translation ought to rank as one of the noblest
of human undertakings, as in its broadest aspects it attempts to
understand, systematize and predict not just one aspect of life but all
human understanding itself. Measured against such a goal, even its
shortcomings have a great deal to tell us. Perhaps one day it will
succeed in such a quest and lead us all out of the jungle of language
and into some better place, although for all the reasons I have
mentioned this appears somewhat unlikely.
Despite having expressed a certain pessimism, I foresee, in fact,
a very optimistic future for those computer projects which respect
some of the reservations I have mentioned and seek limited,
reasonable goals in the service of translation. These will include
computer-aided systems with genuinely user-friendly interfaces,
batch systems which best deal with the problem of making
corrections and—for those translators who dictate their work—the
new voice-processing systems I have mentioned. There also seems
to be considerable scope for using AI to resolve ambiguities in
technical translation with a relatively limited vocabulary. Beyond
this, I am naturally describing my reactions based on a specific
moment in the development of computers and could, of course,
turn out to be quite mistaken. In a field where so many
developments move with such remarkable speed, no one can lay
claim to any real omniscience, and so I will settle at present for
guarded optimism over specific improvements, which will not be
long in overtaking us.

NOTES
* I wish to express my gratitude to the following individuals, who read
this piece in an earlier version and assisted me with their comments and
criticisms: John Báez, Professor of Mathematics, Wellesley College;
Alan Brody, computer consultant and journalist; Sandra Celt,
translator and editor; André Chassigneux, translator and Maître de
Conférences at the Sorbonne’s École Supérieure d’Interprètes et de
Traducteurs (L’ESIT); Harald Hille, English terminologist, United
Nations; Joseph Murphy, Director, Bergen Language Institute; Lisa
Raphals, computer consultant and linguist; Laurie Treuhaft, English
Translation Department, United Nations; Vieri Tucci, computer
consultant and translator; Peter Wheeler, Director, Antler Translation
Services; and Apollo Wu, reviser, Chinese Department, United
Nations.
124 Alex Gross

I would also like to extend my warmest thanks to John Newton, the

editor of this volume, for his many helpful comments.
1 In 1947 Alan Turing began work on his paper ‘Intelligent machinery,’
published the following year. Based on his wartime experience in
decoding German naval and general staff messages, this work foresaw
the use of ‘television cameras, microphones, loudspeakers, wheels and
“handling servo-mechanisms” as well as some sort of “electric brain.”’
It would be capable of:
(i) various games…
(ii) the learning of languages
(iii) translation of languages [my emphasis]
(iv) cryptography
(v) mathematics
(Hodges on Turing 1985)
Further details on Turing’s role are found in Hodges (1985). The best
overview of this entire period, as well as of the entire history of
translating computers is, of course, provided by Hutchins (1986).
2 See especially Weaver (1955).
3 Typical among these have been advertisements for Netherlands-based
Distributed Language Technology, which read in part:
DLT represents the safe route to truly automatic translation: without
assistance from bilinguals, polyglots, or post-editors. But meeting the
quality standards of professional translators—no less …. The aim is a
translation machine that understands, that knows how to tell sense from
nonsense…. In this way, DLT will surpass the limitations of formal
grammar or man-made dictionaries.
At various times during its long development, this system has boasted
the use of pre-editing, gigantic bilingual knowledge banks, an
Esperanto interlingual architecture, artificial intelligence and the ability
to handle ‘a vast range of texts on general and special subjects.’
(Source: advertisements in Language Technology/Electric Word, in most
1989 issues.) On the research side, Jaime G.Carbonell and Masaru
Tomita announced in 1987 that Carnegie Mellon University ‘has begun
a project for the development of a new generation of MT systems
whose capabilities range far beyond the current technology.’ They
further specified that, with these systems, ‘…unlike current MT
systems, no human translator should be required to check and correct
the translated text’ (Carbonell and Tomita 1987). This treatment is
found in Sergei Nirenburg’s excellent although somewhat technical
anthology (Nirenburg 1987).
4 According to an influential book in the United States, innumeracy is as
great a threat to human understanding as illiteracy (Paulos 1989).
5 Newman as included in Vasconcellos 1988a. In addition to this
excellent piece, those obtaining this volume will also want to read Jean
Datta’s candid advice on why computer translation techniques should
be introduced into a business or institution slowly and carefully, Muriel
 Limitations of computers 125

Vasconcellos’ own practical thoughts on where the field is headed and

Fred Klein’s dose of healthy skepticism.
6 ‘In the testing phase, some 5,000 pages of documentation, in three
types of text, will be processed, and the result compared with human
translation of the same text in terms of quantity, time taken, deadlines
met, and cost’, (Kingscott 1990). This piece describes the B’VITAL/
ARIANE method now being used by the French documentation giant
SITE. To my knowledge, this is the first reasonably thorough test
proposed comparing human and machine translation. Yet, it is limited
to one system in one country under conditions which, after the fact,
will most probably be challenged by one party or another. Human
translators will certainly demand to know whether full set-up costs, on-
the-job training courses and software maintenance expenses have been
fully amortized. For their part, machine translation advocates might
conceivably ask how human translators were chosen for the test and/or
what level of training was provided. These are all questions which
merit further consideration if a fair discussion comparing computer and
human translation is to take place.
7 Both Sapir and Whorf carried out extensive studies of American Indian
languages and together evolved what has come to be called the Whorf-
Sapir hypothesis. Briefly stated, this theory states that what humans
see, do and know is to a greater or lesser extent based on the structure of
their language and the categories of thought it encourages or excludes.
The prolonged and spirited debate around this hypothesis has largely
centered on the meaning of the phrase to a greater or lesser extent. Even the
theory’s most outright opponents concede it may have validity in some
cases, although they see something resembling strict determinism in
applying it too broadly and point out that translation between
languages would not be possible if the Whorf-Sapir hypothesis were
true. Defenders of the theory charge that its critics may not have
learned any one language thoroughly enough to become fully aware of
how it can hobble and limit human thinking and further reply that
some translation tasks are far more difficult than others, sometimes
bordering on the impossible.
8 Bloomfield, ‘Secondary and tertiary responses to language’ (Hockett
1970:412–29). This piece originally appeared in Language 20:45–55.
The author’s major work in the field of linguistics was Language (1933/
1984).
9 As so many people in so many countries speak two or more languages,
it might be imagined that there is a broad, widely-shared body of
accurate knowledge about such people. In point of fact there is not, and
the first reasonably accessible book-length account of this subject is by
Grosjean. Some of this book’s major points, still poorly appreciated by
society at large:
Relatively few bilingual people are able to translate between their two
languages with ease. Some who try complain of headaches, many
cannot do it at all, many others do it badly but are not aware of this.
Thus, bilingualism and translation skills are two quite
126 Alex Gross

different abilities, perhaps related to different neurological processes.

No bilinguals possess perfectly equal skills in both their languages. All
favor the one or the other at least slightly, whether in reading,
writing, or speaking. Thus, the notion of being brought up perfectly
bilingual is a myth—much of bilingualism must be actively achieved in
both languages.
One does not have to be born bilingual to qualify as such. Those who
learn a second language later, even as adults, can be considered
bilingual to some extent, provided they actively or passively use a
second language in some area of their lives.
(Grosjean 1982)

10 Although presented here in summarized form, these ideas all form part
of the well-known Chomskian process and can be found elaborated in
various stages of complexity in many works by Chomsky and his
followers (Chomsky 1957, 1965 and 1975).
11 The bloodied battlefields of past scholarly warfare waged over these
issues are easily enough uncovered. In 1968 Charles Hockett, a noted
follower of Bloomfield, launched a full-scale attack on Chomsky
(Hockett 1968). Those who wish to follow this line of debate further
can use his bibliography as a starting point. Hostilities even spilled over
into a New Yorker piece and a book of the same name (Mehta 1971).
Other starting points are the works of Chomsky’s teacher (Harris 1951)
or a unique point of view related to computer translation (Lehmann
1987). Throughout this debate, there have been those who questioned
why these transformational linguists, who claim so much knowledge of
language, should write such dense and unclear English. When
questioned on this, Mehta relates Chomsky’s reply as follows:

‘I assume that the writing in linguistics is no worse than the writing in

any other academic field’, Chomsky says. The ability to use language
well is very different from the ability to study it. Once the Slavic
department at Harvard was thinking of offering Vladimir Nabokov an
appointment. Roman Jakobson, the linguist, who was in the
department then, said that he didn’t have anything against elephants
but he wouldn’t appoint one a professor of zoology’. Chomsky laughs.
(Mehta 1971)

12 See, for example, Fodor (Fodor and Katz 1964) or Chisholm (Chisholm
1981).
13 See Note 3 for reference to Esperanto. The South American Indian
language Aymara has been proposed and partially implemented as a
basis for multilingual machine translation by the Bolivian
mathematician Iván Guzmán de Rojas, who claims that its special
syntactic and logical structures make it an ideal vehicle for such a
purpose. On a surface analysis, such a notion sounds remarkably close
to Bloomfieldian secondary responses about the ideal characteristics of
 Limitations of computers 127

the Sequoyan language, long before computers entered the picture

(Guzmán de Rojas 1985).
14 The principal work encouraging a search for ‘universal’ aspects of
language is by Greenberg. Its findings are suggestive but inconclusive
(Greenberg 1963).
15 This section first appeared in a different form as a discussion between
Sandra Celt and the author (Gelt and Gross 1987).
16 Most of Marvin Minsky’s thoughts on language follow a strictly
Chomskian framework—thus, we can perhaps refer to the overall
outlook of his school as a Minskian-Chomskian one (Minsky 1986:
Sections 19–26).
17 See Hunn for a considerably expanded treatment (Hunn 1977).
18 A rich literature expanding on this theme can be found in the
bibliography of the book mentioned in the preceding note (Hunn 1977).
19 Glossematics is, in the USA, a relatively obscure school of linguistics,
founded by two Danes, Louis Hjelmslev and Hans Jorgen Uldall,
earlier this century. Its basic thesis has much in common with thinking
about computers and their possible architectures. It starts from the
premise that any theory about language must take into account all
possible languages that have ever existed or can exist, that this is the
absolute minimum requirement for creating a science of linguistics. To
objections that this is unknowable and impossible, its proponents reply
that mathematicians regularly deal with comparable unknowables and
are still able to make meaningful generalizations about them. From this
foundation emerges the interesting speculation that linguistics as a
whole may be even larger than mathematics as a whole, and that
‘linguistics’ may not be that science which deals with language but that
the various so-called sciences with their imperfect boundaries and
distinctions may in fact be those branches of linguistics that deal for the
time being with various domains of linguistics. Out of this emerges the
corollary that taxonomy is the primary science, and that only by
naming things correctly can one hope to understand them more fully.
Goncomitant with these notions also arises an idea that ought to have
attracted computer translation researchers, that a glossematic approach
could lay down the basis for creating culture-independent maps of
words and realities through various languages, assigning precise
addresses for each ‘word’ and ‘meaning’, although it would require a
truly vast system for its completion and even then would probably only
provide lists of possible translations rather than final translated
versions. The major theoretical text of Glossematics, somewhat
difficult to follow like many linguistic source books, is Hjelmslev
(Hjelmslev 1961). One excellent brief summary in English is by
Whitfield (Whitfield 1969); another, available only in Spanish or
Swedish, is by Malmberg (Malmberg 1967).
20 Different strands of this argument may be pursued (Nagel and
Newman 1989; Harel 1987; Godel 1931).
21 In correct academic terms, artificial intelligence is not some lesser topic
related to machine translation, rather machine translation is a branch of
artificial intelligence. Some other branches are natural-language
128 Alex Gross

understanding, voice recognition, machine vision and robotics. The

successes and failures of AI constitute a very different story and a well-
publicized one at that—it can be followed in the bibliography provided
by Minsky (Minsky 1986). On AI and translation, see Wilks (Wilks
1984).
22 Neural nets are once again being promoted as a means of capturing
knowledge in electronic form, especially where language is concerned.
The source book most often cited is Parallel Distributed Processing
(Rumelhart and McClelland 1987).
23 Hidden Markov Models, considered by some as merely a different
form of neural nets but by others as a new technology in their own
right, are also being mentioned as having possibilities for machine
translation. They have, as noted, proved quite effective in facilitating
computerassisted voice transcription techniques.
24 The theory of nanotechnology visualizes a further miniaturization in
computers, similar to what took place during the movement from tubes
to chips but in this case actually using internal parts of molecules and
even atoms to store and process information. Regarded with skepticism
by some, this theory also has its fervent advocates (Drexler 1986).

BIBLIOGRAPHY
Bloomfield, L. (1933) Language, New York: Holt, Rinehart & Winston
(reprinted in great part in 1984, University of Chicago).
——(1944) ‘Secondary and tertiary responses to language’, in Language
20:45–55 and in C.F.Hockett (ed.) (1987) A Leonard Bloomfield Anthology,
Chicago: University of Chicago Press.
Booth, A.D. (ed.) (1967) Machine Translation, Amsterdam: North Holland.
Brower, R.A. (ed.) (1959) On Translation, Cambridge, Mass.: Harvard
University Press.
Carbonell, J.G. and Tomita, M. (1987) ‘Knowledge-based machine
translation, the CMU approach’, in S.Nirenburg (ed.) Machine Translation:
Theoretical and Methodological Issues, Cambridge: Cambridge University
Press.
Celt, S. and Gross, A. (1987) ‘The challenge of translating Chinese
medicine’, Language Monthly 43:19–21.
Chisholm, W.S. Jr. (1981) Elements of English Linguistics, London: Longman.
Chomsky, N. (1957) Syntactic Structures, The Hague: Mouton.
——(1965) Aspects of the Theory of Syntax, Cambridge, Mass.: MIT Press.
——(1975) The Logical Structure of Linguistic Theory, Chicago: University of
Chicago Press.
Coughlin, J. (1988) ‘Artificial intelligence and machine translation: present
developments and future prospects’, Babel 34:1, 1–9.
Datta, J. (1988) ‘MT in large organizations: revolution in the workplace’, in
M.Vasconcellos (ed.) Technology as Translation Strategy (American
Translators Association Scholarly Monog raph Series, vol. I I),
Binghamton, New York: State University of New York Press.
Drexler, E.K. (1986) Engines of Creation, New York: Anchor Press.
 Limitations of computers 129

Fodor, J.A. and Katz, J.J. (1964) The Structure of Language, New York:
Prentice-Hall.
Godel, K. (1931) ‘Uber formal unentscheidbare Sätze der Principia
Mathematica und verwandte Systeme I’, Monatshefte für Mathematik und
Physik 38:173–98.
Greenberg, J. (1963) Universals of Language, Cambridge, Mass.: MIT Press.
Grosjean, F. (1982) Life with Two Languages: an Introduction to Bilingualism,
Cambridge, Mass.: Harvard University Press.
Guzmán de Rojas, I. (1985) ‘Logical and linguistic problems of social
communication with the Aymara people’, Ottawa: The International
Development Research Center.
Harel, D. (1987) Algorithmics: The Spirit of Computing, Addison-Wesley.
Harris, Z. (1951) Structural Linguistics, Chicago: University of Chicago Press.
Hjelmslev, L. (1961) Prolegomena to a Theory of Language, Madison: University
of Wisconsin Press.
Hockett, C.F. (1968) The State of the Art, The Hague: Mouton.
——(ed.) (1987) A Leonard Bloomfield Anthology, Chicago: University of
Chicago Press.
Hodges, A. (1983) Alan Turing: The Enigma, New York: Simon & Schuster.
Hunn, E.S. (1977) Tzeltal Folk Zoology: The Classification of Discontinuities in
Nature, New York: Academic Press.
Hutchins, W.J. (1986) Machine Translation: Past, Present, Future, Chichester:
Ellis Horwood.
Jakobson, R. (1959) ‘On linguistic aspects of translation’, in R.A. Brower
(ed.) On Translation, Cambridge, Mass.: Harvard University Press.
Kay, M. (1982) ‘Machine translation’, American Journal of Computational
Linguistics, April-June, 74–8.
Kingscott, G. (1990) ‘SITE buys B’Vital: relaunch of French national
MTproject’, Language International, April.
Klein, F. (1988) ‘Factors in the evaluation of MT: a pragmatic approach’, in
M.Vasconcellos (ed.) Technology as Translation Strategy (American
Translators Association Scholarly Monog raph Series, vol. I I),
Bingghamton, New York: State of New York University Press.
Lehmann, W.P. (1987) ‘The context of machine translation’, Computers and
Translation 2.
Malmberg, B. (1967) ‘Los nuevos caminos de la lingüística’, Siglo Veintiuno,
Mexico: 154–74.
Mehta, V. (1971) John is Easy to Please, New York: Ferrar, Straus & Giroux.
Minsky, M. (1986) The Society of Mind, New York: Simon & Schuster.
Nagel, E. and Newman, J.R. (1989) Godel’s Proof, New York: New York
University Press.
Newman, P.E. (1988) ‘Information-only machine translation: a feasibility
study’, in M.Vasconcellos (ed.) Technology as Translation Strategy (American
Translators Association Scholarly Monog raph Series, vol. I I),
Binghamton, New York: State University of New York Press.
Nirenburg, S. (ed.) (1987) Machine Translation: Theoretical and Methodological
Issues, Cambridge: Cambridge University Press.
Paulos, J.A. (1989) Innumeracy, Mathematical Illiteracy and its Consequences, New
York: Hill & Wang.
130 Alex Gross

Rumelhart, D.E. and McClelland, J.L. (1987) Parallel Distributed Processing,

Cambridge, Mass.: MIT Press.
Sapir, E. (1921) Language: An Introduction to the Study of Speech, New York:
Harcourt, Brace & World.
Saussure, F. de (1916) Cours de linguistique generate, Paris: Payot.
Slocum, J. (ed.) (1988) Machine Translation Systems, Cambridge: Cambridge
University Press.
Vasconcellos, M. (ed.) (1988a) Technology as Translation Strategy (American
Translators Association Scholarly Monog raph Series, vol. I I),
Binghamton, New York: State University of New York.
——(1988b) ‘Factors in the evaluation of MT: formal vs functional
approaches’, in M.Vasconcellos (ed.) Technology and Translation Strategy
(American Translators Association Scholarly Monograph Series, vol. II),
Binghamton, New York: State University of New York.
Vinay, J.-P. and Darbelnet, J. (1977) Stylistique comparée du français et del’anglais,
Paris: Didier.
Weaver, W. (1955) ‘Translation’, in W.N.Locke and A.D.Booth (eds)
Machine Translation of Languages, New York: Wiley, 15–23.
Whitfield, F. (1969) ‘Glossematics’, in A.A.Hill (ed.) Linguistics, Voice of
America Forum Lectures.
Whorf, B.L. (1956) Language, Thought and Reality (collected papers),
Cambridge, Mass.: MIT Press.
Wilks, Y. (1984) ‘Machine translation and the artificial intelligence
paradigm of language processes’, Computers in Language Research 2.
Chapter 8

Computerized term banks and

translation
Patricia Thomas

Advances in the coding of messages in the Second World War and

the application of computers to areas other than mathematics led to
the notion that translation could be done ‘automatically’. Early
computing research workers who were not linguists tended to think
that translation could be effected by substituting a word in one
language for its lexical equivalent in another and, while morphology
was addressed, syntax and semantics were hardly considered; the
frequently hilarious results have been well documented (Knowles
1979, inter alia). Two more or less parallel developments to produce
computerized aids for translators began to take shape: the creation of
machine translation systems in the late 1940s and the building of
computerized dictionaries in the form of multilingual terminology
data banks (term banks) in the early 1960s. Initially these
developments took place independently but in recent years there has
been a merging of the two forms, in which the terminographical
information in a term bank may be incorporated into machine
translation systems in the form of ‘back-up’ dictionaries. Latterly, the
information has been used to provide the data for expert systems
which may help, for example, with the semantic problems in
machine translation.

TERM BANKS
Generally, little is known about term banks as, apart from one or
two, they have not received the same press as machine translation
(MT). Why is this? There seem to be three main reasons: first, it is
only now becoming possible to buy a term bank ‘off the shelf as one
might a personal computer (PC) version of an MT system. Second,

131
132 Patricia Thomas

many are ‘in-house’ developments which are only available to

specific users. Third, there seems to be reluctance on the part of the
general public, at least in the UK, to explore the possibilities
available to them from, for example, British Telecom via a telephone
and a modem.
What sort of help can term banks provide? The principal
functions of term banks are the storage of terms in large numbers,
ease of updating, rapid retrieval and, probably most important, their
standardization or indication of preferred usage. They may provide
domain classification, relationships with other terms, definitions,
examples of terms in context, bibliographic references for further
information and indication of copyright.
The first term banks evolved as computerized dictionaries and
were mono- or multilingual. In 1976, a survey of scientific and
technical databases was carried out jointly by the Association for
Information Management (Aslib) in the UK and the European
Association of Scientific Information Dissemination Centres
(EUSIDIC). It emerged that 66 per cent of the databases had
English as a ‘carrier’ language; that is, they were built using English
as the base language and their query language is English, whereas
only 7 per cent had some multilingual functions; that is, with links
between the languages as opposed to plurilingual, where there is no
link between the languages (Iljon 1977). There are currently some
seventy-five term banks in operation, a number of which transcend
national boundaries (Terminological Data Banks 1989). Early
examples are the European Commission’s EURODICAUTOM, the
Federal Republic of Germany’s LEXIS at the Bundessprachenamt
and the Canadian government’s TERMIUM, the development of
which was accelerated by the adoption of the laws on bilingualism in
Canada between 1969 and 1977. Many have been developed by large
institutions such as the International Monetary Fund and the United
Nations and by international commercial enterprises to satisfy in-
house needs. In recent years, more term banks have been developed
at national levels to cover not only the specialized and mainly
commercial needs within a country, such as TERMDOK in Sweden
and NoTe-Bergen in Norway (the latter being developed for the
Norwegian North Sea oil industry), but also in response to the
requirements of increased trade between countries, necessitating
technical manuals and other documentation in the buyer’s language.
Access to term banks, which are generally available on-line, may
be gained by a subscriber paying an annual fee, plus ‘connection’
 Computerized term banks 133

time for each search; however, on-line access to

EURODICAUTOM and to TERMIUM is free. TERMIUM is
also available via an off-line print service or by telephone query;
since 1990 it has been available on Compact Disc Read-Only
Memory (CD-ROM). From the UK, access to EURODICAUTOM
is via the International Packet Switch Stream (IPSS) from British
Telecom’s PSS Dialplus.1 The user requires a telephone, a modem
and a terminal or microcomputer and needs to obtain a password
(£60 plus £15 quarterly in 1990) from British Telecom who will
provide an agreement which gives access to more than 130 networks
worldwide, including the European Commission Host Organisation
(ECHO) which offers on-line access to a range of EC-sponsored
databases. There is a ‘volume’ charge which for Europe amounted to
around 9 p per full screen of information in 1990. A quicker method
for users who already have access to a packet switch stream (PSS)
system is a telephone call to the EURODICAUTOM help-line in
Luxembourg.2 Their personnel will provide a temporary password
to enable the database to be searched until a permanent contract can
be arranged. EURODICAUTOM’s database is in French but
queries are posed in English. Answers may correspond exactly to
what has been requested, or contain several answers if the query is
for a compound term; the search mechanism has an automatic
truncation function. The content of E U RODICAUTOM is
described in the next section.
France is particularly well served via Minitel which is also
available in the rest of Europe and the USA at US$10 per hour
(1990 rate), giving access to NORMATERM, the database of the
French standards organization, to FRANTEXT, the database of the
National Institute of the French Language and to LEXICOM, the
term bank of the Audiovisual and Computerised Data Centre for
Social Communication, Analysis and Diffusion (DAICADIF). The
system is, however, slow because it was designed as a service
network and requires a separate action to be taken to work from
word processor to Minitel.
Specific domains may be requested from a number of term banks
in the form of glossaries on floppy disks. Some organizations have
adopted software and, in certain cases, terminology, from another
term bank (e.g. Ruhrgas from EURODICAUTOM; the Deutsches
Institut für Normung, DIN (the German standards institute) from
Siemens’ TEAM), yet so far there has been little attempt at
cooperation between the compilers of term banks to incorporate
134 Patricia Thomas

standardized or prescribed terminology in order to prevent

ambiguities. Any attempt at prescription carries with it inherent
problems of national pride and the changing nature of language, yet
it is desirable in technical and scientific domains where safety is often
an important factor.3
Some term banks record the standardized terminology of a
country; these are, perhaps not surprisingly, countries with a long-
standing tradition of preserving the quality of their language, such as
France with NORMATE RM at the Association Française de
Normalisation (AFNOR), Germany with DI N and the former
USSR with ASITO at the VNIIKI/GOSSTANDART (The All-
Union Research Institute). Although the storage of terminology was
a prime motive in the development of early term banks, it is clear
that standardized terminology is necessary for translators and that
term banks provide a valuable tool, giving them recourse to
previously translated texts and access to up-to-date terminology.
However, surveys of the ways translators work show that very few
use term banks, although many expressed a wish to do so (Smith
and Tyldesley 1986; Fulford et al. 1990), as they regard terminology
work as a ‘time waster’.

Term acquisition and representation: the record format

The point of departure in terminology work is the acquisition of
terms. This involves consulting expert sources to define and delimit
the chosen domain and to decide which word or words are used to
express the concepts and thus constitute the terms of that domain.
The next step is to provide a means of recording the terms and their
associated data; this is done by devising a record format for each
term. The two figures which follow give examples of record formats
of varying degrees of complexity. Figure 8.1 shows the descriptive,
linguistic and documentary information which accompanies each
term and which comprises the record format of
EURODICAUTOM, which contains some 500,000 records of
which 90,000 are acronyms and abbreviations. It is strong in the
domains of agriculture, medicine and law but, interestingly, there is
comparatively little on defence.
Figure 8.2 shows an early example of the format used at the
University of Surrey, based on that of the Danish term bank,
DANTERM. Within a domain, this format will be the same for
each term and will comprise a number of fields which may be
 Computerized term banks 135

Figure 8.1 Information provided for each term in EURODICAUTOM.

grouped into different types of data such as administrative (e.g.

date of entry, update, initials of terminologist), linguistic (e.g.
grammar), conceptual (e.g. those which help to ‘site’ the term in a
thesaurus-like structure, such as broader, narrower and related
terms, as well as synonyms and antonyms) and those which enlarge
the user’s knowledge of the field (e.g. bibliographic references); it is
possible for a database of these texts to be available for immediate
access. These categories are not exhaustive and other data types, or
fields, will be found in different term banks. TERMIUM, the
Canadian government’s term bank, contains data in English,
French, German and Spanish, with a maximum of two languages
being dealt with on a terminology record. It covers all domains,
with scientific, technical, administrative and economics being the
most actively used.
With term record formats being developed more or less
simultaneously in different term banks for disparate needs, it is
inevitable that the fields deemed necessary to provide information in
any domain should sometimes be adequate, are usually
comprehensive, but will invariably be widely divergent. However,
with the need for exchanges of data between term banks both
nationally and internationally, it is clear that a universal record
Figure 8.2 Example of a completed record for the entry term ‘measles virus’ in the Surrey term bank with British English
as the source language (SL). Not all record fields are necessarily completed. (The record format is under revision.)
138 Patricia Thomas

format is required which contains a mandatory common core of

fields, to which others may be added as required.

Exchange of data between term banks

The magnetic tape exchange format (MATER) was instigated by
DIN and developed by the International Standards Organization
Technical Committee (I SO TC 37) in conjunction with
terminologists working on term banks in many different countries
(ISO 6156, 1987). MATER is a standard terminological record
format comprising a common core of mandatory fields, and further
optional fields, which is used for exchanging, on tapes, term bank
data stored on mainframe computers by means of an input/output
facility which links the format of the individual bank to that of
MATER.
However, to answer the needs of microcomputers,
MicroMATER has been developed at Brigham Young University,
Provo, in cooperation with Infoterm, Vienna, and this format can
be adapted using various software tools (Melby forthcoming). Its
files comprise selected standard 7-bit ASCII characters to allow
transmission by electronic mail. It is designed to be flexible,
consisting of term records, each of which has an open-ended
number of fields of varying lengths. Each record begins with an
asterisk and each field name is enclosed in curly brackets. The
record is terminated by a commerical ‘at’ and an exclamation mark
(@!). Escape characters preceded by ‘@’ are used to represent
diacritics and are linked to IBM extended (8-bit) ASCII characters.
MicroMATE R, unlike MATE R which is used only for the
exchange of data, can be manipulated to provide, for example,
reversal of a bilingual glossary4 with its corresponding information;
this step, however, needs careful assessment of synonyms and
quasi-synonymous terms.

The future of term banks

It was soon realized that the potential of term banks was greater
than their use solely as a repository for terminology. By using a
relational database, links could be made between linguistic aspects
of terms such as synonyms and antonyms, and between
terminological aspects such as logical and generic relations (ISO
2788, 1974); in other words, hierarchical structures which lend
 Computerized term banks 139

themselves to representation in this type of database. Thus the

scope for retrieval of the information contained within a relational
database was greatly enhanced. Given that a number of terms may
be embedded in the definition of a single term, it became evident
that the knowledge contained within the text of the definition
needed to be structured and deployed for the purpose of
‘intelligent’ retrieval, which allows the machine to make deductions
from the information it contains and to propose options to the user.
The birth of the next generation of term banks is being witnessed.
These can provide for the data relating to a term, which have been
structured hierarchically, to be channelled directly into an expert
system which in turn will enable greater interaction between
translators and the data available in the term bank. Skuce and
Meyer discuss the evolution of term banks into multifactional
knowledge bases. These authors also state that ‘it is still extremely
rare to find in terminology banks or publications any systematic
attempts to represent the conceptual organization of a field,
whether textually or graphically’ (Skuce and Meyer 1990:188–9),
and stress the importance of good graphical representations for
terminology as ‘in our experience, the terminologists referred to
the graph constantly’.
Hence the term ‘record format’, i.e. the ‘fields’ it contains and, in
particular, the structure of the definition, assumes a vital role because
of the growing need to supply terminology for ‘treatment’ by
artificial intelligence. The current research in text analysis is
facilitated if careful consideration has been given to the input stage of
the term record.

Current developments
Since in the context of translation term banks are often linked into
machine translation systems, the current situation of both types of
help for translators will be reviewed, together with related aids.
Renewed interest in MT and particularly machine-assisted
translation (MAT) in recent years is due to a greater insistence on
the use of the mother tongue because of export marketing, with a
trend towards a lesser use of English. Industry’s need to penetrate
foreign markets has created an increasing and largely unsatisfied
demand for multilingual documentation. In addition to the
established term banks, and as an intermediate stage between these
and large MT systems, there is a flourishing growth in interactive
140 Patricia Thomas

computerized dictionary ‘tools’ of varying sophistication available

to the translator who likes to have dictionaries integrated with
word processors (McNaught 1988) and to be able to compile
personal glossaries. Most translators, in fact, limit their use of
computers to word processing, while, as already mentioned, a small
number have access to term banks. However, in addition to these
tools, a wide range of facilities is available such as access to both
general and specialized dictionaries, optical character readers
(OCR) for reading texts into a computer, external communications
(telex, fax, electronic mail), text editing (including spelling,
grammar and style checkers in various languages, and
hyphenation) and text analysis (concordancing and word frequency
programs). This list is not exhaustive, nor are all the facilities
found in all the systems; one system being developed under the
aegis of the European Commission is the translator’s workbench,
ESPRIT II Project no:2315 (Ahmad et al. 1990).
Among commercially available systems, Ericsson Vertriebs-
Partner in Stuttgart has produced INTERDOC as an electronic
language-independent support for translators, as the company
considers fully automatic systems to be too limited and electronic
support more rewarding to use. I NTE RD OC is language
independent because Ericsson does not provide terminology but it
provides font generation for the user to create non-Latin scripts and
characters. In addition to word-processing and communication
facilities, a terminologist allocates points on a scale from 0 to 9 for
close to distant relationships between terms, forming a ‘semantic net’
to help the translator identify synonymous terms in the same text,
even when these are unknown to the translator; it is, for example,
quite common to find different names being allocated to the same
chemical by different departments in the same company.
Among the most used of other commercial products are ‘INK
TextTools’ (Netherlands) which also run on I B M PCs and
compatibles, and have specialized dictionaries compiled by
translators, as well as a general dictionary, together with a text
analyser, ‘TEXAN’, and text editor, ‘TED’; ‘TERM EX’, now
known as ‘MTX’, from Linguatech (Provo, USA), which can draw
data from the Siemens TEAM (Germany) terminology data bank
and for which it is planned to do the same from
EURODICAUTOM (EC) and from Collins On-Line Electronic
Dictionaries (available on floppy disks for IBM PC XTs, ATs and
true compatibles); ‘Tron’ from Transword (Netherlands), limited to
 Computerized term banks 141

bilingual dictionaries; and S ITE (France) with P H E N IX

terminology management and AQUILA technical dictionaries.
There are drawbacks, however: for example, some systems with text
analysers only work out of English and in some cases their screen
layout and manuals leave much to be desired.
One significant development is Harrap’s CD-ROM Multilingual
Dictionary database which gives very rapid access via English to
eighteen general and specialized (scientific/technical and business)
dictionaries in twelve languages, mostly bilingual, and includes
Chinese and Japanese characters. This program is memory-
resident, requiring 170 kilobytes (Kb) of random-access memory
(RAM) in addition to the disk operating system (DOS) and word-
processing software, and a Hitachi, Philips, Sanyo, Sony or Toshiba
CD-ROM drive for interfacing to an IBM PC XT, AT, PS/2 or
true compatible.
In addition, desk-top publishing (DTP) is an innovation which
will speed translation work to a final product. One example is that
available from Arrow Technical Translations, which can function in
fourteen languages, post-translation, with a screen which operates on
the WYSIWYG (‘what you see is what you get’) principle. In the
U K, the University of Manchester Institute of Science and
Technology (UMIST) is developing a system for a user who does
not know the target language (TL) (Hutchins and Somers 1992)
and, in France, C.Boitet’s team in Grenoble aims to develop a
similar, interactive product called LIDIA (large internationalization
of documents by interacting with their authors) using HyperCard on
an Apple Macintosh to prompt the writer to ‘standardize’ input text
either on screen or using synthetic speech.5
MT research and development activity is intense in Japan,
where it is allied to research in AI and is committed to producing
massive ‘knowledge bases’. The Fifth Generation project
announced in 1982 has immense Japanese government and
industrial support, as well as individual support in the USA and
Europe, to enable Japan to keep abreast of Western developments
and to export its goods effectively in a world market. Translations
of documentation are almost entirely from Japanese into English
and vice versa. About twenty groups, some large by European
standards, have full-time researchers working on MT; the
universities involved are Kyushu, Kyoto, Osaka and Fukuoka, and
interested commercial organizations, where most research is
undertaken, include Fujitsu, Toshiba, NEC, Nippon Telegraph and
142 Patricia Thomas

Telephone, and a joint venture by Hitachi and Quick. All these

systems have dictionaries of approximately 30,000 words, typed by
students! The Toshiba system has a particularly good interface
with a bilingual word processor. A linguistic machine editor can
change sentences from the active to the passive voice and, where
ambiguities occur, alternative translations are proffered. It can run
interactively or off-line, in the latter case producing more than one
default translation for ambiguous words, terms or structures. The
Fujitsu system is less flexible but contains Japanese-German and
Japanese-Korean, using English as the interlingua. A
semanticsbased MT system, ATLAS, aiming at translating 20,000
words an hour, is under development.
The major current development in Europe is the European
Commission’s sequel to SYSTRAN, the EUROTRA multilingual
MT system for which design planning started in February 1978.
EUROTRA is based on a monolingual analysis of the source text,
with generation via multilingual transfer modules to a monolingual
generation in the target language. It is planned to be the world’s
largest MT system and is destined to handle translation in the nine
official Community languages, i.e. seventy-two language pairs.
Results from the early stages have been evaluated and, while it is
clear that an advanced MT system will not be realized by the end of
phase III of the project, Danzin et at. recommend in their assessment
report (Danzin et al. 1990) that the research in computational
linguistics (CL) be continued and that the next phase in the language
development of EU ROTRA should concentrate primarily on
monolingual applications. One of these could be the production of
terminologies.

The future of MT and term banks

What is likely to be the structure of future systems? In MT,
research is being continued in two disciplines, AI and CL. It seems
likely that a surface syntactic analysis of a source language (SL)
will be underpinned by a semantic analysis, which could be used
for comparison against prototypes in the form of ‘frames’ or
‘scripts’ in an expert system. To provide material to complete the
‘slots’ for frames and scripts, scanners or OCRs may be used to
‘read in’ texts from which an event could be inferred from partial
information given; here concordancing could play a role in the
provision of terms for the term bank which is an essential
 Computerized term banks 143

component in these operations. Clearly, the relationships in

terminology such as material, cause and function, and inheritance,
have a part to play in this process.
Japan is forging ahead to link AI and CL and so provide the third
generation of MT. The USA is advanced in AI techniques but shows
little interest in MT, whereas in Europe, Philips in Eindhoven is
probably the only enterprise engaged in long-term research; this is
based on Montague grammar

which defines a language by specifying a set of expressions and

their grammatical categories, and a set of syntactic rules
prescribing how these expressions may combine to form new
expressions and what the grammatical category of the new
expression will be.
(Hutchins 1986:287)

Research projects are delving into semantic representation using AI

techniques which may be language-independent, with translation
being one application, perhaps as a test of an AI system. The
enormous volume of technical translation handled in such large
international and multilingual organizations as the EEC illustrates
the need for improved MT systems. The techniques of AI will seek
to incorporate awareness and experience, or so-called ‘world
knowledge’. The inclusion of contrastive linguistics covering a
wider range of language types than hitherto will help in the
assessment of other possible forms of interlinguae used as
intermediaries in MT. Parallel processing techniques and neural
networks will be considered, although linguistic problems are
unlikely to be solved by speed and power. Certainly a sharing of
resources such as term banks and other lexical databases seems
desirable.
The picture would be incomplete without mention of advances
in speech synthesis; tests are being undertaken in restricted
domains such as banking, where recognition of numbers can be
made without the need for them to be voice specific. Other systems
are AIRSPEAK and SEASPEAK and, following research by E.
Johnson et at. at Wolfson College, Cambridge, POLICESPEAK is
being developed by Kent County Constabulary and British
Telecom in the U K, to cover traffic and emergencies in the
Channel Tunnel. Speech systems can also be ‘trained’ to recognize
the voice of an individual.
144 Patricia Thomas

Pocket-sized foreign-language translators aimed at travellers,

with dictionary, thesaurus-type suggestions and pronunciation
facilities, while not at present providing a useful tool for the
professional translator, nevertheless have potential because of their
compact size and improvements in storage capacity. A glimpse of
future potential in this field is provided by a hand-held English-
Japanese translating machine from Fuji-Xerox which has a
vocabulary of 30,000 words and phrases, which, when stroked
over English text at up to 20cm per second, produces Japanese
characters on the liquid crystal display (LCD). It is easy to
envisage a number of the features mentioned being incorporated
into such a convenient form which would be much easier to handle
and more portable than the average dictionary. Large systems,
because of cost and lack of portability, are not generally available
to the freelance translator who hitherto has probably not had
access to any computerized help other than a word processing
package on a personal computer with a printer attached.
What advances may be envisaged to improve the lot of the
translator? A workstation which will minimize the time spent on
library and other searches may comprise the following: a word
processor with a multi-window screen from which one or more term
banks may be consulted and access given to an MT system, with the
possibility of creating personal lexica; administrative and accounting
facilities; style, grammar and spelling checkers for text, with the
possibility of interactive prompting either on screen or via speech
synthesis; desk-top publishing for the final product; and the ability to
receive and transmit text through electronic mail networks or
facsimile transmission.

NOTES
1 Contact British Telecom UK Sales Operation, 8th Floor, Tenter House,
45 Moorfields, London EC2Y 9TH, tel: 0800 282444, telex: 8952558
NSSAL G, fax: 071 250 8343.
2 Tel: 010 352 488 041.
3 One area in which cooperation is being achieved is work on subject
classification which began in the Scandinavian countries with the
development of N ORDTE RM. Representatives include
organizations in Denmark, Finland, Iceland, Norway, Sweden,
Germany and the Netherlands. The classification system is
hierarchical, consisting of a letter plus four digits, and it is hoped that
it will be implemented within the next few years. Its importance lies
in its effectiveness in providing small, tightly defined domains which
 Computerized term banks 145

help to clarify the exact meaning of homographs which, when

appearing in more than one domain, vary semantically in each
(Thomas 1988).
4 To reverse the two parts which make up a bilingual dictionary, the
entry terms for the first part (for example, English, with equivalents in
French) are transposed so that the translation equivalents from the first
part become the entry terms for the second part (the original English
entry terms then become the translation equivalents in the French-
English part). The fact that the semantic fields of Equivalents’ drawn
from different languages rarely (if ever) coincide exactly makes this a
dubious basis for dictionary creation in anything but the most restricted
domains.
5 For a comprehensive resource guide to the then available
wordprocessing software, localized DTP, on-line multilingual term
banks and dictionaries see the special insert in LT/Electric Word, issue
13, May/June 1989.

BIBLIOGRAPHY
Ahmad, K., Fulford, H., Holmes-Higgin, P., Rogers, M. and Thomas, P.
(1990) ‘The translator’s workbench project’, in C.Picken (ed.)
Translating and the Computer 11: Preparing for the Next Decade, London:
Aslib, 9–19.
Danzin, A., Allén, S., Coltof, H., Recoque, A., Steusloff, H. and O’Leary,
M. (1990) ‘Eurotra Programme Assessment Report’, Commission of the
European Communities, March 1990.
Fulford, H., Höge, M. and Ahmad, K. (1990) ‘User requirements study’,
European Commission Esprit I I Project no. 2315, Translator’s
Workbench Project, Final Report on Workpackage 3.3.
Hutchins, W.J. (1986) Machine Translation: Past, Present, Future, Chichester:
Ellis Horwood.
Hutchins, W.J. and Somers, H.L. (1992) An Introduction to Machine Translation,
London: Academic Press.
Iljon, A. (1977) ‘Scientific and technical databases in a multilingual society’,
in Proceedings of Third European Congress on Information Systems, Munich:
Commission of the European Communities.
ISO 2788 (1974) ‘Documentation—guidelines for the establishment and
development of multilingual thesauri’.
ISO 6156 (1987) ‘Magnetic tape exchange format for terminological/
lexicographical records (MATER)’.
Knowles, F. (1979) ‘Error analysis of Systran output: a suggested criterion
for “internal” evaluation of translation quality and a possible corrective
design’, in B.M.Snell (ed.) (1979) Translation and the Computer, Amsterdam:
North-Holland, 109–33.
McNaught, J. (1988) ‘A survey of termbanks worldwide’, in C.Picken
(ed.) Translating and the Computer 9: Potential and Practice, London: Aslib,
112–29.
146 Patricia Thomas

Melby, A. (forthcoming) ‘MicroMATER: a proposed standard format for

exchanging lexical/terminological data files’, META (special issue).
Pogson, G. (1989) ‘The LT/Electric Word multilingual wordworker’s resource
guide’, LT/Electric Word 13, Amsterdam: Language Technology BV.
Skuce, D. and Meyer, I. (1990) ‘Computer-assisted concept analysis: an
essential component of a terminologist’s workstation’, in H.Czap and
W.Nedobity (eds) TKE ’90: Terminology and Knowledge Engineering,
Frankfurt: INDEKS Verlag, 187–99.
Smith, D. and Tyldesley, D. (1986) ‘Translation Practices Report’, Reading:
Digital Equipment Corporation.
‘Terminological Data Banks’ (1989) TermNet News 24, Infoterm, Vienna:
25–30.
Thomas, P. (1988) ‘Analysis of an English and French LS P: some
comparisons with English general text corpora’, ALSED-UNESCO
Newsletter 11, 1 (26), 2–10.
Chapter 9

The translator workstation

Alan Melby

A decade ago, when Martin Kay wrote of the need to develop

software for a translator workstation (Kay 1980), most translators
were either still using typewriters or, in the case of dictating
translators, their secretaries were using typewriters to transcribe
tapes. Today, however, most translators (or their secretaries) are
using word processing software, and, in addition, some translators
have even added other translator workstation functions to their
word processors. The 1990s promise widespread use of translator
workstations on new, more powerful, yet affordable
microcomputers.

PREVIEW
The functions of a translator workstation can be divided into three
levels (Melby 1982) as follows:

• level one includes word processing, telecommunications, and

terminology management;
• level two adds text analysis, automatic dictionary look-up, and
synchronized bilingual text retrieval;
• level three provides an interface to machine translation systems.

These various functions deserve a more detailed explanation, but

before elaborating, I will address the question of whether machine
translation is likely to eliminate the need for multi-level translator
workstations in the near future.

147
148 Alan Melby

FULLY AUTOMATIC MACHINE TRANSLATION VERSUS

TRANSLATOR WORKSTATIONS
The primary objection to investing in the development of translator
workstations has always been that available resources would be
better spent on the development of automatic machine translation
systems. A background assumption for this objection is that machine
translation will soon replace human translators altogether. This
chapter will challenge that assumption. If we examine the history of
machine translation since the early 1950s, we see that its success is
highly dependent on the nature of the source text and the intended
use of the translation.

Indicative machine translation

If the purpose is simply to obtain a rough indication of the source
text content, and not a careful, finished translation by human
standards, then fully automatic machine translation may be in order.
Raw, low-quality output which is not intended to be edited into a
high-quality translation is sometimes called ‘indicative’ translation.
Indicative translation has been used for many years by the US Air
Force to avoid unnecessary human translation. A scientist can use an
indicative translation to decide whether a foreign language
document is sufficiently interesting to warrant a careful human
translation. To obtain indicative translations, translator workstations
are not really needed.

Translation for publication or distribution

If, however, the purpose of the translation is to produce a high-quality
document for publication and distribution, then a translator workstation
is useful. The case of dialogue-based machine translation (such as the
translation of telephone conversations or machine translation involving
on-line interaction between a human and a computer), although very
interesting, will not be treated in this chapter.
There are many types of text translated for publication and
distribution, the extremes being sublanguage text and literary text.
A sublanguage text is restricted in several ways, including
vocabulary, syntax and universe of discourse. Perhaps the best
known example of naturally occurring sublanguage text is weather
bulletins. The M ETEO machine translation system translates
 The translator workstation 149

Canadian weather bulletins throughout the day. In this case, human

translators review the machine-translated output at a translator
workstation. This workstation needs mainly level-three functions as
only minor corrections to the raw machine-translated output are
necessary. Any true sublanguage in which there is a large, constant
flow of text is a good candidate for machine translation.
The other extreme of translation for publication and distribution
is literary text. Here there are no predefined restrictions on
vocabulary, syntax and universe of discourse. A literary text is often
filled with creative metaphors and words used in slightly ‘twisted’
ways. Far from being negative factors, freedom and flexibility give
life to the text and help stimulate the reader to read on. General
background knowledge is drawn on implicitly to avoid ambiguity.
Sometimes ambiguity is intentional, providing two levels of
understanding in the same passage.
For literary texts, both prose and poetry, a translator workstation
can be useful but only at levels one and two, and then only in a
limited way, since literary texts are not filled with standardized
terminology. Nevertheless, the word processor can be as useful as for
any writer, and the terminology management function can be used
to record notes about interesting expressions and solutions to their
translation problems. Even the text analysis function could be used
by a literary translator to look for other occurrences of an interesting
expression encountered in the source text. However, it is generally
accepted that current machine translation systems should not be
applied to literary text.
The problem, of course, is that most texts are somewhere between
the extremes of literary text and sublanguage text. We will use Snell-
Hornby’s terms of ‘general language’ and ‘special language’ (Snell-
Hornby 1988) to describe two points along the spectrum of text
types. General-language texts include newspaper articles, light fiction
and advertising texts. Special-language texts, which represent the
bulk of the translation done by professional translators, are a mixture
of the lexical flexibility found in general texts and the precise
terminology found in sublanguage texts. Special-language texts
include the vast number of scientific reports and the huge volume of
technical documentation written and translated every year. One
could predict that machine translation applied to such texts would
give mixed results, and this is indeed the case. Raw machine
translation of special-language text requires so much human revision
that it is often not worth post-editing for publication and
150 Alan Melby

distribution. Indeed, it is in the translation of special-language texts

that the translator workstation comes into its own. Level-one and
level-two functions are often very effective tools in the hands of
human translators for producing high-quality translations of
scientific or technical documents.
It would be worthwhile conducting a careful study of the
worldwide translation market in terms of the types of translation
discussed above. But, failing that, we can look at the results of many
years of development of the SYSTRAN machine translation system
in the European Community (EC). It is well known that translation
represents an important portion of the EC budget and that
SYSTRAN has been refined and tailored to the needs of the EC.
There is a motivation to use SYSTRAN when it is more efficient
than human translation; however, SYSTRAN is currently used on
fewer than 5 per cent of the texts translated by the EC. This
reticence reflects the increased difficulties involved in providing high-
quality machine translations of special-language texts as opposed to
pure sublanguage texts.
Studies of the worldwide translation market consistently estimate
that the amount of money spent annually on translations to be
published and distributed totals several thousands of millions of US
dollars. And, assuming that most of the text translated for these vast
sums of money is special-language text, as in the case of the EC, it
would be reasonable to conclude that the major portion of today’s
translation market would benefit from increased use of translator
workstations. The most important unresolved issue is whether
machine translation will soon be applicable to most special-language
texts to be translated, thus reducing the motivation to develop level-
one and level-two functions.
The claim of this chapter is that translator workstations are not
just a stopgap measure to improve translator productivity until
human translators are made superfluous by fully automatic high-
quality machine-translation systems. It is believed by some that such
systems will eventually be produced through the refinement of
current design techniques. However, research in the philosophy of
language indicates that this will not be so. Recent thinking in
philosophy and translation theory suggests that special-language
texts draw on the following two types of semantic networks: general
semantic networks, which correspond to the general and literary
texts described above, and domain-specific semantic networks,
which correspond to sublanguage texts or to standardized
 The translator workstation 151

terminology in special-language texts (Melby forthcoming). Thus,

special-language texts consist of a mixture of words linked to general
semantic networks and terms linked to domain-specific semantic
networks.
Present techniques of machine translation implicitly assume that
entire texts draw solely on domain-specific networks. Both transfer
approaches and ‘interlingual’ approaches (Hutchins 1986) are
based on the assumptions of modern, mainstream generative
grammar. They treat language as a formal system in which
meaning is computed compositionally from the bottom up (i.e. by
combining word meanings to form larger semantic units). These
word meanings are drawn from lexicons containing lists of
predefined, distinct word senses. For each of these senses, one
literal meaning is held to be more basic, while additional,
metaphorical meanings are considered to be more derivative. As
Lakoff has demonstrated so well (Lakoff 1987), mainstream
generative grammar is firmly based on objectivism, which is being
called into question in the various branches and re-evaluations of
post-modern philosophy.
It exceeds the scope of this chapter to attempt to explore current
trends in the philosophy of language. But it is appropriate to point
out that if Lakoff and others are right, there is no reason to assume
that current ‘objectivist’ techniques in machine translation will ever
provide an adequate solution for general texts or even for special-
language texts that contain a significant incidence of general
language. For predominantly general texts, entirely new techniques
will have to be developed. Some theorists say that neural networks
are the answer; others maintain that the simplifications made in
current individual simulated neurons will lead to a dead end and
will require a new start. All admit that a neural network capable of
dealing with full-blown natural language is still only a distant goal.
Clearly, at this time, there exists no foreseeable new formal
linguistic model that will allow machines to deal with general
language.
The implications of these findings with regard to the difference
between the two types of semantic networks are clear. Current
techniques of machine translation should only be applied to
sublanguage texts and special-language texts that are predominantly
based on domain networks. Instead of wasting effort applying
current techniques to texts which are too heavily tied to general
semantic networks, more funding should be provided for basic
152 Alan Melby

research in non-objectivist linguistics. How do we design machines

that can do what humans do, that is, draw on an unlimited amount
of context and background information to interpret new expressions,
such as the phrase ‘a jaw for news’ (creatively based on the standard
expression ‘a nose for news’)?
Given the relatively small amount of progress that has been
made in machine translation in the past forty years, and the
expanding market for high-quality translation for publication and
distribution, along with the relatively modest investment required
to develop effective tools for human translators, it seems entirely
reasonable to expect that substantial funding should also be made
available for the development of sophisticated level-one and level-
two functions for translator workstations. The return on
investment from such development will probably be high. Funding
should be directed at both translator workstation and machine
translation systems. However, neither is likely to provide an easy,
total solution to the problem of translating every type of text.
Experimentation will reveal which type of text is best translated
with each type of system.
The complexity of natural language and the difficulty of building
systems that deal with it effectively help me appreciate the wisdom of
some advice I received in 1972 from David G.Hays, the father of
computational linguistics. I was a ‘young turk’ out to solve the
problem of machine translation, and he told me that no matter what
syntactic-semantic model we used, a good machine translation system
would still require mostly plain hard work. At the time, his comment
irritated me because I was convinced that using the right formal
linguistic model would make the problems melt before my eyes.
Objectivism is so pervasive that we normally do not even recognize it.
Now I realize that I was so wrapped up in the assumptions of
objectivism that I could not see either the beauty or the implications of
the impossibility of distinctly listing all the senses associated with a
word of general language. A good lexicographer knows that the word
senses in a general dictionary do not represent a fixed, unchanging set
of all the possible meanings of a word carved into a marble pillar, but
rather some of the common landing points for a free-floating butterfly.

SOM E TRANSLATOR WORKSTATION FUNCTIONS

The rest of this chapter will describe level-one and level-two
functions in more detail, then explain the need for a standard
 The translator workstation 153

format to facilitate exchange of terminology files and finally

situate the translator workstation in the total document creation
process.

Word processing, hardware and operating environments:

considerations of an ‘ideal’ translator workstation
The core function of a translator workstation is word processing. As
more and more translators use word processing, they accept without
question its superiority over typing and retyping texts on a
typewriter. In addition, the use of dedicated word processors has
given way to the use of word processing software on general-purpose
microcomputers. The advantage of using such general-purpose
microcomputers is their flexibility. A dedicated unit, because of
proprietary design, normally cannot allow the user to add the other
software packages required for a higher level translator workstation.
Years ago, the word processing software available for use on stand-
alone units provided a greater number of more powerful functions.
Today, however, the word processing software available for general-
purpose microcomputers is superior.
If the translator workstation is to be a practical tool for a typical
real-life translator, it must be affordable. In the 1970s, when
translator workstations were only a pipe dream, some of us
dreamers predicted that translators would start buying translator
workstations when the cost of a complete workstation, including
the computer, a printer, the operating system and essential
applications software, became less than the cost of a modest
automobile. At that point, it was predicted, a workstation would
become a truly viable choice for individual translators. In an area
where many people get along using public transportation, it
becomes a real alternative to owning an automobile; in an area
where most people do own automobiles and many families have
two, it becomes a real alternative to a second automobile. Clearly,
with today’s microcomputers and the increased cost of
automobiles, we have arrived at that predicted day, as the cost of
workstations using both IBM personal computers and compatibles
(PCs), and even Macintoshes (Macs), is comparable to or less than
the cost of a modest automobile.
The second important consideration for a translator workstation
is the wise choice of an operating system. This choice must be based
not only on the processing capabilities of the operating system and
154 Alan Melby

its level of user-friendliness, but also on the application software that

is available under that operating system.
In 1980, the dominant operating system running on affordable
microcomputers was CP/ M for 8080-type microprocessors, an
operating system with too many inherent limitations to become a
viable foundation for translator workstation software. Xerox
Corporation was already developing experimental computers using a
graphical user interface, multiple windows, and a mouse, but these
computers were very expensive and thus beyond the reach of the
typical translator.
Then, during the 1980s, three significant events occurred in the
microcomputer world.

1 The IBM PC with its PC-DOS operating system appeared; it took

the microcomputer world by storm. More importantly, the open
architecture of the IBM PC was accepted as a standard and
gradually most other microcomputer hardware vendors began
building IBM PC compatibles, bringing the price down to an
affordable range and the power up to an acceptable level. There are
now tens of millions of IBM PCs and PC compatibles worldwide.
2 A few years after the introduction of the PC, the Apple Macintosh
brought the Xerox user-interface technology to the attention of
the general public. Although the first Macintoshes suffered from
memory and disk space limitations, like the first PCs, the current
Mac II series hardware and operating system (system 6 and
beyond) are powerful enough to support adequate translator
workstation software. Unfortunately, because the closed
architecture of the Macintosh has prevented clones from being
made, the price of a Mac II is higher than that of a comparable
PC based on an 80386-type microprocessor (or its more powerful
successors).
3 A few years after the introduction of the Mac, a similar graphical
interface was introduced for PCs, called MicroSoft Windows
(Windows). Like many new products, it was hardly usable at first,
but its later versions (version 3 and above), in conjunction with a
PC, are much more usable. Currently, a Mac, operating with the
standard Apple operating system, and a powerful PC, operating
with DOS plus Windows, are both viable, affordable ‘platforms’
for a translator workstation.

An important element of this sequence of developments is that the

 The translator workstation 155

Mac and the PC+Windows platforms each utilize a standard user-

interface style to which all application programs must conform. This
is highly important to translators because of their need for a
consistent user interface across the various functions of their
translator workstations. For a number of years, there was no
standard user interface in the world of PC-DOS text-based
programs. Because of this deficiency, the only way to get a consistent
user interface across a set of functions was for one software vendor
to write all of them. For a while, ‘integrated’ packages, those which
combined word processing, spreadsheet and other functions within
one program, appeared to be the answer in the business software
world. However, users soon realized that no one vendor could write
all of the best application software. The world of translator
workstation-related software development has also suffered from the
lack of a standard because it is too expensive to develop a word
processor from scratch that specifically meets translators’ needs. So
translators must use general-purpose word processors and adapt
them to their needs. The various graphical user interfaces (GUIs) in
use today are, fortunately, becoming more rather than less similar.
This means that the translator workstation of the 1990s will
probably use a GUI and compatible software components developed
by various vendors.

Further hardware developments

Another hardware development trend emerging during the 1980s
that affects translator work groups is the appearance of inexpensive
high-capacity disk storage, the rise of local area networks, and the
development of intelligent GUI Unix workstations.
If terminology and translation quality are to remain consistent
across various documents translated by different humans, there
needs to be some sharing of terminology throughout a work
group. This can be accomplished in various ways. Shared
information can be placed on each separate workstation using
either a magnetic or an optical disk, or the information can be
shared by connecting a workstation to a local area network or
central computer. Each method has advantages and
disadvantages. Until recently, being connected to a central
computer was far from user-friendly. Now, however,
microcomputers can be used as terminals to a central computer
while retaining a GUI interface. The main progress in this regard
156 Alan Melby

is the development of the X-windows interface for the Unix

operating system.
An important spin-off of GUI development is an increased ability
to handle multiple languages on the screen at the same time. On the
Mac, since characters are generated in software, there can be
characters from several different alphabets on the screen at the same
time. On a PC running DOS alone, it is difficult to have more than
256 different characters available at any given time, since each
character is represented as one ‘byte’, a basic unit of information
which can have only 256 different values. Extensions to this system
of using one byte for each character become very specific to one
piece of software. But with the standard operating system on the
Mac and with Windows on the PC, extended alphabets are handled
in a more global fashion.
Having established the need for translator workstations and
having examined some likely hardware/software platforms for a
translator workstation, I will now look at some of the individual
level-one functions (word processing, telecommunications and
terminology management) and level-two functions (text analysis,
automatic [dictionary] look-up and synchronized [bilingual] text
retrieval) of a translator workstation in more detail.

LEVEL-ONE FUNCTIONS

Improvements needed in word processing

Current word processing and desk-top publishing software is very

sophisticated. The first improvement that may be suggested is in the
area of spell checking. Systems that provide spell checking for
languages other than English are currently limited. However,
pressure from bilingual users of word-processing software will
probably prompt the enhancement of this type of software.
Another needed improvement in word processing is in
morphology-based search and replace. All word processors have a
search-and-replace function which allows the user to find all
instances of a character string. Sometimes the search and replace
can be made case insensitive so that it will find a string such as
‘Break’ at the beginning of a sentence, even though the search
string was ‘break’. But what if the user wants to replace all
occurrences of the verb ‘break’ with ‘fracture’? The typical search-
 The translator workstation 157

and-replace function will not find ‘broke’ because it does not know
enough about morphology to identify past, plural and other forms
of words. Again, translators would not be the only clients to use
this feature, so it is likely that normal market pressures will
eventually result in general-purpose word processors with
morphology-based search-and-replace functions becoming
available.
An interesting enhancement to word processing proposed by
Professor Gregory Shreve (personal communication) of Kent State
University is a database of prototypical texts. When translating a
patent, for example, a translator would find it helpful to have ready
access to a typical patent in the target language and country as well
as a description of the elements of such a document. We see the
beginnings of functions to support this in the style sheets available in
some word processors.
Also, SGML, a very flexible mark-up language, promises to make
document structure descriptions available in a universal form. Once
SGML document descriptions are widely adopted, it will make it
easier to transfer texts among different word processing software
packages.
SG M L, which stands for Standard Generalized Mark-up
Language, became an international standard (ISO 8879) in 1986.
The text encoding initiative (TEI) is an international effort
including representatives from the Association for Computing in
the Humanities and the Association for Computational
Linguistics. The TEI group is producing an application of SGM L
which will create a standardized system of embedded format
codes. Eventually, through end-user pressure, word processing
software packages will include utilities to import to and export
from TEI-SGML format. This will solve one of the problems
with word processing for translators today: some clients want the
final version of their translation returned to them containing the
internal format codes for one word processor while others prefer
that it contain those of another. Translators should not have to
switch back and forth between different word processors to meet
these requirements. However, lacking any other option, they
sometimes turn to utilities, either supplied with the word
processing software or by a third party, that switch between all
kinds of word processing formats. However, these utilities need to
be updated constantly for new releases of word processing
software and thus tend to be incomplete. The final solution to
158 Alan Melby

this dilemma is for each word processing vendor to support the

TEI-SGML format.

Telecommunications
Telecommunications is a basic need for most translators today. The
apparent exception is the in-house translator who is handed a
source text from someone else in the organization and who hands
the translation to someone else. But even in this situation, it is
likely that some text transfer using telecommunications takes place
either as the document travels from the technical writers to the
organization or from the translation office to an office in another
country.
It seems that mail is just not fast enough for many text deliveries
today. One reason for this is the crucial need for business to reduce
the time lag between launching a new product in one language
market and making it available in another. Corporations will
probably find the problem of reducing this interval even more
critical in the 1990s, particularly after 1992 in Europe.
The most common form of telecommunications used in the late
1980s was text transfer by means of a fax machine. As useful as fax
has become, it is only a partial solution for the translator. Text
transfer of target-language texts via fax presents two important
problems: first, a fax cannot be edited using a word processor;
second, a faxed copy of a text is not camera-ready. Perhaps the
optimum solution to this problem would be the widespread use of
file transfer by modem with automatic error detection and
correction. Fortunately, there are several such error-calculating
transmission protocols available, such as X-modem and Kermit.
Another important translation-related function of
telecommunications is the obtaining of access to remote databases,
such as bibliographic databases, for doing research. This research
may relate either to single terms or to general research about the
domains of the source text and would be carried out by locating
related or explanatory documents.

Terminology management
Terminology management is extraordinarily important to
translation of special-language texts containing many technical
terms. Terminology is standardized at multiple levels. Some terms
 The translator workstation 159

are coined by individuals working in the forefront of a branch of

technology as they describe a newly invented concept. These terms
may work their way ‘up’ to larger and larger groups of users, until
they are adopted as an international standard term defining that
concept. Others are imposed by a larger group and filter
‘downward’ to the increasingly smaller groups in the individual
organizations. Of course, standardization can also take place at any
point between the two extremes described. This process takes some
time and different terms may be used to describe the same concept
in different organizations at any one time. Translators who do
work for several different requestors cannot and should not be
expected to carry these differing translation standards for technical
terms in their heads.
Since there is another chapter devoted specifically to terminology
management, I will only mention a few enhancements that need to
be made to current terminology management software over the next
few years.
First, we must get away from the notion of externally defined
record layouts which is nearly universal in today’s sophisticated
database-management packages. This notion is fine for business data
processing, where you can pre-define what an employee record will
look like and what fields it will have. However, in lexical and
terminological data management, one record in a given electronic
glossary may have five fields and another may have fifty. No one can
dictate that all terms will have the same number of alternative forms
and the same number of notes. Similarly, no one can dictate that all
the words in a general-purpose dictionary will have the same number
of definitions. Ultimately, there will need to be a new kind of data
management tailored to the needs of handling many records of
‘structured text’ with internally defined fields. Several groups,
including one at the University of Waterloo (Canada) which is
pioneering work on software to access the new Oxford English
Dictionary, are working on this problem.
Terminology-management software will also have to be able to
associate several files containing different types of data. For
example, a translator might define a team of files which consists
of a bilingual reference, two appropriate monolingual references,
and a user-annotation file. Once the system has activated all of
these files, the user would type in a word or multi-word
expression only once and the system would be able to access all
the entries keyed to that word or expression within any of those
160 Alan Melby

reference files, passing automatically through hypertext cross-

reference links if appropriate. Perhaps the initial search would
locate the related entries corresponding to that term or expression
in the language in which it was entered. A second search for user-
selected correspondences in the target language could b e
activated either by a limited number of keystrokes or with a
single mouse click. This second search would also provide
information from all of the selected reference files that comprise
another user-selected team.
Another important feature for terminology software is the
availability of ‘filters’ and ‘views’. A filter makes only qualifying
records visible to the user, and a view makes only a selected subset
of the fields in a record visible. These functions allow the user to see
what is of most interest at the time and reduce the amount of
scrolling that needs to be done.
Also, there should be multiple indexes on a file so that a record
can be accessed rapidly in several ways. In the area of the
development of indexing programs, the problem of how to deal
effectively with words that appear in a file either dozens or hundreds
of times has not as yet been effectively solved. Current software
programs often present all of these multiple occurrences in an
unstructured list, which can easily result in information overload for
the user.
In working on these various enhancements, much important
information can be gleaned from standard relational database
systems and fourth-generation programming languages, but we
must avoid using the externally defined record layouts that are
usually associated with such database systems. Avoiding
externally defined record layouts without sacrificing the retrieval
and manipulation power of relational databases is the central
challenge for terminology-management software in the 1990s: the
trick is to provide additional power without compromising
flexibility.
Another significant challenge is where to obtain the data to put
into a bilingual terminology file. An important benefit from such
work is that the translator can become knowledgeable in a
particular domain more easily and more efficiently. A major source
would be monolingual conceptual knowledge bases in each
domain. The central obstacle to this effort is the astounding rate of
creation of new knowledge in today’s scientific and technological
world.
 The translator workstation 161

Unfortunately, there is no obvious practical solution to the

problem of how to codify knowledge more efficiently. The most
obvious ideal solution would be for technical writers to create, as
a matter of course, easily accessible concept-based terminology
files containing all of the new terms they create while writing. But
this change in conventions will likely occur very slowly. A
particular solution may be in more sophisticated text-analysis
tools, which brings us to the level-two functions of a translator
workstation.

LEVEL-TWO FUNCTIONS
Level one does not assume that the source text is available in
machine-readable form. The target text can be created on a word
processor, various terminology files can be consulted and the
translation can be sent to the requestor by electronic file transfer, all
without the source text being available to the translator in a
compatible electronic form. Although most translation produced in
the 1980s was produced from hard copy only, this situation should
change during the 1990s.
When the source text is available in compatible electronic form,
three new functions can be added to the translator workstation: text
analysis, automatic term look-up in electronic terminology files and
synchronized retrieval of source and target texts.
The distinction between level-one functions and level-two
functions is simply that level one is restricted to what can be done
when the source text is only available in hard-copy form. Level two
comprises all functions included in level one plus additional
functions for processing the source text in various ways.

Text analysis
One basic text analysis tool is a dynamic concordance system which
indexes all the words in the document and which allows the user to
request all occurrences of a word or combination of words within the
document. This type of analysis may assist in the translation of a
long document because it allows the translator to quickly see how
troublesome terms are used in various contexts throughout the
document.
Several text-analysis software packages are available commercially
for PCs. There are two basic types: one which searches a text ‘on-
162 Alan Melby

the-fly’ without any pre-processing; and one which requires a pre-

processing pass to create an index of all the words of the text (except
a few common words like ‘and’ and ‘the’, called ‘stop’ words). A
long document is probably better studied by preprocessing it during
a lunch hour or some other break and then looking up words and
combinations very quickly using the index. These rapid searches
could be done before starting translation, or, more likely, during
translation as questions arise.
The most important addition to a dynamic concordance system
would be access to various monolingual references through the
terminology-management function of level one. Using some kind
of dynamic data exchange between programs, the user should be
able to look for the occurrences of a term in the source text and
then ask the terminology-management software to find references
to that term in a reference dictionary or terminology file without
retyping the term. The problem, of course, is when the term does
not appear in the reference and is not clearly defined in the source
text. At some point, it should become the responsibility of
requestors of translation to provide a terminology file with each
source text that contains the definitions of new terms that appear in
the text as well as how new terms fit conceptually with related
terms in that domain. Until that responsibility is widely accepted,
the terminologist/translator must be content with keyword searches
in bibliographic databases and access to the minds of colleagues by
telephone, telefax or electronic mail.
One international project that could perhaps benefit
communication as much as anything else would be the creation and
maintenance of a roster of experts in many (perhaps thousands of )
domains in dozens of major languages. A pay-per-use service bureau
would accept an unknown term with a definition and specification of
domain and desired target language and consult an appropriate
expert for a translation equivalent. All communication would, of
course, be done by electronic mail.

Automatic lookup
Once terminology files are available and routines are in place to find
the basic forms of inflected words (i.e. to perform morphological
analysis), it is a straightforward matter to identify the words in a
piece of source text and automatically look them up in a terminology
file. Terms found in the file would be displayed on the screen along
 The translator workstation 163

with the corresponding target-language terms, and the human

translator could copy selected terms into the target text.

Synchronized retrieval
Another level-two spin-off of morphological analysis could be the
automatic creation of synchronized (i.e. parallel) bilingual text files.
When a document had been translated and revised, the final version,
as well as its source text, would be stored in such a way that each
unit of source text was linked to a corresponding unit of target text.
The units would generally be sentences, except in cases where one
sentence in the source text becomes two in the target text or vice
versa. The benefits of synchronized bilingual text retrieval are
manifold with appropriate software. A translator beginning a
revision of a document could automatically incorporate unmodified
units taken from a previous translation into the revision with a
minimum of effort. Such a system has been developed and marketed
by Alpnet but it requires that the text be marked and segmented by
the software and then translated segment by segment. The next
generation of such software would automatically synchronize source
and target units ‘after the fact’, along the lines of several
experimental software research projects.
Another benefit of synchronized bilingual retrieval would be the
creation of large bilingual databases of previously translated texts.
Thus, an individual requestor, service bureau or, most interestingly,
a large corporation or agency, could provide to its translators a
database (perhaps on CD-ROM) showing how various terms had
been translated in the organization’s documents over the past several
years. Not only would such information be invaluable for helping
the translator or terminoiogist choose the appropriate equivalent
terms for the purposes of the current text, it would also have the
added benefit of instantly allowing them to determine whether or
not a particular use of a term was new.

The long-range plan

The long-range plan to use translator workstations effectively
calls for requestors of translation to cooperate with
terminologists, translators and publishers in organizing the
knowledge of each domain and providing it in an accessible form
to all of the technical writers, terminologists and translators who
164 Alan Melby

handle their documents from creation to publication. The

translator workstation then becomes, unsurprisingly, an extended
version of a technical writer workstation, and the ‘extra’ work
involved pays off in more readable, more consistent and thus
higher-quality source and target language documents. The
training required to use a translator workstation effectively
should pay off in increased productivity. However, it is important
to introduce new technology sensitively and for a translator to
train on one function at a time.
The advanced translator workstation (Shreve and Vinciquerra
1990) would give the translator access to a ‘knowledge base’
integrating terminological information, encyclopedic information
(thesauri and bibliographic databases), textual information (such as
prototype texts) and strategic information (such as source-target unit-
pairs retrieved from synchronized bilingual databases).
Once the necessity for terminology work has been recognized at
the highest levels, groups of requestors and ultimately all requestors
of translation need to cooperate by making standardized terminology
available to each other, holding back only terms proprietary to each
organization. For this purpose, the Brigham Young University
Translation Research Group (c/o Alan Melby, Department of
Linguistics), in cooperation with the Terminology Committee of the
American Translators Association (ATA), Infoterm (an international
coordination organization), and the Steering Committee of the Text
Encoding Initiative is developing a universal terminological data-
exchange format called MicroMATER. (MATER is the name of
ISO standard 6156 for terminology exchange.) In this way, the
format for terminology exchange can be standardized in a format
compatible with SGML.
The translator workstation, rather than being a short-term
solution, is indeed a long-term solution to international
communication problems. Level-two functions (which include level-
one functions) will eventually become the dominant tools used for
special-language texts when high-quality translation is required. And
level-three functions (which now include the submission of texts to
machine translation systems) will become common for
predominantly domain texts. Eventually, translator workstations will
even interface with machine translation systems, providing them
with terminology information in order to avoid costly duplication of
effort, as machine translation systems also need constantly updated
bilingual term equivalents.
 The translator workstation 165

BIBLIOGRAPHY
Hutchins, W.J. (1986) Machine Translation: Past, Present, Future, Chichester,
Ellis Horwood.
Kay, M. (1980) ‘The proper place of men and machines’, in Language
Translation, Research Report, Xerox Palo Alto Research Center, Palo Alto,
California.
Lakoff, G. (1987) Women, Fire, and Dangerous Things: What Categories Reveal
about the Mind, Chicago: University of Chicago Press.
Melby, A. (1982) ‘Multi-level translation aids in a distributed system’, in
J.Horecky (ed.) Proceedings of COLlNG-82, Prague, July 1982, Amsterdam:
North-Holland.
——(forthcoming) ‘Causes and effects of partial asymmetry between semantic
networks tied to natural languages’ (a lecture series given at the Collège
de France, Paris, February, 1990), to appear in Les Cahiers de Lexicologie,
Paris.
Shreve, G.M. and Vinciquerra, K.J. (1990) ‘Hypertext knowledge-bases for
computer-assisted translation: organization and structure’, in A.L.
Wilson (ed.) Proceedings of the 31st Annual Conference of the American
Translators Association, New Orleans, October 1990: Learned Information,
Inc.
Snell-Hornby, M. (198 8) Translation Studies: An Integrated Approach,
Amsterdam: John Benjamins.
Chapter 10

SYSTRAN: It obviously works but

how much can It be improved?
Yorick Wilks

INTRODUCTION

To this day the SYSTRAN system (Toma 1976) remains the existence
proof of machine translation (MT). When people argue (as they
sometimes still do) that usable MT (of a quality that benefits a large
class of consumers) does not exist, one can simply point to the
existence of SYSTRAN and, in particular, to its twenty-year history at
the Foreign Technology Division (FTD) in Dayton, Ohio, where it
translates large numbers of Russian scientific and engineering theses
every month.
At the present time, there is a resurgence of MT research based
only on text statistics at IBM, New York, a revived technique
(Brown et al. 1990) that has attracted both interest and funding.
Their claim that MT can and should be done on the basis of
correlations of English and foreign words, established between very
large bilingual corpora, assumes that ‘symbolic,’ non-quantitative,
MT cannot do the job. The quick answer to them is again
SYSTRAN, which I B M’s ‘proportion of sentences correct’
percentage (40 per cent versus SYSTRAN’S 60–70 percent success
rate) lags far behind, with no evidence beyond hope, energy and
application of ever closing the gap. IBM’s argument simply ignores
what constitutes a refutation of their claims, namely, SYSTRAN.
A more detached observer might say of this clash of opinion that,
while SYSTRAN has twenty years of work to its name, the IBM
results would still be important even if all they could do was reach
the same levels of accuracy as SYSTRAN, simply because the IBM
procedures would be wholly automatic, requiring no linguistics,
translations, text marking, rules, dictionaries or even foreign-
language speakers.

166
 SYSTRAN 167

However, the fallacy here is that it ignores SYSTRAN’S long

history of tackling the most obvious language pairs (English-French,
English-Japanese, etc.). The purported economic gain for statistical
methods would have to be found in less common language pairs (in
terms of translation volume): but these are, almost by definition, the
very pairs for which the very large bilingual corpora the IBM
method requires will not be available. So the assumption, even if
true, will not yield much.
The IBM researchers would like to claim that their method will
eventually outperform SYSTRAN but as it is not obvious that their
method is as yet incrementally improvable as SYSTRAN has been
(and that is the core subject of this chapter), there is no reason to
believe their claims. It may well be that the I B M statistical
techniques have already done all they can, and that their success
rate (46 percent) of correctly translated sentences is simply
inadequate, remarkable though it may be given their non-standard
assumptions.
I do not wish to suggest that the only challenge to SYSTRAN in
MT comes from the use of statistical techniques. On the contrary, a
number of researchers in linguistics and artificial intelligence (AI)
continue to claim that advances in linguistics, semantics and
knowledge representation during the past decades now permit the
construction of a wholly new MT system that will be able to break
through the quality ceiling of about 65–70 percent correctly
translated sentences established by SYSTRAN. I am of that school of
thought, as well as being a participant in a new, large-scale, effort
(Wilks et al. 1990) to achieve just that result with a cooperative
alliance of three American university laboratories.
The last effort to surpass SYSTRAN on a substantial scale was
EUROTRA (Johnson et al. 1985). While one must admit that its
results after ten years have been undistinguished, this can be largely
attributed to failures in the management of an overly large-scale
international cooperative venture, and to the inappropriateness of
using classical linguistic methods, as opposed to AI or knowledge-
based methods. But as with everything else, time will tell whether
those efforts will bear fruit or not.
The purpose of this pocket history of MT is to emphasize
SYSTRAN’S central role, and to establish it as the point of reference.
Twenty years after its inception, SYSTRAN is still in the running to
become the European Community’s rough translator of
memoranda: it is being used in Luxembourg to some degree, while
168 Yorick Wilks

EU ROTRA is not. In spite of all its advantages, however,

SYSTRAN’S methods and nature are both in dispute, as is the
quality of its repeated performance evaluations. This shadow of
doubt has many sources, including the suspicious nature of many
early MT demonstrations; some doubt clings to any system, like
SYSTRAN, whose origins go back to those days.
Yet since that time, there have been many evaluations of
SYSTRAN, in the course of which it has performed well, and it is
one of these (which I carried out for the US Air Force in 1979–80)
that I wish to describe in this chapter, since I believe it to have been
of considerable importance (I have received repeated requests for it
over the years). This was not only because of continuing interest in
SYSTRAN but also because it was an evaluation that assumed a priori
that SYSTRAN performed MT at a level suitable for some classes of
customers, and sought further to investigate the issue of how far
revisions to the system done for a new type of text (in this case,
political texts as opposed to the scientific Russian texts upon which
the SYSTRAN Russian-English system had been built) transferred to
more (unseen) texts of that type.
This test answered what is to me the key question about MT
systems: how improvable are they, and what are their optimization
ceilings, when new errors introduced by revisions to the lexicon or
grammar equal or outweigh their benefits (i.e. sentences worsened
versus those improved)? In other words, it demonstrated that, while
all MT systems are certain to fail in many cases, what counts is their
flexibility and improvability. In this and other ways, I believe,
SYSTRAN has discovered answers while evolving methods more
familiar to AI than to linguistics proper. This may sound to some an
absurd remark since, by conventional standards, SYSTRAN has
neither conventional linguistic-syntactic rules, nor any explicit
representation of world knowledge of the kind found in many AI
systems. This is because, as I argued elsewhere (Wilks 1990),
SYSTRAN confirms two ‘principles’ of MT:

a any theory (however absurd) can be the basis of an MT system;

and
b MT systems, if they have substantial coverage, normally do not
operate by means of their stated principles.

While SYSTRAN translates to a reasonably high degree of

proficiency, it has no underlying theory that a theoretical linguist
 SYSTRAN 169

would acknowledge as such: hence we have (a) above. SYSTRAN

has been the subject of few published descriptions (though see Toma
1976), most of which have described it in terms of multiple passes
through texts and the extraction of phrases and clauses by a ‘partial
parser.’ But, in fact, there is good reason to believe that SYSTRAN’S
original Russian-English performance is as good as it is because of
the very large number of long word-collocations in Russian (about
300,000, together with a 300,000-word stem vocabulary): hence we
have principle (b) above.
One can view this as an early application of what in AI is now
called ‘case-based techniques’: the storage of very large numbers of
individual cases of usage. Again, partial parsing is a technique that
is constantly being rediscovered in AI (e.g. Jacobs et al. 1991), as
are techniques for controlling and keeping in use large bodies of
usable but antique software whose internal structure is no longer
modifiable. In the case of SYSTRAN, of course, it is its core
routines that are no longer modifiable but which are simply
imported en bloc when a new language pair is started. It is for this
reason (the reusability of blocks of analysis and generation
software, even if they have little function in reality) that SYSTRAN
has been described by its owners as a transfer rather than a direct
system, even though, in fact, it has no true separable transfer
lexicon for a language pair.
In conclusion, the power of SYSTRAN lies in its well-
established and relentless system of lexicon modification and
augmentation in the face of bad translation results. It is these
cycles that yield its subsequently high performance levels; the
role of the following experiments was to see how far the process
could go when SYSTRAN’S capabilities were applied to a totally
new subject area.

BACKGROUND: ASSESSMENT OF THE FTD/MT

SYSTEM S
What follows is different from previous assessments of the
SYSTRAN system in two ways:

1 the subject matter tested (political articles and books) differs

greatly from the scientific areas for which the SYSTRAN FTD/
MT system was originally designed;
2 the methodology of the present study differs from all existing
170 Yorick Wilks

studies with the exception of the Battelle Memorial Institute

report (1977), henceforth referred to as BR.

The present project’s goals in evaluating the US Foreign Technology

Division’s (FTD) version of SYSTRAN committed it to the BR
methodology, at the core of which is an important and valuable point
(I shall touch on some of BR’s shortcomings later). Given a realistic
and developing system like SYSTRAN, a sensible questionnaire
would not require an evaluator to assign an arbitrary value to a
translated sentence (such as six or seven on a scale of ten) but rather
would reduce evaluator judgments to the form:

Sentence X is better than sentence Y in terms of quality Z

where Z might be translation accuracy or some other wholly internal

quality of the target language material, such as intelligibility or
naturalness of the language.
Such a method is appropriate to a system like FTD’s SYSTRAN
which is constantly being updated and improved in the face of new
requirements (i.e. new Russian language usages). According to BR, it
is reasonable to ask an evaluator: ‘Is version X better than version Y
in terms of quality Z?’ where X and Y are two different target-
language translations of the same source language string (English,
we shall assume here, translated from Russian) and Y was produced
after the correction of errors found in a body of sentences that may
or may not have included X. If the corpus whose errors were
corrected did include X, then the improvement of Y over X is to be
expected. If X was not in the corpus update but Y improved anyway,
then that would be very interesting to us indeed, and this is the heart
of the BR method.
Our testing method had two such corpora: texts that were
updated with corrections after a ‘first run’, and whose corrections
were fed into the system’s dictionaries and rules, and those whose
improvement was not examined after the first run but after the
second run only. This latter was the control text, the most significant
for our study: if great translation improvements were subsequently
achieved in this non-updated corpus, that would indicate that a
continually expanding and updating MT system may be
economically viable.
Restricting oneself to this form of qualitative assessment, as the
BR method does, rules out the use of some of the more original
 SYSTRAN 171

methods of MT assessment, such as Carroll’s, in which an evaluator

judged whether a translation from FTD was informative by
comparing it to a normative (correct) translation (Carroll 1966). The
assumption here was that if the translation was informative
(compared to the original) it must be, to that degree, wrong.
Monolinguals could perform that function, and bilinguals could
compare the machine translation to the original source text as well.
Carroll’s method was only appropriate for relatively small text
samples for which high-quality normative translations could be
given, however. And, more importantly, his methodology also
assumed that the question, ‘How well does the FTD/MT system
translate?’ had been answered elsewhere. The present study assumes
that too and, that having satisfactorily answered that question, we
are free to discover how far updating improves the translation of
unseen sentences.
A key element in the BR methodology was the prominence given
to monolinguals. The origin of such an emphasis can be found in the
following passages from the ALPAC report.

The results from the ratings by bilinguals contribute nothing

more to the differentiation of the translations than is obtainable
with the monolinguals’ ratings…one is inclined to give more
credence to the results from the monolinguals because
monolinguals are more representative of potential users of
translations and are not influenced by knowledge of the source
language.
(ALPAC 1966:72)

This passage will undoubtedly elicit conflicting reactions in

readers: on the one hand, given that a high correlation between
fidelity and intelligibility is well established with regard to
translation, the above seems reasonable enough. And, as we shall
see, some of the bilinguals’ odd behavior in the present study can
only be explained if we, too, assume that they are adversely
influenced by their knowledge of the source text when making
certain judgments. On the other hand, it is quite risky to base one’s
judgment of MT largely on the performance of monolinguals, if
only because sentences may be intelligible and coherent without
being faithful translations, something a monolingual would have
no way of spotting.
In this study, we used three monolinguals and six bilinguals and,
172 Yorick Wilks

despite the discrepancies among the former group’s judgments, all

agreed on the improvements of the control text sentences. It turned
out that our results relied more heavily than we expected upon the
evidence provided by the bilingual evaluators. It also should be
noted that nothing in our study challenged the hypothesized
correlation between fidelity and intelligibility: the statistics for the
present study were not analyzed in such a way as to allow that to be
tested for individual sentences.

THE BATTELLE REPORT: ITS METHODOLOGY

AND SHORTCOMINGS
The BR had two objectives: it surveyed existing methods of machine
translation evaluation, and then applied a version of the updating
methodology sketched above to a sample of scientific texts. In this
section I will examine discussions of monolingual evaluation and its
relation to categories such as intelligibility.
The arguments for monolingual evaluation in the BR survey of
evaluation methods were twofold: first, that estimates of the fidelity
(correctness) of a translation strongly correlate to estimates of its
quality in monolingual judgments of the output. And, second, that a
monolingual expert can be expected to judge the overall coherence
of an output text since, as a text lengthens, the chances of its being
both coherent and incorrect approach zero.
B R counted as distinct the following three concepts that it
completely failed to distinguish: intelligibility, comprehensibility and
readability (Battelle 1977:10–11). At first glance, it might seem that
the difference between these categories is one of scale (with only
comprehensibility applying to entire texts) but the intelligibility test is
also applied by it to long sequences of output. Likewise, readability
which ‘measures the appropriate overall contextual cohesivness’ of a
text (Battelle 1977:14) has little obvious contrast to the previous two
categories. Indeed, the three separate tests given (one rating output
on a ‘clarity scale’, the second asking questions about the content of
the output and the third ‘Cloze technique’ requiring a subject to fill
in word gaps left at regular intervals in the output) could be applied
equally to any of the three concepts with no change in the results.
What is actually being discussed here are three different methods of
measuring coherence, nothing more.
Battelle also seems to miss the significance of its essentially
monolingual tests when it asserts:
 SYSTRAN 173

Although results obtained from these methods may correlate

well with quality of translation [monolingual quality assessment:
YW], many of them do not really test the correctness of
translation, the basic purpose of both an MT system and an
evaluation method.
(Battelle 1977:23)

This completely contradicts its earlier remark that, in a significant

class of cases, monolingually judged quality and correctness of
translation correlate strongly.
Despite these ambiguities, Battelle’s survey of monolingual tests
gives us a useful notion of test coherence that can be assessed by
experts ignorant of the source language. Moreover, we can be
confident that the results of such tests may well continue to correlate
strongly with bilingually assessed translation correctness, which
brings us to our adaptation of its experimental design.

METHODOLOGY OF THE PRESENT TEST

The test had the following stages:

i Text materials containing one-and-a-half million Russian words

were sent from FTD to LATSEC, the SYSTRAN company in
California.
ii Of these, 150,000 were chosen by a random procedure and
divided into two roughly equal groups of documents: the object
or update text (O) and the control text (C).
iii Both texts were keypunched and the object text translated by an
existing copy of the FTD/MT Russian-English MT system.
iv The control text was left unexamined and untouched, while
errors in the translation of the object text were analyzed at
LATSEC, and dictionary and program corrections implemented
for them. This process took four-and-a-half months, during
which 2,750 stem dictionary entries and 2,800 expression
dictionary entries were updated.
v With the update system inserted, a second version of the MT
program was created which was then run on both object and
control texts.
vi The first copy of the MT system (without the updating) was
then run on the control text, thus creating four sets of output:
first runs of object and control texts and second runs of both.
174 Yorick Wilks

vii A comparator program took the first and second runs of each text
and listed only those sentences that were changed between runs,
viii The two outputs of the comparator program (one object, one
control) were each divided into three parts,
ix Evaluators were chosen at three sites as follows:
A at Essex University, two bilinguals familiar with Rus-sian-
English translation but not with MT, plus one monolingual
with qualifications in political science, were named A1, A2,
A3, respectively;
B at FTD, three bilinguals familiar with MT were named B1,
B2, B3, respectively;
C at LATS EC, the inverse of (A) was done: one non-MT
bilingual plus two monolinguals familiar with the subject
matter were chosen, called C3, C1, C2, respectively.
Each evaluator with a given digit in their name code received
that same one-third of the change object and control text.
x Each evaluator received the same instructions and questionnaire
(see below). The sentences came to them in the form of a
Russian sentence, a first-run English sentence or a second-run
English sentence. These last two sentences were randomly
ordered, so as to avoid any assumption that the second was
‘better’. For each of three questions the evaluator was asked to
choose one of the four answers A, B, C, or D. Their choice was
indicated by circling one of the letters on a computer form
containing the number of the sentence and the letters A to D.
Answer sheets were mailed directly back to LATSEC.
xi A totalizator program compiled the results from each evaluator
for each set of texts, plus monolingual and bilingual totals and
these in turn were subjected to statistical analysis.
xii The evaluators were asked to give their reactions to the test and
questionnaire, and some were asked to review sample sentences,
answering with different choice orders, and to count the Russian
words that survived translation (for the significance of this, see
questionnaire below).

Precautions taken against bias and to ensure security of the

data
For anyone, especially those who are sceptical about MT or the
FTD/MT system in particular, it is critical to be sure that none of the
 SYSTRAN 175

updatings were performed on the control text, for the ‘unplanned’

improvement of the control text (or carry-over effect) is at the very
heart of the study. And, while no methodology is foolproof, we
believe that ours took all reasonable precautions against obvious
sources of bias in its results, and against any criticisms that must
inevitably arise about the security of the control text.

Test selection
A keypuncher at LATSEC prepared a card for each document in the
one-and-a-half million word Russian corpus. On each card was
punched the number of one of the documents and a random number
(these being taken from the standard library copy of a random
number table, starting on the first page with the first number, and
continuing in order from that page). The pack of cards prepared for
all the documents in the corpus went to Teledyne Ryan who ran it
through a standard program that sorts random numbers in
ascending order. LATSEC then keypunched the Russian documents
by taking their numbers from the ordered list provided by Teledyne
Ryan (taking the document numbers in turn which corresponded
one-to-one in that they were on the same card) to the random
numbers, now in numerical sequence.
For security, the original card pack was then sent to FTD so that
the whole procedure could be verified later with any standard
sorting program. We believe this procedure gave a random selection
of 150,000 Russian words by LATSEC keypunching down the list
until that total was reached (the first 75,000 becoming the object text
so that translation and updating could start immediately, and the
second 75,000 becoming the control text). While this method was
perhaps overly detailed, it yielded a significant sample of the corpus
by any normal statistical criteria, one which compared very well in
terms of sample size with experiments referred to in other surveys.

Anonymity and spread of evaluators

The nine evaluators were marked on the totalized output by their
code names only (A1,…C3). They communicated directly with
LATSEC, and their identities were not divulged to the project
director. As previously noted, the evaluators were not only of three
types but were from three sites: six out of nine had no previous
connection with LATSEC.
176 Yorick Wilks

Order of A, B, C, D choices and 1, 2, 3 questions

To prevent the order of questions, and answer types within
questions, influencing the evaluators (such as the suggestion that
both translations might be too poor to assess their relationship), the
evaluators at FTD answered the questions in the orders 2 1 3, 3 2 1,
3 2 1, while all other evaluators used orders 1 2 3; Essex evaluators
further redid a sample of their answers to the questions with the
choices re-ordered as BCDA.

Order of presentation of the English translations

An obvious source of bias would have been the ordering of first-and
second-run English translations in regular patterns, creating an order
bias in quality assessment. To avoid such a bias, the two translations
were presented in random order on the data sheets, and the input
answers corrected for this by the totalizator program question by
question.

Security of the control text

The first-run MT program was copied and sent to FTD so that it
could be tested later to ensure that it was in fact the program that ran
on the control text. More substantially, each sheet used by LATSEC
in updating the object text (like all the originals, these were sent to
FTD at the end of the project) was labeled with the text and sentence
number that gave rise to update for later reference.
Finally, FTD was sent a copy of all the updates made during the
project for incorporation into its system, so that it could run the
following smallscale check: taking a number of updates from the
whole set, it traced back (via the hand-marked sheets and the texts)
to ensure that they did indeed arise from a text in the object sample
rather than the control text.

RESULTS OF THE STUDY AND THEIR

SIGNIFICANCE
Let us begin by asking the main question: has this procedure
improved translation quality as judged by bilinguals?
Question 1: fidelity of translation as judged by FTD and non-FTD
bilinguals
 SYSTRAN 177

Of 3,655 sentences analyzed by both FTD and non-FTD bilinguals,

the standard application of confidence intervals around any of the
percentages for responses A through D shows them to be statistically
significant, due to the large sample size of the texts (see the appendix
for the A, B, C, D codes). Using the orthodox formula for
confidence intervals yielded a maximum error (at the 95 percent
confidence level) of less than 2 percent either way for any of these
categories. It was thus established that the technique did improve the
translation, with an improvement rate of 47 ± 2 percent.
The pattern and significance levels were the same for both FTD
and non-FTD evaluators, with a large variance within both sets of
results. The percentage of C codes given varies inside both teams by
around 14per cent. The FTD team had a high of 53per cent
improvement by one evaluator, while the non-FTD team went from 51
percent to 38 percent accuracy, a difference that cannot be accounted
for by the varying difficulty of the text batches (the batch which for B2
improved by 38 percent caused A2 to find 47 percent improvement).
These differences of proportion (Z test=5.89) are statistically
significant at any known level. As the frequency of B code judgments
also varied (though not as much), the most obvious measure of
success (percentage improved minus percentage made worse)
fluctuated considerably. The table below shows this:

The average, it should be noted, was for a net improvement rate of

36 per cent. We now turn to the control text results for the carryover
effect on question 1.
Once more, the basic table shows us the consistency between the
aggregate results of the two teams, how the results demonstrated a
statistically significant improvement and that the A/D distinction was an
unreliable one, although the combined A+D response categories were
nearly constant between the two teams (a matter I shall return to below).
178 Yorick Wilks

Note: Some sets of figures do not add up to 100 because

of rounding errors. The margin of error for any of these
percentages was again less than ± 2 per cent.

The fluctuation range was much the same over evaluators as before,
though its net impact is less important. The table below gives the
balance, and, of course, displays the range of fluctuation:

Note: Some sets of figures do not add up to 100 because

of rounding errors.

Question 2: intelligibility
The results may be summarized as follows:

Note: Some sets of figures do not add up to 100 because of rounding errors.
All percentages accurate to ± 1%.
 SYSTRAN 179

These results generally confirm that for both object (O) and control (C)
texts, there was a significant improvement in the second translation. For
the O text, this should be assessed as 44 percent to 53 percent
improvement. (All C code percentages have a margin of error of no
more than 2 percent.) Little attention should be paid to the monolingual
value (especially where it differs in the C text), as the three monolinguals
disagreed with each other sharply. The overall result for the bilinguals
can be seen by pooling the FTD and non-FTD teams as follows:

Note: Some sets of figures do not add

up to 100 because of rounding errors.
All percentages accurate to ± 1%.

This shows us an average improvement (C-B%) of 40 percent for the

object text and 24 percent for the control text, subject to the limits shown.
Question 3: naturalness of the English
In some ways, these were the most interesting results, in that they
clearly showed the difference between monolinguals and bilinguals.
The table below gives the percentages for the two bilingual teams
and the monolinguals separately:

Note: Some sets of figures do not add up to 100 because of founding errors.
All percentages accurate to ± 2%.
180 Yorick Wilks

(All percentages in this table have a statistical margin of error of less

than 5 percent.) Here we see that bilinguals were unprepared to treat
the sentences as good enough to make judgments about them on this
dimension, although they did judge there to be a significant
improvement (20 percent). The monolinguals, on the other hand,
found the judgement relatively easy to make, and found nearly 50
percent improvement for the O text, reduced to just under 40
percent for the C text (I shall return to this contrast later).
To summarize the results in one great table (see Appendix for the
questions asked):

Proportion of improvement (i.e., column C responses, not

improvement balance C-B)

Note: All these figures were subject to very small error margins (1.5–2.5 per
cent), and are statistically significant. The slots followed by an asterisk are
those in which the FTD/non-FTD differences are insignificant.

The effect of question and choice order

The complete results show us that inter-evaluator disagreement was
high and that question order might have been one cause of this, as
the FTD bilinguals answered not in order 1 2 3 but 2 1 3, 3 2 1 and
3 2 1, respectively. While question order as one possible cause of the
variance cannot be ruled out, I do not believe it explains much of it.
The table below demonstrates that there is no significant difference
between improvement percentages among any of the orderings in
question 1.
 SYSTRAN 181

Note: Some sets of figures do not add up to 100 because

of founding errors. All percentages accurate to ± 1%.

One might argue that the difference in question 3’s results was
dependent on ordering: those who answered it last (non-FTD and
B1 at FTD) were less likely to find improvement (only 18 percent of
the cases), while B2 and 3 (at FTD), who answered it fiirst, found a
28 percent improvement. But if we look further, we find that there
was actually more disagreement within this order group (B2 = 19
percent, B3 = 33 percent) than between the two groups!
Furthermore, A3 (non-FTD) found as high an improvement score
(32 per cent) as did B3 (FTD), while C3 (non-FTD) found an even
higher one (38 percent).
To investigate the effect of choice order, the Essex (A) group re-
did a large sample of their data sheets in the choice order B C D A.
The average difference they found was around 4 per cent: most
likely an insignificant figure.

Evaluator variance
As already noted, a striking feature of the results is the high level
of evaluator variance. The standard deviation of the twenty-four
individual judgments made by nine evaluators on three questions
(monos did not answer the first question) is very high: 17.2
percent (19.4 percent for the control text) for the proportion
deemed judgeable (the sum of B+C percent). While this is
unfortunate, it is compensated for by the much lower standard
deviation for those judgments that were made, around 4.8–4.9
percent. In other words, we should attach little importance to the
figures when a sentence translation pair (i.e. first run, second run)
could not be judged as different, but considerable reliability to the
70–80 percent figure for improvement when a decision could be
made. Thus while the average unjudgeable (A+D) proportion was
50 percent for O text and 60 percent for C text, the range within
182 Yorick Wilks

which the true figure lies is much greater, for the margin of error
was +7 percent. But for the actual judgments, we can confidently
state that the error range was less than 2 percent. This is
reassuring because had the reverse result occurred (i.e. had the
evaluations of improvement varied greatly in a subjective way),
we would have had cause to doubt our entire methodology of
evaluation.
Further confirmation of our method came from examining
correlations between evaluations of O and C texts. Reassuringly,
the correlation coefficient over the 25 judgements made on each
text is not significantly different from zero. On the other hand,
the tendency to deem sentences unjudgeable was shown to arise
from evaluators’ individual differences in outlook; the correlation
between each evaluator’s unjudgeability evaluations for each
question between O and C text was amazingly high (0.913). As
this clearly did not represent any actual link between such
evaluation items (the texts having been drawn at random), the
average level and the variance of these decisions should not
concern us.

SHORTCOMINGS OF THE PRESENT STUDY

After the completion of their task, the evaluators were invited to
comment on the questionnaire and the entire study in which they
had taken part. Their comments are distilled in the following
discussion of a number of interconnected problems:

The A/D choice

Even for the bilinguals, the distinction betwen codes A and D
seemed unreliable (monolinguals had no way of expressing their
preference for a D-choice sentence, as the bilinguals could by
choosing A, for instance). The FTD team found 7 percent more
sentences to be unjudgeable because of translation errors (choice
A), and 7 percent fewer sentences undifferentiable (choice D).
This pattern was further highlighted by the individual evaluator’s
distinctions: there was a high negative correlation (-0.715)
between categories A and D. The ambiguity follows from the fact
that the choices Code A (either i: both English sentences are so
bad as translations that no choice can be made between them; or
ii: both English sentences contain Russian words) or Code D (iii:
 SYSTRAN 183

no preference) are subjective. Of these, (ii) is an objective

assessment, while (i) and (iii) are not mutually exclusive. Our
assumption was that choice D applied to two equally acceptable
translations, and choice A to two equally unacceptable
translations but, unfortunately, the evaluators appear not to have
acted according to our assumption.
The basic error here was not a lack of strict definition for the
categories (as some monolinguals felt) but asymmetry between mono
and bilinguals. This may have contributed to the wide variance (and
hence downgrading) of the monolinguals’ evidence. To compensate
for this effect, we simply combined choices A and D into an
‘unjudgeable’ category, leaving us with three possibilities: worsened,
improved and unjudgeable translations.

The ‘question 3 effect’

Certainly, the strong negative correlation between choices A and
D cannot be explained away by a mistaken blurring of the two
categories (as can be seen by the strikingly different A/D behavior
of the mono and bilinguals between questions 2 and 3). As I
remarked earlier, in responding to question 3 the bilinguals
simply declined to judge the naturalness of the English, and often
took refuge in choice A, even though they were quite prepared to
accept these same translations in question 2 on the basis of
intelligibility (A choices for question 2 for the two bilingual
groups were 37 percent and 21 percent versus 65 percent and 88
percent for question 3), while the monolinguals had no problem
making the choice, a fact that confirms the value of monolingual
judgments.

The original sampling of O and C texts

Our process standardized the total number of words in the
documents but not the number of documents in each sample.
Thus, if document length varied greatly, word samples drawn
from a smaller number of long texts would be chosen alongside
word samples drawn from a large number of short texts, which
creates a problem if there is, in fact, a relationship between the
length of a text and the nature of its language. By making the
documents sampled units, a random selection could have
represented them proportionately to the frequency with which
184 Yorick Wilks

different text lengths co-occurred in the corpus. But happily, no

such problem arose.

The ‘Russian word problem’

Part of the A choices refer to the presence of Russian words in the
output but, as we saw with B1, an evaluator may be tempted to
interpret this loosely (considering the word’s meaning obvious, or
unimportant), or deeming a Cyrillic misspelling not a Russian word.
On the other hand, some monolinguals can guess the meaning of a
Russian word, especially if it is close to the Latin spelling, as in
N’YU S DEY (‘Newsday’ in 78054/177). Furthermore, the
questionnaire forces the monolingual to choose the sentence without
Russian words as more natural English. Finally, there were many
sentences where the program translated a proper name into English,
also creating confusion in the mind of the monolinguals, who might
well believe that (wholly) English sentences contained Russian
words.
In summary, what it meant for a sentence to contain a Russian
word was not made totally clear, presenting a genuine difficulty
for clarity. As fewer than 10 percent of the sentences contained
Russian words, this had a negligible effect on our results. In
future tests, the problem could b e avoided by removing
untranslated items in the sentences from the data seen by the
evaluators.

Monolingual expertise
It is generally accepted that the value of monolingual evaluation in
scientific subjets depends on monolingual subject expertise. While
our monolingual evaluators all had some expertise in the field of
political science, this simply did not transfer from Russian to English
in the way that a universally understood area like physics would. To
some degree, this explains the high variance among the
monolinguals, and their consequently diminished role compared to
the BR study of scientific texts.

CONCLUSION: SUMMARY OF FINDINGS

1 While the Battelle Report was poorly argued and statistically
 SYSTRAN 185

flawed, it provided us with the methodological basis for a new

study.
2 The most significant finding was the 20 percent carry-over effect
from updated to control text (balance of improvement: 30 percent
of sentences improved minus 10 percent worsened) in a very
different subject area from the one for which the system was
originally developed.
3 There was a very high variance among evaluators, especially
monolinguals. This was reduced to a significant result by
distinguishing between sentences deemed judgeable and, of those
judgeable, taking those deemed improved. While variance as to
what was judgeable remained high, in the vital category of which
judgeables were improved, variance was minimal: a strong,
indirect confirmation of our methodology.
4 Since the question of naturalness of English output produced an
odd response in the bilinguals, it is better ignored, especially as
this notion is of little importance to the ultimate monolingual user,
in any case.

APPENDIX

Notes for evaluators

It is most important that you study these notes and the

questionnaire before reading the data.
In the body of the data you will find sets of three items: a Russian
sentence followed by two English sentences. If you do not know
any Russian, you should simply ignore the former, and concentrate
on the latter. The English sentences which contain items that are
not really in English (those that contain numbers, for instance) are
easily spotted, and will be referred to in the questions as Russian
words, even though they are written mostly in the English
alphabet.
When looking at the English sentences you should ignore all
questions of stylistic nicety, such as differences between British and
American English, and think only in terms of what is natural English
for you.
186 Yorick Wilks

When looking at both the Russian and the English, you should be
careful not to assume that the sets of sentences form coherent,
continuous texts; rather, you should treat each triplet individually.
Also, do not assume that the second English sentence is better than
the first: the order of the sentence pairs is entirely random.
The difference between ‘understandable’ and ‘natural’ can be
illustrated as follows: the sentences ‘To John gave I the apple’ and
‘I want you go now’ are both understandable but not natural
English.

Questionnaire
Each line on the form corresponds to one sentence triplet (one
Russian and two English) by number. Each numbered section below
corresponds to a column on the answer form. You should circle one
and only one letter (A, B, C or D) for each question and then do that
for each sentence triplet in the data.

1 (enter in column 1 for each triplet) Look at the set of three

sentences and consider (if you can) whether the English sentences
are accurate translations of the Russian one.

Circle
A if you do not speak Russian, OR if you speak Russian and
consider both English sentences to be such bad translations that
no choice can be made between them, OR if you speak Russian
and can see that BOTH English sentences contain Russian
words
B if you prefer the first sentence as an accurate translation
C if you prefer the second sentence
D if you have no preference

2 (enter in column 2 for each triplet) Now look at only the English
sentences in the triplet, and ask yourself if you can comprehend
them as such, accounting for your knowledge of the subject
matter.

Circle
A if you speak Rusian, but consider both texts to be such bad
translations that you decline to form a judgment, OR if both
 SYSTRAN 187

English sentences contain Russian words (this option is available

to non-Russian speakers, as well)
B if you prefer the first sentence for its understandability
C if you prefer the second sentence
D if you have no preference

3 (enter in column 3 for each triplet) Consider the English

sentences alone once more, and judge their naturalness of
language (word order, word choice and so forth).

Circle
A if you speak Russian and consider both sentences such bad
translations of the Russian that you decline to make this
judgment, OR if both English sentences contain Russian words
(once again, you can select this if you do not speak Russian but
should NOT do so if only one of the English sentences contains
Russian words)
B if you prefer the first sentence for the naturalness of its English
C if you prefer the second sentence
D if you have no preference

BIBLIOGRAPHY
ALPAC (1966) Language and Machines: Computers in Translation and Linguistics
(Report by the Automatic Language Processing Advisory Committee,
Division of Behavioral Sciences, National Research Council),
Washington, DC: National Academy of Sciences.
Battelle Columbus Laboratories (1977) ‘The evaluation and systems
analysis of the SYSTRAN machine translation system’, RADC-TR-76–
399 Technical Report.
Brown, P.P., Cocke, J., Della Pietra, S.A., Della Pietra, V.J., Jelinek, F.,
Lafferty, J.D., Mercer, R.L. and Roossin, P.S. (1990) ‘A statistical
approach to machine translation’, Computational Linguistics 16:79–85.
Carroll, J.B. (1966) ‘An experiment in evaluating the quality of translations’,
Mechanical Translation and Computational Linguistics 9 (3 & 4): 55–66.
Jacobs, P., Krupka, G. and Rau, L. (1991) ‘Lexico-semantic pattern
matching as a companion to parsing in text understanding’, in Proceedings
of the DARPA Speech and Natural Language Workshop, Monterey, California.
Johnson, R., King, M. and des Tombe, L. (1985) ‘E U ROTRA: a
multilingual system under development’, in Computational Linguistics 11(2–
3): 155–69.
Toma, P. (1976) ‘An operational machine translation system’, in R.W.
Brislin (ed.) Translation: Applications and Research, New York: Gardner, 247–
59.
188 Yorick Wilks

Wilks, Y. (1990) ‘Form and content in semantics’, in Synthèse 82:329–51.

Wilks, Y., Carbonell, J., Farwell, D., Hovy, E. and Nirenburg, S. (1990)
‘Machine translation again?’, in Proceedings of the DARPA Speech and Natural
Language Workshop, Monterey, California.
Chapter 11

Current research in machine

translation
Harold L.Somers

INTRODUCTION

The purpose of this chapter is to give a view of current research in

machine translation (MT). It is written on the assumption that readers
are, in fact, more or less familiar with most of the wellknown current
MT projects, or else can find out more about them by following up the
references given. The research described here will be divided into two
sets: the members of the first set have in common a direct line of
descent from the classical ‘second generation’ design. The second set
forms a significantly different set and heterogeneous group of current
MT research projects—mostly rather less well known—having in
common only the feature that they in some sense reject the
conventional orthodoxy that typifies the first group.
Part of the background to this division is the feeling that basic
research within the second-generation paradigm has reached its limits:
for example, it could be said that in the twenty-six years since the
ALPAC report of 1966, the second-generation architecture has led to
only slightly better results than the architecture it replaced; so from
a research point of view it is timely to question all the assumptions
that have been accepted unequivocally in that period, just to see what
would happen. I will return to this view later.

PROBLEMS WITH THE CLASSICAL SECOND-

GENERATION ARCHITECTURE
Let me start by considering the classical second-generation
architecture. Examples would be GETA’s ARIANE system (Boitet
and Nedobejkine 1981; Vauquois 1985; Vauquois and Boitet 1985),

189
190 Harold L.Somers

TAUM’s METEO (Chevalier et al. 1981; Lehrberger and Bourbeau

1988), the European Commission’s EUROTRA (Raw et al. 1988,
1989; Steiner (ed.) 1991) and several other systems which
incorporate most of the typical design features. These include the
well-known notions of linguistic rule-writing formalisms with
software implemented independently of the linguistic procedures,
stratificational analysis and generation, and an intermediate
linguistically motivated representation which may or may not
involve the direct application of contrastive linguistic knowledge.
The key unifying feature is modularity, both ‘horizontal’ and
‘vertical’: the linguistic formalisms are supposed to be declarative, so
that linguistic and computational issues are separated; and the whole
process is divided up into computationally and/or linguistically
convenient modules.
While these are admirable design features, at least insofar as they
seem to address the perceived problems of MT system design pre-
ALPAC, they also lead to several general or specific deficiencies in
design.
In general, they reflect the preferred computational and linguistic
techniques of the late 1960s and early 1970s, which have to a great
extent been superseded. There are now several viable alternatives to
the procedural algorithmic strictly typed programming style; while
in linguistics the transformational-generative paradigm and its
associated stratificational view of linguistic processing (morphology—
surface syntax—deep syntax) has become somewhat old-fashioned.
A serious problem with the stratificational approach is the extent
to which it encourages an approach to translation which I have
called ‘structure-preserving translation as first choice’ (Somers et al.
1988:5). This stems from the commitment to compositionality in
translation, i.e. that the translation of the whole is some not too
complex function over the translations of the parts. This leads to a
strategy which embodies the motto, ‘Let’s produce translations that
are as literal as we can get away with’ (Somers 1986:84). Notice that
this is in direct contrast to the human translator’s view, which is
roughly ‘structure-preserving translation as a last resort’. This
attitude can be seen again in discussions of the need to limit
‘structural transfer’ and to build systems which are essentially
interlingual systems with lexical transfer. But we know very well the
difficulties of designing an interlingua, even if we remove the burden
of a ‘conceptual lexicon’.
I must admit that I do not have a ready solution here. But it seems
 Current research 191

to me important to recognize the limitations and pitfalls of the now

traditional stratified linguistic approach to both processing and
representation, so that even the apparently well-established
technique should not necessarily be assumed as a ‘given’ in MT
system design.
I will end this section by making one other observation. This is
that all MT systems so far have been designed with the assumption
that the source text contains enough information to permit
translation. This is obviously true of non-interactive systems; but it
is also true even of the few systems which interact with a user
during processing in order to disambiguate the source text or to make
decisions (usually regarding lexical choice) about the target text.
Notice, by the way, that I want to distinguish here between truly
interactive systems, and those which merely incorporate some sort
of interactive post-editing. In fact, very few research systems are
truly interactive in this sense (e.g. Ntran—see below). However, the
point I want to make concerns how MT researchers view this
problem: it is seen as a deficiency of the system—that is to say,
either the linguistic theory used, or its implementation—rather than
of the text. Consequently, the solutions offered almost inevitably
involve trying to enhance the performance of the part of the system
seen to be at fault: incorporating a better semantic theory, dealing
with translation units bigger than single sentences, trying to take
account of contextual or real-word knowledge. Of course, these are
all worthy research aims but I think the extent to which they will
address the problems they are supposed to solve is generally
exaggerated.

CURRENT RESEARCH DIRECTLY DESCENDED

FROM THAT ARCH ITECTURE
I want to look now at current research projects which I take to be
directly descended from the second-generation architecture, and
which, therefore, in a sense can be said to be subject to the same
criticisms. The research projects in this group can be divided into
subgroups according to which specific part of the problem of MT, as
traditionally viewed, they try to address. So we have projects which
address the problem of insufficient contextual and real-world
knowledge; projects which seek a more elegant linguistic or
computational linguistic framework; and projects where translation
quality is enhanced by constraining the input.
192 Harold L.Somers

For a while now it has been the conventional wisdom that the
next advance in MT design—the ‘third generation’—would involve
the incorporation of techniques from AI. In his instant classic,
Hutchins is typical in this respect:

The difficulties and past ‘failures’ of linguistics-oriented MT point

to the need for AI semantics-based approaches: semantic parsers,
preference semantics, knowledge databases, inference routines,
expert systems, and the rest of the AI techniques.
(Hutchins 1986:327)

He goes on to say:

There is no denying the basic AI argument that at some stage

translation involves the ‘understanding’ of a [source language]
text in order to convey its ‘meaning’ in a [target language] text.
(idem)

In fact, this assertion has been questioned by several commentators

(Johnson 1983:37; Slocum 1985:16), as Hutchins himself notes.

Incorporating AI techniques
Returning to the question of AI-oriented ‘third-generation’ MT
systems, it is probably fair to say that the most notable example of
this approach is at the Center for Machine Translation at Carnegie
Mellon University (CMU), where a significantly sized research team
was expressly set up to pursue the question of ‘knowledgebased MT’
(KB MT) (Carbonell and Tomita 1987). Similar work on the
Japanese ATR project has recently been reported (Kudo 1990).
What then are the ‘AI techniques’ which the CMU team has
incorporated into its MT system, and how do we judge them?
In the Nirenburg and Carbonell description of KB MT, the
emphasis seems to be on the need to integrate discourse pragmatics
in order to get pronouns and anaphora right (Nirenburg and
Carbonell 1987). This requires texts to be mapped onto a
corresponding knowledge representation in the form of a framebased
conceptual interlingua. More recent descriptions of the project
(Nirenburg 1989; Nirenburg and Levin 1989) stress the use of
domain knowledge. In Kudo’s case, local ‘cohesive’ knowledge (in a
dialogue) is stressed. These are well-respected techniques in the
 Current research 193

general field of AI, and we cannot gainsay their application to MT.

But as twenty years of AI research has shown, the step up from a
prototype ‘toy’ implementation to a more fully practical
implementation is a huge one. And there still remain doubts as to
whether the improvement of quality achieved by these AI techniques
is commensurate with the additional computation they involve.

Better linguistic theories

It is normally said that a major design advance from the first to the
second generation of MT systems was the incorporation of better
linguistic theories, and there is certainly a group of current research
projects which can be said to be focusing on this aspect. This is
especially true if we extend the term ‘linguistic’ to include
‘computational linguistic’ theories. The scientific significance of the
biggest of all the MT research projects—EUROTRA—can be seen as
primarily in its development of existing linguistic models, and
notable innovations include the work on the representation of tense
(van Eynde 1988), work on homogeneous representation of
heterogeneous linguistic phenomena (especially through the idea of
‘featurization of purely surface syntactic elements, and a coherent
theory of ‘canonical form’ (Durand et al. 1991), as well as, in some
cases, the first ever wide-coverage formal (i.e. computational)
descriptions of several European languages. As much as anything
else, EUROTRA has shown the possibilities of an openly eclectic
approach to computational linguistic engineering. Nevertheless,
‘Eurotrians’ will be the first to admit that the list of remaining
problems is longer than the list of problems solved or even half
solved. ‘Lexical gaps’, usually illustrated by the well-worn example
of like/gern, modality and determination are just a few more or less
purely linguistic problems that remain, before we even think of
anaphora resolution, use of contextual and real-world knowledge
and so on, already discussed.
Several research projects have taken a more doctrinaire view of
linguistics in that they have explicitly set out to use MT as a testing
ground for some computational linguistic theory. Most notable of
these is Rosetta (Landsbergen 1987a,b) based on Montague
grammar but we could also mention Ntran (Whitelock et al. 1986;
Wood and Chandler 1988), which uses a combination of Lexical
Functional Grammar (LFG) and Generalized Phrase Structure
Grammar (G PSG) in analysis, and Categorial Grammar for
194 Harold L.Somers

generation. There are several other research projects based on

specific linguistic theories, notably LFG (Rohrer 1986; Alam 1986;
Kudo and Nomura 1986; Kaplan et al. 1989; Sadler et al. 1990; Zajac
1990) but also GPSG (Hauenschild 1986), Head-driven Phrase
Structure Grammar (HPSG) (van Noord et al. 1990), Government
and Binding (Wehrli 1990), Systemic functional Grammar (Bateman
1990), Tree-adjoining Grammars (Abeillé et al. 1990), categorial
grammar (Beaven and Whitelock 1988), Functional Grammar (van
der Korst 1989), Situation semantics (Rupp 1989), and, although it
may be regarded as more of a programming technique than a
linguistic ‘theory’ as such, Logic Grammar (Dymetman and Isabelle
1988; Farwell and Wilks 1989; McCord 1989). Given the right
theory and programming environment it is possible to develop very
quickly a reasonable state-of-the-art toy system: for example, Amores
Carredano was able, in only three man-months, to build an LFG-
based system programmed in P ROLO G with a coverage
comparable to early systems which took many man-years to develop
(Amores Carredano 1990). But this says as much about our
expectations of MT systems as it does about the suitability of LFG
or PROLOG. In all these cases, I think it is fair to say that under the
stress of use in a real practical application, the linguistic models,
whose original developers were more interested in a general
approach than in working out all the fine details, inevitably crack.

Sublanguage
Obviously, the most successful MT story of all is that of the
METEO system, which translates daily more than 30,000 words of
weather bulletins from English into French at a cost of less than 0.5¢
(Canadian) per word, with an accuracy rate of 95 per cent
(Chandioux 1987/9:169). The system performs a translation task too
boring for any human doing it to last for more than a few months,
yet sufficiently constrained to allow an MT system to be devised
which only makes mistakes when the input is ill formed. Some
research groups have looked for similarly constrained domains.
Alternatively, the idea of imposing constraints on authors has a long
history of association with MT. At the 1978 Aslib conference,
Elliston showed how at Rank Xerox acceptable output could be got
out of SYSTRAN by forcing technical writers to write in a style that
would not catch the system out (Elliston 1979). It is interesting to see
much the same experience reported again ten years later, at the same
 Current research 195

forum, but this time using Weidner’s MicroCat (Pym 1990). This
rather haphazard activity has fortunately been ‘legitimized’ by its
association with research in the field of language for special purposes
(LSP), and the word ‘sublanguage’ is starting to be widely used in
MT circles (e.g. Kosaka et al. 1988; Luckhardt 1991). In fact, I see
this as a positive move, as long as ‘sublanguage’ is not just used as a
convenient term to camouflage the same old MT design but with
simplified grammar and a reduced lexicon.
Studies of sublanguage (e.g. Kittredge and Lehrberger 1982)
remind us that the topic is much more complex than that: should a
sublanguage be defined prescriptively (or even proscriptively) as in
the Elliston and Pym examples, or descriptively, on the basis of some
corpus judged to be a homogeneous example of the sublanguage in
question? And note that even the term ‘sublanguage’ itself can be
misleading: in most of the literature on the subject, the term is taken
to mean ‘special language of a particular domain’ as in ‘the
sublanguage (of) meteorology’. Yet a more intuitive interpretation of
the term, especially from the point of view of MT system designers,
would be something like ‘the grammar, lexicon, etc. of a particular
text-type in a particular domain’, as in ‘the sublanguage of
meteorological reports as given on the radio’, which might share
some of the lexis of, say, ‘the sublanguage of scientific papers on
meteorology’, although clearly not (all) the grammar. By the same
token, scientific papers on various subjects might share a common
grammar, while differing in lexicon. Furthermore, there is the
question of whether the notion of a ‘core’ grammar or lexicon is
useful or even practical. Some of these questions are being addressed
as part of one of the MT projects started in 1990 at UMI ST
(Manchester), concerning the design of an architecture for a system
which interacts with various types of experts to ‘generate’ a
sublanguage MT system: I will begin my final section with a brief
description of this research.

SOM E ALTERNATIVE AVENUES OF RESEARCH

In this final section, I would like to mention some research projects
which have come to my attention which, I think, have in common
that they reject, at least partially, the orthodoxy of the ‘second-
generation and derivative’ design, or in some other way incorporate
some ideas which I think significantly broaden the scope of MT
research.
196 Harold L.Somers

Sublanguage plus
One system recently proposed at UMIST is a sublanguage MT
system for the Matsushita company (Ananiadou et al. 1990). The
design is for a system with which individual sublanguage MT
systems can be created, on the basis of a bilingual corpus of ‘typical’
texts. The system therefore has two components: a core MT engine,
which is to a certain extent not unlike a typical second-generation
MT system, with explicitly separate linguistic and computational
components; and a set of expert systems which interact with humans
in order to extract from the corpus of texts the grammar and lexicon
that the linguistic part of the MT system will use. The expertise of
the expert systems and the human users is divided between domain
expertise and linguistic expertise, corresponding to the separate
domain knowledge and linguistic knowledge (i.e. of grammars,
lexicons and contrastive knowledge). Using various statistical
methods (see below), the linguistic expert system will attempt to
infer the grammar and lexicon of the sublanguage, on the
assumption that the corpus is fully representative (and approaches
closure). From our observation of other statistics-based approaches
to MT, we concluded that the statistical methods needed to be
‘primed’ with linguistic knowledge, for example, concerning the
nature of linguistic categories, morphological processes and so on.
We have investigated the extent to which this can be done without
going as far as to posit a core grammar, as we are uneasy about the
idea that a sublanguage be defined in terms of deviation from some
standard. The system will make hypotheses about the grammar and
lexicon, to be confirmed by a human user, who must clearly be a
linguist rather than, say, the end user. In the same way, the
contrastive linguistic knowledge is extracted from the corpus, to be
confirmed by interaction with a (probably different) human. Again,
some ‘priming’ is almost always necessary.

Automatic grammar updating

A research project which I find particularly appealing concerns an
MT system which revises its own grammars in response to having its
output post-edited (Nishida et al. 1988; Nishida and Takamatsu
1990). A common complaint from post-editors is that postediting
MT output is frustrating, not least because the same errors are
repeated time and time again (e.g. Green 1982). The idea that such
 Current research 197

errors can somehow be corrected by feedback from posteditors is

obviously one worth pursuing vigorously.
The idea is roughly as follows: there is a fairly traditional second-
generation type English-Japanese MT system (MAPTRAN) whose
output is post-edited interactively. The post-editing system PECOF
(Post-Editor’s COrrection Feedback) asks post-editors to identify
which of the basic post-editing operations (replacement, insertion,
deletion, movement and exchange) each correction involves, with
optionally a reason expressed in terms of other words in the text, or
some primitive linguistic features, e.g. ‘replace n1 by “new word
sequence” where n1 conflicts with n2 in terms of feature’. What
PECOF does with such a correction is try to locate the linguistic rule
in the MT system responsible for the error, and then propose a
revision of it (typically an extension to the general rule to cover the
particular instance identified), which must be confirmed by the post-
editor.
The way it locates the error is also of interest. Translation errors
are assumed, given the architecture of MAPTRAN, to come from
errors in analysis, lexical transfer or structural transfer. In order to
locate which of these modules is responsible, the corrected text is
subjected to reverse translation back into English using an MT
system which is an exact mirror image of MAPTRAN, i.e. structural
transfer precedes lexical transfer. The intermediate representations at
each stage are compared with the corresponding original
representations, and in this way the discrepancy is highlighted.
Then, the appropriate rule can be located, and amended.
I believe that research on this system is still at an early stage, and
it is obvious that only a certain category of translation error can be
dealt with in this way. But it seems to be a useful way of extending
the grammar and lexicon of the system to account for ‘special cases’
on the basis of experience, rather than relying on linguists to
somehow predict things. If PECOF can also interact with the post-
editor to see how generalizable a given correction is, then this is
clearly an excellent way of developing a large-scale MT system.

Dialogue MT
A recent research direction to emerge is an MT system aimed at a
user who is the original author of a text to be composed in a foreign
language. Two such systems are Ntran, mentioned above, and
Huang’s system (Huang 1990), both for use by a monolingual writer
198 Harold L.Somers

of the source language, which embeds this idea in a fairly standard

interactive MT environment, where interaction with the machine is
aimed at disambiguating the input text. An alternative scenario is
one where the interaction takes place before text is input, in the form
of a dialogue between the system and the user, in which the text to
be translated is worked out, taking into account the user’s
communicative goals and the system’s translation ability.
The idea of automatic composition of foreign language texts was
suggested by Saito and Tomita (1986), and is the basis of work done
at UMIST for British Telecom (Jones and Tsujii 1990). In this
system, the user collaborates with the machine to produce high-
quality ‘translations’ of business correspondence on the basis of pre-
translated fragments of stereotypical texts with slots in them which
are filled in by interaction. The advantage is that the system only
translates what it ‘knows’ it can translate accurately, with the result
that the system shows what MT can do, rather than what it cannot,
as in traditional MT. Obviously, though, this strength is also a
weakness in the sense of the severe limitation on what the system can
be used for.
However, we can extend the idea to make it more flexible, and
conceive of a system which has more scope concerning the range
of things it can translate, with corresponding degrees of
confidence about translation quality. This is the case in our
dialogue-MT system (Somers et al. 1990) which we are working
on in collaboration with the Japanese ATR research organization:
we are constructing a system which will act as a bilingual
intermediary for the user in a dialogue with a conference office,
where the user wants to get information about a forthcoming
conference. It is thus a ‘dialogue-MT system’ both in the sense
that it enters into a dialogue with the user about the translation
(see Boitet 1989), and in that the object of the translation is the
user’s contribution to a dialogue. Dialogue is a particularly good
example of the problem, inherent in MT, that the translation of
the text depends to a greater or lesser extent on the surrounding
context (Tsujii and Nagao 1988). In other words, the source text
alone does not carry sufficient information to ensure a good
translation. We envisage a sort of MT ‘expert system’ which can
play the role of an ‘intelligent secretary with knowledge of the
foreign language’, gathering the information necessary to
formulate the target text by asking the user questions, pushing the
user towards a formulation of the ‘source’ text that the system can
 Current research 199

be confident of translating correctly, on the basis of some existing

partial ‘model translations’ which have been supplied by a human
expert beforehand.
The fact that the object of translation is also part of a dialogue
(with another user) adds another dimension of complexity to the
project described in Somers et al. (1990) but the idea of dialogue MT
in general is an interesting development away from the current
situation where the MT system makes the best of what it is given
(and cannot really be sure whether or not its translation is good)
towards a situation where quality can be assured by the fact that the
system knows what it can do and will steer the user to the safe
ground within those limitations.

Corpus-based MT
The approaches to be described in this final section have in common
the idea that a pre-existing large corpus of already translated text
could be used in some way to construct an MT system. They can be
divided into three types, called ‘memory-based’, ‘example-based’ and
‘statistics-based’ translation.

Memory-based translation
The most ‘linguistic’ of the corpus-based approaches is ‘memory-
based translation’ (Sato and Nagao 1990): here, example translations
are used as the basis of new translations. The idea—first suggested by
Nagao (1984)—is that translation is achieved by imitating the
translation of a similar example in a database. The task becomes one
of matching new input to the appropriate stored translation. In this
connection, a secondary problem is the question of the most
appropriate means of storing the examples. As they believe that
combining fragments of sentences is an essential feature, it is natural
for Sato and Nagao to think in terms of storing linguistic objects—
notably partial syntactic trees (in their case, dependency trees). An
element of statistical manipulation is introduced by the need for a
scoring mechanism to choose between competing candidates.
Advantages of this system are ease of modification—notably by
changing or adding to the examples—and the high quality of
translation seen here again, as above, as a result of translations being
established a priori rather than compositionally (although there is an
element of compositionality in Sato and Nagao’s approach). The
200 Harold L.Somers

major disadvantage is the great deal of computation involved,

especially in matching partial dependency trees.

Example-based translation
A similar approach which overcomes this major demerit has been
developed quite independently by two groups of researchers at ATR
in Japan (Sumita et al. 1990), and at UMIST in Manchester (Carroll
1990). In both cases, the central point of interest is the development
of ‘distance’ or ‘similarity’ measures for sentences or parts of
sentences, which permit the input sentence to be translated to be
matched rapidly against a large corpus of existing translations. In
Carroll’s case, the measure can be ‘programmed’ to take account of
grammatical function words and punctuation, which has the effect of
making the algorithm apparently sensitive to syntactic structure
without actually parsing the input as such. While Sumita et al.’s
intention is to provide a single correct translation by this approach,
Carroll’s measure is used in an interactive environment as a
translator’s aid, selecting a set of apparently similar sentences from
the corpus, to guide the translator in the choice of the appropriate
translation. For this reason, spurious or inappropriate selections of
examples can be tolerated as long as the correct selections are also
made at the same time.

Statistics-based approaches
Other corpus-based approaches have been more overtly statistical or
mathematical. The most notable of these is the work at IBM (Brown
et al. 1988a,b, 1990). These researchers, encouraged by the success
of statistics-based approaches to speech recognition and parsing,
decided to apply similar methods to translation. Taking a huge
corpus of bilingual text available in machine-readable form (3 million
sentences selected from the Canadian Hansard), the probability that
any one word in a sentence in one language corresponds to zero, one
or two words in the translation is calculated. The glossary of word
equivalences so established consists of lists of translation possibilities
for every word, each with a corresponding probability. For example,
the translates as le with a probability of 0.610, as la with probability
0.178, and so on. These probabilities can be combined in various
ways, and the highest-scoring combination will determine the words
which will make up the target text. An algorithm to get the target
 Current research 201

words in the right order is now needed. This can be calculated using
rather well-known statistical methods for measuring the probabilities
of word-pairs-triples, etc.
The results of this experiment are certainly interesting.
Translations which were either the same as or preserved the meaning
of the official translations were achieved in about 48 per cent of the
cases. Although at first glance this level of success would not seem to
make this method viable as it stands, it is to be noted that not many
commercial MT systems achieve a significantly better quality. More
interesting is to consider the near-miss cases in the IBM experiment:
incorrect translations were often the result of the fact that the system
contains no linguistic ‘knowledge’ at all. Brown et al. admit that
serious problems arise when the translation of one word depends on
the translation of others, and suggest (Brown et al. 1988a:11–12) that
some simple morphological and/or syntactic analysis, also based on
probabilistic methods, would greatly improve the quality of the
translation.
As part of sublanguage MT research at UMIST, Arad (Arad
1991) investigated the possibility of using statistical methods to
derive morphological and syntactic grammars and mono- and
bilingual lexica from bilingual corpora. A similar goal is reported by
Kay and Röscheisen (1988). That the text type is restricted permits
Arad to work with lower thresholds of statistical significance, and
hence smaller corpora. In the future she intends to prime the system
with a certain amount of a priori linguistic knowledge of a very basic
kind, for example, what sort of morphological processes are likely?
(e.g. for English mainly non-agglutinative suffixes, stems should
contain a vowel and be longer than their affixes), typical
characteristics of phrase structure (notions of open- and closed-class
words, headedness and so on). Arad has resisted the idea of priming
the system with a core grammar but recognizes that this may prove
to be a necessary step.

CONCLUSIONS
In this chapter I have given a view—at times personal—of current
research in MT. Of course, there are probably numerous research
projects that I have omitted to mention, generally because I have not
been able to get information about them, or simply because they
have not come to my notice or have emerged while this book was in
press. I am conscious that some readers of this chapter will be
202 Harold L.Somers

relative newcomers to the field, and so I should stress that my

coverage of the subject here has been from my own viewpoint rather
than as a neutral reporter. For those readers, let me end by indicating
some possible future sources of information on MT research.
The conference series ‘Theoretical and Methodological Issues in
Machine Translation of Natural Languages’, which recently held its
fourth meeting in Montreal in 1992, is rapidly becoming an
established event for MT researchers. To a lesser extent, the
International Conference on Computational Linguistics (COLING)
and meetings of the ACL likewise provide a much-appreciated forum
for publicizing ongoing research. There are typically one or two MT-
related events in Japan every year, which generally repay the effort
(and expense) of attending them; the annual Aslib Translating and
the computer conference in London views MT from the translator’s
point of view. Finally, there are two specialist quarterly journals,
Machine Translation, and Applied Computer Translation, published by
Sigma Press, which, as their names suggest, carry articles of direct
relevance.

BIBLIOGRAPHY
Abeillé, A., Schabes, Y. and Joshi, A.K. (1990) ‘Using lexicalized tags for
machine translation’, in H.Karlgren (ed.) COLING-90: Papers Presented to
the 13th International Conference on Computational Linguistics, Helsinki:
Yliopistopaino, vol. 3:1–6.
Alam, Y.S. (1986) ‘A lexical-functional approach to Japanese for the purpose
of machine translation’, Computers and Translation 1:199–214.
ALPAC (1966) Language and Machines: Computers in Translation and Linguistics
(Report by the Automatic Language Processing Advisory Committee,
Division of Behavioral Sciences, National Research Council),
Washington, DC: National Academy of Sciences.
Amores Carredano, J.G. (1990) ‘An LFG machine translation system in
DCG-PROLOG’, unpublished MSc dissertation, UMIST, Manchester.
Ananiadou, S., Carroll, J.J. and Phillips, J.D. (1990) ‘Methodologies for
development of sublanguage MT systems’, CCL Report 90/10, Centre
for Computational Linguistics, UMIST, Manchester.
Arad, I. (1991) ‘A quasi-statistical approach to automatic generation of
linguistic knowledge, PhD thesis, UMIST, Machnester.
Bateman, J.A. (1990) ‘Finding translation equivalents: an application of
grammatical metaphor’, in H.Karlgren (ed.) COLING-90: Papers Presented
to the 13th International Conference on Computational Linguistics, Helsinki:
Yliopistopaino, vol. 3:13–18.
Beaven, J.L. and Whitelock, P. (1988) ‘Machine translation using
isomorphic UCGs’, in D.Vargha (ed.) COLING Budapest: Proceedings of the
12th International Conference on Computational Linguistics, Budapest: John von
 Current research 203

Neumann Society for Computing Sciences, 32–5.

Boitet, Ch. (1989) ‘Speech synthesis and dialogue based machine
translation’, in ATR Symposium on Basic Research for Telephone Interpretation,
Kyoto, Japan, Proceedings, 6/5/1–9.
Boitet, Ch. and Nedobejkine, N. (1981) ‘Recent developments in Russian-
French machine translation at Grenoble’, Linguistics 19:199–271.
Brown, P.P., Cocke, J., Della Pietra, S.A., Della Pietra, V.J., Jelinek, J.,
Lafferty, J.D., Mercer, R.L. and Roossin, P.S. (1990) ‘A statistical
approach to machine translation’, Computational Linguistics 16:79–85.
Brown, P.P., Cocke, J., Della Pietra, S.A., Della Pietra, V.J., Jelinek, J.,
Mercer, R.L. and Roossin, P.S. (1988a) ‘A statistical approach to French/
English translation’, Proceedings, Second International Conference on Theoretical
and Methodological Issues in Machine Translation of Natural Languages, Carnegie
Mellon University, Pittsburgh, Pennsylvania.
——(1988b) ‘A statistical approach to language translation’, in D. Vargha
(ed.) COLING Budapest: Proceedings of the 12th InternationalConference on
Computational Linguistics, Budapest: John von Neumann Society for
Computing Sciences, 71–6.
Carbonell, J.G. and Tomita, M. (1987) ‘Knowledge-based machine
translation, the CMU approach’, in S.Nirenburg (ed.) Machine Translation:
Theoretical and Methodological Issues, Cambridge: Cambridge University
Press, 68–89.
Carroll, J.J. (1990) ‘Repetitions processing using a metric space and the
angle of similarity’, CCL Report 90/3, Centre for Computational
Linguistics, UMIST, Manchester.
Chandioux, J. (1987/9) ‘10 ans de METEO (MD)’ (texte présenté lors du
Congrès du Conseil des traducteurs et interprètes du Canada, octobre
1987) reprinted in A.Abbou (ed.) La Traduction Assistée par Ordinateur:
Perspectives technologiques, industrielles et économiques envisageables à l’horizon
1990: l’offre, la demande, les marchés et les évolutions en cours, Actes du
séminaire international (Paris, March 1988), Paris: DAICADIF, 169–73.
Chevalier, M., Isabelle, P., Labelle, F. and Lainé, C. (198 1) ‘La
traductologie appliquée à la traduction automatique’, Meta 26:35–47.
Durand, J., Bennett, P., Allegranza, V., van Eynde, F., Humphreys, L.,
Schmidt, P. and Steiner, E. (1991) ‘The Eurotra linguistic specifications:
an overview’, Machine Translation, 6:103–47.
Dymetman, M. and Isabelle, P. (1988) ‘Reversible logic grammars for
machine translation’, Proceedings, Second International Conference on Theoretical
and Methodological Issues in Machine Translation of Natural Langauges, Carnegie
Mellon University, Pittsburgh, Pennsylvania.
Elliston, J.S.G. (1979) ‘Computer-aided translation: a business viewpoint’,
in B.M.Snell (ed.) Translating and the Computer, Amsterdam: North-
Holland, 149–58.
Farwell, D. and Wilks, Y. (1989) ‘Ultra: a multilingual machine translator’,
Research Report, Computing Research Laboratory, New Mexico State
University, Las Cruces, New Mexico.
Green, R. (1982) ‘The MT errors which cause most trouble to posteditors’,
in V.Lawson (ed.) Practical Experience of Machine Translation, Amsterdam:
North-Holland, 101–4.
204 Harold L.Somers

Hauenschild, C. (1986) ‘KIT/NASEV oder die Problematik des Transfers

bei der maschinellen Übersetzung’, in I.Bátori and H.J.Weber (eds) Neue
Ansätze in Maschineller Sprachübersetzung: Wissensrepräsentation und Textbezug,
Tübingen: Niemeyer Verlag, 167–95.
Huang, X. (1990) ‘Machine translation in a monolingual environment’, in
Third International Conference on Theoretical and Methodological Issues in Machine
Translation of Natural Languages, Austin, Texas, 33–41.
Hutchins, W.J. (1986) Machine Translation: Past, Present, Future, Chichester:
Ellis Horwood.
Johnson, R.L. (1983) ‘Parsing—an MT perspective’, in K.Sparck Jones and
Y.Wilks (eds) Automatic Natural Language Parsing, Chichester: Ellis
Horwood, 32–8.
Jones, D. and Tsujii, J. (1990) ‘Interactive high-quality machine translation
for monolinguals’, in Third International Conference on Theoretical and
Methodological Issues in Machine Translation of Natural Languages, Austin,
Texas, 43–46.
Kaplan, R.M., Netter, K., Wederkind, J. and Zaenen, A. (1989) ‘Translation
by structural correspondences’, Fourth Conference of the European
Chapter of the Association for Computational Linguistics, Manchester,
Proceedings, 272–81.
Karlgren, H. (ed.) (1990) COLING-90: Papers Presented to the 13th International
Conference on Computational Linguistics, Helsinki: Yliopistopaino.
Kay, M. and Röscheisen, M. (1988) ‘Text-translation alignment’, Research
Memo, Xerox Palo Alto Research Center, Palo Alto, California.
Kittredge, R. and Lehrberger, J. (1982) Sublanguage: Studies of Language in
Restricted Semantic Domains, Berlin: de Gruyter.
Kosaka, M., Teller, V. and Grishman, R. (1988) ‘A sublanguage approach to
Japanese-English machine translation’, in D.Maxwell, K.Schubert and
T.Witkam (eds) New Directions in Machine Translation, Dordrecht: Foris,
109–20.
Kudo, I. (1990) ‘Local cohesive knowledge for a dialogue—machine
translation system’, in H.Karlgren (ed.) COLING-90: Papers presented to the
13th International Conference on Computational Linguistics, Helsinki:
Yliopistopaino, vol. 3:391–3.
Kudo, I. and Nomura, H. (1986) ‘Lexical-functional transfer: a transfer
framework in a machine translation system based on LFG’, 11th
International Conference on Computational Linguistics, Proceedings of
COLING-86, Bonn: 112–14.
Landsbergen, J. (1987a) ‘Isomorphic grammars and their use in the
ROSETTA translation system’, in M.King (ed.) Machine Translation
Today: the State of the Art, Edinburgh: Edinburgh University Press, 351–
72.
——(1987b) ‘Montague grammar and machine translation’, in P. Whitelock,
M.MacG. Wood, H.L.Somers, R.Johnson and P.Bennett (eds) Linguistic
Theory and Computer Applications, London: Academic Press, 113–47.
Lehrberger, J. and Bourbeau, L. (1988) Machine Translation: Linguistic
Characteristics of MT Systems and General Methodology of Evaluation,
Amsterdam: John Benjamins.
Luckhardt, H.D. (1991) ‘Sublanguages in machine translation’, Fifth
 Current research 205

conference of the European chapter of the Association for

Computational Linguistics, Berlin, Proceedings 306–8.
McCord, M.C. (1989) ‘Design of LMT: a Prolog-based machine translation
system’, Computational Linguistics 15:33–52.
Nagao, M. (1984) ‘A framework of a mechanical translation between
Japanese and English by analogy principle’, in A.Elithorn and R. Banerji
(eds) Artificial and Human Intelligence, Amsterdam, North-Holland, 172–80.
Nirenburg, S. (ed.) (1987) Machine Translation: Theoretical and Methodological
Issues, Cambridge: Cambridge University Press.
——(1989) ‘Knowledge-based machine translation’, Machine Translation 4:
5–24.
Nirenburg, S. and Carbonell, J.G. (1987) ‘Integrating discourse pragmatics
and propositional knowledge for multilingual natural language
processing’, Computers and Translation 2:105–16.
Nirenburg, S. and Levin, L. (1989) ‘Knowledge representation support’,
Machine Translation 4:25–52.
Nishida, F. and Takamatsu, S. (1990) ‘Automated procedures for the
improvement of a machine translation system by feedback from
postediting’, Machine Translation 5:223–46.
Nishida, F., Takamatsu, S., Tani, T. and Doi, T. (1988) ‘Feedback of
correcting information in postediting to a machine translation system’, in
D.Vargha (ed.) COLING Budapest: Proceedings of the 12th International
Conference on Computational Linguistics, Budapest: John von Neumann
Society for Computing Sciences, 476–81.
Pym, P.J. (1990) ‘Pre-editing and the use of simplified writing for MT: an
engineer’s experience of operating an MT system’, in P.Mayorcas (ed.)
Translating and the Computer 10: The Translation Environment 10 Years On,
London: Aslib, 80–96.
Raw, A., Vandecapelle, B. and van Eynde, F. (1988) ‘Eurotra: an overview’,
Interface 3:5–32.
Raw, A., van Eynde, F., ten Hacken, P., Hoekstra, H. and Vandecapelle, B.
(1989) ‘An introduction to the Eurotra machine translation system’,
Working Papers in Natural Language Processing 1, TAAL Technologie,
Utrecht and Katholieke Universiteit Leuven.
Rohrer, C. (1986) ‘Maschinelle Übersetzung mit Unifikationsgrammatiken’,
in I.Bátori and H.J.Web er (eds) Neue Ansätze in Maschineller
Sprachübersetzung: Wissensrepräsentation und Textbezug, Tübingen: Niemeyer
Verlag, 75–99.
Rupp, C.J. (1989) ‘Situation semantics and machine translation’, Fourth
Conference of the European Chapter of the Association for
Computational Linguistics, Manchester, Proceedings, 308–18.
Sadler, L., Crookston, I., Arnold, D. and Way, A. (1990) ‘LFG and
translation’, in Third International Conference on Theoretical and Methodological
Issues in Machine Translation of Natural Languages, Austin, Texas, 121–30.
Saito, H. and Tomita, M. (1986) ‘On automatic composition of stereotypic
documents in foreign languages’, presented at 1st International
Conference on Applications of Artificial Intelligence to Engineering
Problems, Southampton, April 1986; Research Report CMU–CS–86–
107, Department of Computer Science, Carnegie Mellon University.
206 Harold L.Somers

Sato, S. and Nagao, M. (1990) ‘Toward memory-based translation’, in H.

Karlgren (ed.) COLING-90: Papers presented to the 13th International
Conference on Computational Linguistics, Helsinki: Yliopistopaino, vol. 3:247–
52.
Slocum, J. (1985) ‘A survey of machine translation: its history, current
status, and future prospects’, Computational Linguistics 11:1–17.
Somers, H.L. (1986) ‘Some thoughts on interface structure(s)’, in W. Wilss
and K.-D. Schmitz (eds) Maschinelle Übersetzung—Methoden und Werkzeuge,
Tübingen: Max Niemeyer Verlag, 81–99.
Somers, H.[L.], Hirakawa, H., Miike, S. and Amano, S. (1988) ‘The
treatment of complex English nominalizations in machine translation’,
Computers and Translation 3:3–21.
Somers, H.L., Tsujii, J. and Jones, D. (1990) ‘Machine translation without a
source text’, in H.Karlgren (ed.) COLING-90: Papers Presented to the 13th
International Conference on Computational Linguistics, Helsinki: Yliopistopaino,
vol. 3:271–6.
Steiner, E. (ed.) (1991) Special issue on EUROTRA, Machine Translation, 6:2–
3.
Sumita, E., Iida, H. and Kohyama, H. (1990) ‘Translating with examples: a
new approach to machine translation’, Third International Conference on
Theoretical and Methodological Issues in Machine Translation of Natural
Languages, Austin, Texas, 203–12.
Tsujii, J. and Nagao, M. (1988) ‘Dialogue translation vs text translation—
Interpretation based approach’, in D.Vargha (ed.) (1988)
COLI NGBudapest: Proceedings of the 12th International Conference on
Computational Linguistics, Budapest: John von Neumann Society for
Computing Sciences, 688–93.
van der Korst, B. (1989) ‘Functional grammar and machine translation’, in
J.H.Connolly and S.C.Dik (eds) Functional Grammar and the Computer,
Dordrecht: Foris, 289–316.
van Eynde, F. (1988) ‘The analysis of tense and aspect in Eurotra’, in D.
Vargha (ed.) COLING Budapest: Proceedings of the 12th International Conference
on Computational Linguistics, Budapest: John von Neumann Society for
Computing Sciences, 699–704.
van Noord, G., Dorrepaal, J., van der Eijk, P., Florenza, M. and des Tombe,
L. (1990) ‘The MiMo2 research system’, Third International Conference on
Theoretical and Methodological Issues in Machine Translation of Natural
Languages, Austin, Texas, 213–33.
Vargha, D. (ed.) (1988) COLING Budapest: Proceedings of the 12th International
Conference on Computational Linguistics, Budapest: John von Neumann
Society for Computing Sciences.
Vauquois, B. (1985) ‘The approach of Geta to automatic translation:
comparison with some other methods’, paper presented at International
Symposium on Machine Translation, Riyadh; in Ch. Boitet (ed.) (1988)
Bernard Vauquois et la TAO: vingt-cinq ans de traduction automatique— Analectes,
Grenoble: Association Champollion, 631–86.
Vauquois, B. and Boitet, Ch. (1985) ‘Automated translation at Grenoble
University’, Computational Linguistics 11:28–36.
Wehrli, E. (1990) ‘STS: an experimental sentence translation system’, in
 Current research 207

H.Karlgren (ed.) COLlNG-90: Papers Presented to the 13th International

Conference on Computational Linguistics, Helsinki: Yliopistopaino, vol. 1:76–8.
Whitelock, P.J., Wood, M.MacG., Chandler, B.J., Holden, N. and Horsfall,
H.J. (1986) ‘Strategies for interactive machine translation: the experience
and implications of the UMIST Japanese project’, 11th International
Conference on Computational Linguistics, Proceedings of COLING-86, Bonn:
329–34.
Wood, M.M. and Chandler, B.J. (1988) ‘Machine translation for
monolinguals’, in D.Vargha (ed.) COLING Budapest: Proceedings of the 12th
International Conference on Computational Linguistics, Budapest: John von
Neumann Society for Computing Science, 760–63.
Zajac, R. (1990) ‘A relational approach to translation’, Third International
Conference on Theoretical and Methodological Issues in Machine Translation of
Natural Languages, Austin, Texas, 235–54.
Bibliography

Abbou, A. (ed.) (1989) Traduction Assistée par Ordinateur: Perspectives

technologiques, industrielles et économiques envisageables à l’horizon 1990: l’offre, la
demande, les marchés et les évolutions en cours, Actes du Séminaire international
(Paris, March 1988), Paris: DAICADIF.
Abeillé, A., Schabes, Y. and Joshi, A.K. (1990) ‘Using lexicalized tags for
machine translation’, in H.Karlgren (ed.) COLING-90: Papers Presented to
the 13th International Conference on Computational Linguistics, Helsinki:
Yliopistopaino, vol. 3:1–6.
Ahmad, K., Fulford, H., Holmes-Higgin, P., Rogers, M. and Thomas, P.
(1990) ‘The translator’s workbench project’, in C.Picken (ed.) Translating
and the Computer 11: Preparing for the Next Decade, London: ASLIB, 9–19.
Aitchison, J. (1978) Teach Yourself Linguistics, Sevenoaks: Hodder and
Stoughton.
Alam, Y.S. (1986) ‘A lexical-functional approach to Japanese for the purpose
of machine translation’, Computers and Translation 1:199–214.
ALPAC (1966) Language and Machines: Computers in Translation and Linguistics
(Report by the Automatic Language Processing Advisory Committee,
Division of Behavioral Sciences, National Research Council),
Washington, DC: National Academy of Sciences.
Amores Carredano, J.G. (1990) ‘An LFG machine translation system in
DCG-PROLOG’, unpublished MSc dissertation, UMIST, Manchester.
Ananiadou, S., Carroll, J.J. and Phillips, J.D. (1990) ‘Methodologies for
development of sublanguage MT systems’, CCL Report 90/10, Centre
for Computational Linguistics, UMIST, Manchester.
Arad, I (1991) A quasi-statistical approach to automatic generation of
linguistic knowledge, PhD thesis, UMIST, Manchester.
Bar-Hillel, Y. (1951) ‘The state of machine translation in 1951’, American
Documentation 2:229–37.
Bateman, J.A. (1990) ‘Finding translation equivalents: an application of
grammatical metaphor’, in H.Karlgren (ed.) COLING-90: Papers Presented
to the 13th International Conference on Computational Linguistics, Helsinki:
Yliopistopaino, vol. 3:13–18.
Bátori, I. and Weber, H.J. (eds) (1986) Neue Ansätze in Maschineller
Sprachübersetzung: Wissensrepräsentation und Textbezug, Tübingen: Niemeyer
Verlag.

208
 Bibliography 209

Battelle Columbus Laboratories (1977) ‘The evaluation and systems

analysis of the SYSTRAN machine translation system’, RADC-TR-76–
399 Technical Report.
Beaven, J.L. and Whitelock, P. (1988) ‘Machine translation using
isomorphic UCGs’, in D.Vargha (ed.) COLING Budapest: Proceedings of the
12th International Conference on Computational Linguistics, Budapest: John von
Neumann Society for Computing Sciences, 32–5.
Ben-Avi, S. (1977) ‘An investigation into the use of Esperanto as an
intermediate language in a machine translation project’, PhD thesis,
Manchester.
Bennett, W.S. and Slocum, J. (1985) ‘The LRC machine translation system’,
in Computational Linguistics 11:111–21; reprinted in J.Slocum (ed.) Machine
Translation Systems, Cambridge: Cambridge University Press, 1988, 111–
40.
Blanke, D. (1985) Internationale Plansprachen, Berlin: Akademic-Verlag.
Bloomfield, L. (1933) Language, New York: Holt, Rinehart & Winston
(reprinted in great part in 1984, University of Chicago).
——(1944) ‘Secondary and tertiary responses to language’, in Language
20:45–55 and in C.F.Hockett (ed.) (1970) A Leonard Bloomfield Anthology,
Chicago: University of Chicago Press.
Boitet, Ch. (1989) ‘Speech synthesis and dialogue based machine
translation’, in ATR Symposium on Basic Research for Telephone Interpretation,
Kyoto, Japan, Proceedings, 6/5–1/9.
——(1990) ‘Towards personal MT: general design, dialogue structure,
potential role of speech’, in H.Karlgren (ed.) COLING-90: Papers presented
to the 13th International Conference on Computational Linguistics, Helsinki:
Yliopistopaino, vol. 3, 30–35.
Boitet, Ch. and Nedobejkine, N. (1981) ‘Recent developments in Russian-
French machine translation at Grenoble’, Linguistics 19: 199–271.
Booth, A.D. (ed.) (1967) Machine Translation, Amsterdam: North Holland.
Briem, S. (1990) ‘Maskinoversaettelse fra esperanto til islandsk’, in J.Pind
and E.Rögnvaldsson (eds) Papers from the Seventh Scandinavian Conference of
Computational Linguistics (Reykjavík 198 9), Reykjavík: Institute of
Lexicography/Institute of Linguistics, 138–45.
Brower, R.A. (ed.) (1959) On Translation, Cambridge, Mass.: Harvard
University Press.
Brown, P.P., Cocke, J., Della Pietra, S.A., Della Pietra, V.J., Jelinek, F.,
Lafferty, J.D., Mercer, R.L. and Roossin, P.S. (1990) ‘A statistical
approach to machine translation’, Computational Linguistics 16:79–85.
Brown, P.F., Cocke, J., Della Pietra, S.A., Della Pietra, V.J., Jelinek, J.,
Mercer, R.L. and Roossin, P.S. (1988a) ‘A statistical approach to French/
English translation’, Proceedings, Second International Conference on Theoretical
and Methodological Issues in Machine Translation of Natural Languages, Carnegie
Mellon University, Pittsburgh, Pennsylvania.
——(1988b) ‘A statistical approach to language translation’, in D. Vargha
(ed.) COLING Budapest: Proceedings of the 12th International Conference on
Computational Linguistics, Budapest: John von Neumann Society for
Computing Sciences, 71–6.
Carbonell, J.G. (1990) ‘Machine translation technology: status and
210 Bibliography

recommendations’, in T.Valentine (1990) ‘Status of machine translation

(MT) technology: Hearing before the Subcommittee on Science,
Research and Technology of the Committee on Science, Space and
Technology, US House of Representatives, 101st Congress, Second
Session, September 11, 1990’, [no. 153], Chairman: Rep. T.Valentine,
Washington: US Government Printing Office, 119–31.
Carbonell, J.G. and Tomita, M. (1987) ‘Knowledge-based machine
translation, the CM U approach’, in S.Nirenburg (1987) Machine
Translation: Theoretical and Methodological Issues, Cambridge: Cambridge
University Press, 68–9.
Carroll, J.B. (1966) ‘An experiment in evaluating the quality of translations’,
Mechanical Translation and Computational Linguistics 9 (3 & 4): 55–66.
Carroll, J.J. (1990) ‘Repetitions processing using a metric space and the
angle of similarity’, CCL Report 90/3, Centre for Computational
Linguistics, UMIST, Manchester.
Celt, S. and Gross, A. (1987) ‘The challenge of translating Chinese
medicine’, Language Monthly 43:19–21.
Chandioux, J. (1987/9) ‘10 ans de METEO (MD)’ (texte présenté lors du
Congrès du Conseil des traducteurs et interprètes du Canada, octobre
1987) reprinted in A.Abbou (ed.) La Traduction Assistée par Ordinateur:
Perspectives technologiques, industrielles et économiques envisageables à l’horizon
1990: l’offre, la demande, les marchés et les évolutions en cours, Actes du
Séminaire international (Paris, March 1988), Paris: DAICADIF, 169–73.
——(1989) ‘M ETEO: 100 million words later’, in ATA conference
proceedings, 449–53.
Chevalier, M., Isabelle, P., Labelle, F. and Lainé, C. (198 1) ‘La
traductologie appliquée à la traduction automatique’, Meta 26:35–47.
Chisholm, S. Jr. (1981) Elements of English Linguistics, London: Longman.
Chomsky, N. (1957) Syntactic Structures, The Hague: Mouton.
——(1965) Aspects of the Theory of Syntax, Cambridge, Mass.: MIT Press.
——(1975) The Logical Structure of Linguistic Theory, Chicago: University of
Chicago Press.
Coughlin, J. (1988) ‘Artificial intelligence and machine translation:present
developments and future prospects’, in Babel 34:1, 1–9.
Cruz, E. and Ortiz y Ortiz, R. (1981) Translation Strategies/Estrategias para
Traducción, London: Macmillan.
Crystal, D. (1985) Linguistics, Harmondsworth: Penguin Books Limited.
Danzin, A., Allén, S., Coltof, H., Recoque, A., Steusloff, H. and O’Leary,
M. (1990) ‘Eurotra Programme Assessment Report’, Commission of the
European Communities, March 1990.
Datta, J. (1988) ‘MT in large organizations: revolution in the workplace’, in
M.Vasconcellos (ed.) Technology as Translation Strategy (American
Translators Association Scholarly Monog raph Series, vol. I I),
Binghamton, New York: State University of New York, 167–83.
Denisov, P.N. (1965) Principy modelirovanija jazyka (na materiale vspomogatel’nych
jazykov dlja avtomaticeskogo poiska i perevoda), Moscow: Moskovskij Universitet.
Dietze, J. (1986) ‘Projekt der rechnergestützten Erarbeitung eines
Wörterbuchs von Esperantowurzeln’, Wissenschaftliche Zeitschrift der
Universität Halle, 35 [G, 5]: 90–1.
 Bibliography 211

Drexler, E.K. (1986) Engines of Creation, New York: Anchor Press.

Dulicenko, A.D. (1989) ‘Ethnic language and planned language, and its
place in language science’, in K.Schubert (ed.) Interlinguistics. Aspects of the
Science of Planned Languages, Berlin/New York: Mouton de Gruyter, 47–61.
Durand, J., Bennett, P., Allegranza, V., van Eynde, F., Humphreys, L.,
Schmidt, P. and Steiner, E. (1991) ‘The Eurotra linguistic specifications:
an overview’, Machine Translation, 6:103–47.
Dymetman, M. and Isabelle, P. (1988) ‘Reversible logic grammars for
machine translation’, Proceedings, Second International Conference on Theoretical
and Methodological Issues in Machine Translation of Natural Languages, Carnegie
Mellon University, Pittsburgh, Pennsylvania.
EEC (1990) ‘Council Decision of 26 November 1990 adopting a specific
programme concerning the preparation of the development of an
operational Eurotra system’, Official Journal of the European Communities, no.
L 358/84, EEC/664/90, 21.12.90.
Elliston, J.S.G. (1979) ‘Computer-aided translation: a business viewpoint’,
in B.M.Snell (ed.) Translating and the Computer, Amsterdam: North-
Holland, 149–58.
Farwell, D. and Wilks, Y. (1989) ‘Ultra: a multilingual machine translator’,
Research Report, Computing Research Laboratory, New Mexico State
University, Las Cruces, New Mexico.
Fodor, J.A. and Katz, J.J. (1964) The Structure of Language, New York:
Prentice-Hall.
Forster, P.G. (1987) ‘Some social sources of resistance to Esperanto’, in José-
Luis Melena Jiménez et al. (eds) Serta gratvlatoria in honorem Juan Régulo,
vol. 2: Esperantismo, La Laguna: Universidad, 203–11.
Fulford, H., Höge, M. and Ahmad, K. (1990) ‘User requirements study’,
European Commission Esprit I I Project no. 2315, Translator’s
Workbench Project, Final Report on Workpackage 3.3.
García Yebra, V. (1984) Teoría y Práctica de la Traducción 2nd edn, Madrid:
Editorial Gredos.
Godel, K. (1931) ‘Über formal unentscheidbare Sätze der Principia
Mathematica und verwandte Systeme I’, Monatshefte für Mathematik und
Physik 38:173–98.
Gordos, G. (1985) ‘Parol-sintezo kun limigita vortaro’, in I.Koutny (ed.)
Perkomputila tekstoprilaboro, Budapest: Scienca Eldona Centro, 11–29.
Green, R. (1982) ‘The MT errors which cause most trouble to posteditors’,
in V.Lawson (ed.) Practical Experience of Machine Translation, Amsterdam:
North-Holland, 101–4.
Greenberg, J. (1963) Universals of Language, Cambridge, Mass.: MIT Press.
Grimes, J.E. (1975) The Thread of Discourse, The Hague: Mouton.
Grosjean, F. (1982) Life with Two Languages: an Introduction to Bilingualism,
Cambridge, Mass.: Harvard University Press.
Guzmán de Rojas, I. (1985) ‘Logical and linguistic problems of social
communication with the Aymara people’, Ottawa: The International
Development Research Center.
Halliday, M.A.K. (1967) ‘Notes on transitivity and theme in English’, part 2,
Journal of Linguistics 3:199–244.
Harel, D. (ed.) (1987) Algorithmics: The Spirit of Computing, Wokingham:
Addison-Wesley.
212 Bibliography

Harris, Z. (1951) Structural Linguistics, Chicago: University of Chicago Press.

Hartmann, R.R.K and Stork, F.C. (1976) Dictionary and Language and
Linguistics, New York: Wiley.
Hauenschild, C. (1986) ‘KIT/NASEV oder die Problematik des Transfers
bei der maschinellen Übersetzung’, in I.Bátori and H.J.Weber (eds) Neue
Ansätze in Maschineller Sprachübersetzung: Wissenrepräsentation und Textbezug,
Tübingen: Niemeyer Verlag, 167–95.
Hjelmslev, L. (1961) Prolegomena to a Theory of Language, Madison: University
of Wisconsin Press.
——(1963) Sproget, 2nd edn, Copenhagen: Berlingske forlag.
Hockett, C.F. (1968) The State of the Art, The Hague: Mouton.
——(ed.) (1987) A Leonard Bloomfield Anthology, Chicago: University of
Chicago Press.
Hodges, A. (ed.) (1985) Alan Turing: The Enigma of Intelligence, London:
Unwin Paperbacks.
Huang, X. (1990) ‘Machine translation in a monolingual environment’, in
Third International Conference on Theoretical and Methodological Issues in Machine
Translation of Natural Languages, Austin, Texas.
Hunn, E.S. (1977) Tzeltal Folk Zoology: The Classification of Discontinuities in
Nature, New York: Academic Press.
Hutchins, W.J. (1982) ‘The evolution of machine translation systems’, in
V.Lawson (ed.) Practical Experience of Machine Translation, Amsterdam:
North-Holland, 21–37.
——(1986) Machine Translation: Past, Present, Future, Chichester: Ellis Horwood.
——(1988) ‘Future perspectives in translation technologies’, in M.
Vasconcellos (ed.) Technology as Translation Strategy (American Translators
Association, Scholarly Monograph Series, vol. II, Binghamton, New
York: State University of New York, 223–40.
Hutchins, J. and Somers, H. (1992) An Introduction to Machine Translation,
London: Academic Press.
Iida, H. (1989) ‘Advanced dialogue translation techniques: plan-based,
memory-based and parallel approaches’, in ATR Symposium on Basic
Research for Telephone Interpretation, Kyoto, Japan, Proceedings 8/7–8/8.
Iljon, A. (1977) ‘Scientific and technical databases in a multilingual society’,
in Proceedings of Third European Congress on Information Systems, Munich:
Commission of the European Communities.
Isabelle, P. (1989) ‘Bilan et perspectives de la traduction assistée par
ordinateur au Canada’, in A.Abbou (ed.) Traduction Assistée par Ordinateur:
Perspectives technologiques, industrielles et économiques envisageables à l’horizon
1990: l’offre, la demande, les marchés et les évolutions en cours, Actes du
séminaire international (Paris, March 1988), Paris: DAICADIF, 153–8.
ISO 2788 (1974) ‘Documentation—guidelines for the establishment and
development of multilingual thesauri’.
ISO 6156 (1987) ‘Magnetic tape exchange format for terminological/
lexicographical records (MATER)’.
Jacobs, P., Krupka, G. and Rau, L. (1991) ‘Lexico-semantic pattern
matching as a companion to parsing in text understanding’, in Proceedings
of the DARPA Speech and Natural Language Workshop, Monterey, California.
Jakobson, R. (1959) ‘On linguistic aspects of translation’, in R.A.Brower
(ed.) On Translation, Cambridge, Mass.: Harvard University Press.
 Bibliography 213

Janot-Giorgetti, M.T. (1985) ‘Parol-rekono kun limigita vortaro, gia apliko

en la lernado de parolataj lingvoj’, in I.Koutny (ed.) Perkomputila
tekstoprilaboro, Budapest: Scienca Eldona Centro, 57–68.
JEIDA (1989) ‘A Japanese view of machine translation in light of the
considerations and recommendations reported by ALPAC, U SA’,
Tokyo: JEIDA.
Johnson, R.L. (1983) ‘Parsing—an MT perspective’, in K.Sparck Jones and
Y.Wilks (eds) Automatic Natural Language Parsing, Chichester: Ellis
Horwood, 32–8.
Johnson, R., King, M. and des Tombe, L. (1985) ‘E U ROTRA: a
multilingual system under development’, in Computational Linguistics 11
(2–3): 155–69.
Jones, D. and Tsujii, J. (1990) ‘High quality machine-driven texttranslation’,
in Third International Conference on Theoretical and Methodological Issues in
Machine Translation of Natural Languages, Austin, Texas.
Kaplan, R.M., Netter, K., Wederkind, J. and Zaenen, A. (1989) ‘Translation
by structural correspondences’, Fourth Conference of the European
Chapter of the Association for Computational Linguistics, Manchester,
Proceedings, 272–81.
Karlgren, H. (ed.) (1990) COLING-90: Papers Presented to the 13th International
Conference on Computational Linguistics, Helsinki: Yliopistopaino.
Kat, J.O. de (1985) ‘Traduko el la internacia lingvo al naciaj’, in I. Koutny
(ed.) Perkomputila tekstoprilaboro, Budapest: Scienca Eldona Centro, 259–66.
Katumori, H. and Hukuda, M. (1984) ‘Esuperanto o tyûkai-gengo to suru
kikai-honyaku no kokoromi’, WGNL meeting 1984–7–26.
Kay, M. (1980) ‘The proper place of men and machines’, in Language
Translation, Research Report, Xerox Palo Alto Research Center, Palo Alto,
California.
——(1982) ‘Machine translation’, American Journal of Computational Linguistics,
April-June, 74–8.
Kay, M. and Röscheisen, M. (1988) ‘Text-translation alignment’, Research
Memo, Xerox Palo Alto Research Center, Palo Alto, California.
Kelly, I.D.K. (1978) ‘How Esperanto can aid machine translation’, Computer
Weekly 19 January 1978.
King, M. (ed.) (1987) Machine Translation Today: the State of the Art, Edinburgh:
Edinburgh University Press.
——(1989) ‘Perspectives, recherches, besoins, marchés et projets en Suisse’, in
A.Abbou (ed.) Traduction Assistée par Ordinateur: Perspectives technologiques,
industrielles et économiques envisageables à l’horizon 1990: l’offre, la demande, les
marchés et les évolutions en cours, Actes du Séminaire international (Paris,
March 1988), Paris: DAICADIF, 177–9.
Kingscott, G. (1990) ‘SITE buys B’Vital: relaunch of French national MT
project’, Language International, April.
Kittredge, R. and Lehrberger, J. (1982) Sublanguage: Studies of Language in
Restricted Semantic Domains, Berlin: de Gruyter.
Klein, F. (1988) ‘Factors in the evaluation of MT: a pragmatic approach’, in
M.Vasconcellos (ed.) Technology as Translation Strategy (American
Translators Association Scholarly Monog raph Series, vol. I I),
Binghamton, New York: State University of New York.
214 Bibliography

Knowles, F. (1979) ‘Error analysis of Systran output: a suggested criterion

for “internal” evaluation of translation quality and a possible corrective
design’, in B.M.Snell (ed.) Translation and the Computer, Amsterdam: North-
Holland, 109–33.
Kosaka, M., Teller, V. and Grishman, R. (1988) ‘A sublanguage approach to
Japanese-English machine translation’, in D.Maxwell, K.Schubert and
T.Witkam (eds) New Directions in Machine Translation, Dordrecht: Foris,
109–20.
Koutny, I., Olaszy, G. and Kisfaludy, K. (1988) ‘Esperanto speech synthesis
and its application in language learning’, in T.Szende (ed.) From Phonology
to Applied Phonetics, Magyar Fonetikai Füzutek//Hungarian Papers in
Phonetics, 19, Budapest: Linguistics Institute of the Hungarian Academy
of Sciences, 47–54.
Kudo, I. (1990) ‘Local cohesive knowledge for a dialogue-machine
translation system’, in H.Karlgren (ed.) COLING-90: Papers presented to the
13th International Conference on Computational Linguistics, Helsinki:
Yliopistopaino, vol. 3:391–3.
Kudo, I. and Nomura, H. (1986) ‘Lexical-functional transfer: a transfer
framework in a machine translation system based on LFG’, 11th
International Conference on Computational Linguistics, Proceedings of
COLING-86, Bonn: 112–14.
Laffling, J.D. (1990) ‘Machine disambiguation and translation of
polysemous nouns’, PhD thesis, (CNAA) Wolverhampton Polytechnic.
Lakoff, G. (1987) Women, Fire, and Dangerous Things: What Categories Reveal
about the Mind, Chicago: University of Chicago Press.
Landsbergen, J. (1987a) ‘Isomorphic grammars and their use in the
ROSETTA translation system’, in M.King (ed.) Machine Translation Today:
the State of the Art, Edinburgh: Edinburgh University Press, 351–72.
——(1987b) ‘Montague grammar and machine translation’, in P. Whitelock,
M.MacG.Wood, H.L.Somers, R.Johnson and P.Bennett (eds) Linguistic
Theory and Computer Applications, London: Academic Press, 113–47.
Lawson, V. (ed.) (1982) Practical Experience of Machine Translation, Amsterdam:
North-Holland.
Lehmann, W.P. (1987) ‘The context of machine translation’, Computers and
Translation 2.
Lehrberger, J. and Bourbeau, L. (1988) Machine Translation: Linguistic
Characteristics of MT Systems and General Methodology of Evaluation,
Amsterdam: John Benjamins.
León, M. and Schwartz, L.A. (1986) ‘Integrated development of English-
Spanish machine translation: from pilot to full operational capability:
technical report of Grant DPE-5543-G-SS-3048–00 from the US Agency
for International Development’, co-ordinated by M.Vasconcellos,
Washington, DC: Pan American Health Organization.
Lewis, D. (1985) ‘The development and progress of machine translation
systems, ALLC Journal 5:40–52.
Li Wei (1986) ‘Automata tradukado el la internacia en la cinan kaj anglan
lingvojn (Esperanta-angla/cina masintraduka sistemo)’, Grundlagenstudien
aus Kybernetik und Geisteswissenschaft/Humankybernetik 27:147–52.
Luckhardt, H.-D. (1991) ‘Sublanguages in machine translation; Proceedings
 Bibliography 215

of the Fifth Conference of the European Chapter of the Association for

Computational Linguistics. Berlin: 306–8.
Maas, H.D. (1982) ‘Automata tradukado en kaj el Esperanto’, in H.F.
Yashovardhan and B.Frank-Böhringer (eds) Lingvokibernetiko kaj aliaj
internacilingvaj aktoj de la IX-a Internacia Kongreso de Kibernetiko//
Sprachkybernetik und andere internationalsprachige Akten vom IX. Internationalen
Kybernetikerkongreß, Tübingen: Narr, 75–81.
——(1985) ‘Pri kelkaj strategioj por fraz-analizo’, in I.Koutny (ed.)
Perkomputila tekstoprilaboro, Budapest: Scienca Eldona Centro, 175–205.
——(1987a) The MT system SUSY’, in M.King (ed.) Machine Translation
Today: the State of the Art, Edinburgh: Edinburgh University Press, 209–
46, 392–435.
——(1987b) ‘The Saarbrücken automatic translation system (SUSY)’, in
Overcoming the language barrier, Commission of the European
Communities, Munich: Verlag Dokumentation, vol. 1:585–92.
McCord, M.C. (1989) ‘Design of LMT: a Prolog-based machine translation
system’, Computational Linguistics 15:33–52.
McElhaney, T. and Vasconcellos, M. (1988) ‘The translator and the
postediting experience’, in M.Vasconcellos (ed.) Technology as Translation
Strategy (American Translators Association Scholarly Monograph Series,
vol. II), Binghamton, New York: State University of New York.
McNaught, J. (1988) ‘A survey of termbanks worldwide’, in C.Picken (ed.)
Translating and the Computer 9: Potential and Practice, London: Aslib, 112–29.
Makino, S., Hirata, M. and Katumori, H. (1986) ‘Esuperanto o tyûkaigengo
to suru kikai hon’yaku’, WGNL meeting 1986–3–28.
Malmberg, B. (1967) ‘Los nuevos caminos de la lingüística’, Siglo Veintiuno,
Mexico: 154–74.
Mehta, V. (1971) John is easy to please, New York: Ferrar, Straus & Giroux.
Melby, A. (1982) ‘Multi-level translation aids in a distributed system’, in
J.Horecky (ed.) Proceedings of COLI NG-8 2, Prague, July 198 2,
Amsterdam: North-Holland.
——(forthcoming) ‘Causes and effects of partial asymmetry betweensemantic
networks tied to natural languages’ (a lecture series given at the Collège de
France, Paris, February, 1990), to appear in Les Cahiers de Lexicologie, Paris.
——(forthcoming) ‘MicroMATE R: a proposed standard format for
exchanging lexical/terminological data files’, META (special issue).
Minsky, M. (1986) The Society of Mind, New York: Simon & Schuster.
Mohai, L. (1986) ‘Komputila tradukprogramo el Esperanto hungaren’, in
P.Broczkó, I.Koutny and A.Lukács (eds) Language Cybernetics, Educational
Cybernetics//Lingvokibernetiko, pedagogia kibernetiko//Cybernétique de la langue,
cybernétique de l’éducation, Budapest: Neumann Society, 47–51.
Nagao, M. (1984) ‘A framework of a mechanical translation between
Japanese and English by analogy principle’, in A.Elithorn and R.Banerji
(eds) Artificial and Human Intelligence, Amsterdam, North-Holland, 172–80.
Nagao, M., Tsujii, J. and Nakamura, J. (1985) ‘The Japanese government
project for machine translation’, Computational Linguistics 11:91–110.
Nagel, E. and Newman, J.R. (eds) (1990) Godel’s Proof, London: Routledge.
Newman, P.E. (198 8) ‘Information-only machine translation: a
feasibilitystudy’, in M.Vasconcellos (ed.) Technology as Translation Strategy
216 Bibliography

(American Translators Association Scholarly Monograph Series, vol. II),

Binghamton, New York: State University of New York, 178–89.
Newton, J. (1987) ‘An insider’s view of machine translation’, in The Linguist,
vol. 26, no. 1:21–3.
——(1988) ‘Translating software for the national language’, in R.Sharpe (ed.)
Information Technology Manufacturing—Europe 198 8, London: Sterling
Publishing Group, 302–4. Reprinted in The Linguist, vol. 27, no. 3:102–4.
——(1989) ‘The terminology database in the translation department’, The
Linguist, vol. 28, no. 2:46–8.
Nirenburg, S. (ed.) (1987) Machine Translation: Theoretical and Methodological
Issues, Cambridge: Cambridge University Press.
——(1989) ‘Knowledge-based machine translation’, Machine Translation, 4:5–
24.
Nirenburg, S. and Carbonell, J.G. (1987) ‘Integrating discourse pragmatics
and propositional knowledge for multilingual natural language
processing’, Computers and Translation 2:105–16.
Nirenburg, S. and Levin, L. (1989) ‘Knowledge representation support’,
Machine Translation 4:25–52.
Nishida, F. and Takamatsu, S. (forthcoming) ‘Automated procedures for the
improvement of a machine translation system by feedback from
postediting’, to appear in Machine Translation.
Nishida, F., Takamatsu, S., Tani, T. and Doi, T. (1988) ‘Feedback of
correcting information in postediting to a machine translation system’, in
D.Vargha (ed.) COLING Budapest: Proceedings of the 12th International
Conference on Computational Linguistics, Budapest: John von Neumann
Society for Computing Sciences, 476–81.
Pappegaaij, B.C., Sadler, V. and Witkam, A.P.M. (eds) (1986) Word Expert
Semantics: an Interlingual Knowledge-based Approach, (Distributed Language
Translation 1), Dordrecht: Foris.
Paulos, J.A. (1989) Innumeracy, Mathematical Illiteracy and its Consequences, New
York: Hill & Wang.
Picken, C. (1986) Translating and the Computer 7, London: Aslib.
——(ed.) (1989) The Translator’s Handbook, revised edition. London: Aslib.
Piron, C. (1987) ‘Esperanto: l’immagine e la realità’, in A.Chiti-Batelli (ed.)
La comunicazione Internazionale tra politico e glottodidattica, Milan: Marzorati,
68–116.
Pogson, G. (1989) ‘The LT/Electric Word multilingual wordworker’s
resource guide’, LT/Electric Word 13, Amsterdam: Language Technology
BV.
Pym, P.J. (1990) ‘Pre-editing and the use of simplified writing for MT: an
engineer’s experience of operating an MT system’, in P.Mayorcas (ed.)
Translating and the Computer 10: The Translation Environment 10 Years On,
London: Aslib, 80–96.
Raw, A., Vandecapelle, B. and van Eynde, F. (1988) ‘Eurotra: an overview’,
Interface 3:5–32.
Raw, A., van Eynde, F., ten Hacken, P., Hoekstra, H. and Vandecapelle, B.
(1989) ‘An introduction to the Eurotra machine translation system’,
Working Papers in Natural Language Processing 1, TAAL Technologie,
Utrecht and Katholieke Universiteit Leuven.
 Bibliography 217

Riesbeck, C.K. (1975) ‘Conceptual analysis’, in R.C.Shank (ed.) Conceptual

Information Processing, Amsterdam: North-Holland, 83–156.
Rohrer, C. (1986) ‘Maschinelle Übersetzung mit Unifikationsgrammatiken’,
in I.Bátori and H.J.Web er (eds) Neue Ansätze in Maschineller
Sprachübersetzung: Wissenrepräsentation und Textbezug, Tübingen: Niemeyer
Verlag, 75–99.
Rumelhart, D.E. and McClelland, J.L. (1987) Parallel Distributed Processing,
Cambridge, Mass.: MIT Press.
Rupp, C.J. (1989) ‘Situation semantics and machine translation’, Fourth
Conference of the European Chapter of the Association for
Computational Linguistics, Manchester, Proceedings, 308–18.
Sadler, L., Crookston, I., Arnold, D. and Way, A. (1990) ‘LFG and
translation’, in Third International Conference on Theoretical and Methodological
Issues in Machine Translation of Natural Languages, Austin, Texas.
Sadler, V. (1989) Working with Analogical Semantics. Disambiguation Techniques in
DLT, Dordrecht/Providence: Foris.
——(forthcoming) ‘Machine translation project reaches watershed’, Language
Problems and Language Planning 15.
Saito, H. and Tomita, M. (1986) ‘On automatic composition of stereotypic
documents in foreign languages’, presented at 1st International
Conference on Applications of Artificial Intelligence to Engineering
Problems, Southampton, April 1986; Research Report CMU-CS-86–107,
Department of Computer Science, Carnegie Mellon University.
Santangelo, S. (1988) ‘Making an MT system work: perspective of a
translator’, in M.Vasconcellos (ed.) Technology as Translation Strategy
(American Translators Association Scholarly Monograph Series, vol. II),
Binghamton, New York: State University of New York.
Sapir, E. (1921) Language: An Introduction to the Study of Speech, New York:
Harcourt, Brace & World.
Sato, S. (1989) ‘Phonetic form generation in a semantics-to-speech system of
Japanese’, Sendai: Dissertation, Tohoku University.
Sato, S. and Kasuya, H. (1987) ‘Automatic translation/speech synthesis of
Japanese from written Esperanto incorporating a linguistic knowledge
base editor’, in J.Laver and M.A.Jack (eds) European Conference in Speech
Technology (Edinburgh 1987), vol. 2, Edinburgh: CEP, 414–17.
Sato, S. and Nagao, M. (1990) ‘Toward memory-based translation’, in H.
Karlgren (ed.) COLING-90: Papers presented to the 13th International
Conference on Computational Linguistics, Helsinki: Yliopistopaino, vol. 3:247–
52.
Saussure, F. de (1916) Cours de linguistique générale, Paris: Payot.
Schubert, K. (1986) ‘Linguistic and extra-linguistic knowledge’, Computers
and Translation 1:125–52.
——(1987) Metataxis. Contrastive Dependency Syntax for Machine Translation,
Dordrecht/Providence: Foris.
——(1988) ‘Ausdruckskraft und Regelmäßigkeit. Was Esperanto für
automatische Üb ersetzung geeignet macht’, Language Problems
andLanguage Planning 12:130–47.
——(ed.) (1989a) Interlinguistics. Aspects of the Science of Planned Languages, Berlin/
New York: Mouton de Gruyter.
218 Bibliography

——(1989b) ‘Interlinguistics—its aims, its achievements, and its place in

language science’, in K.Schubert (ed.) Interlinguistics. Aspects of the Science of
Planned Languages, Berlin/New York: Mouton de Gruyter, 7–44.
——(1989c) ‘An unplanned development in planned languages. A study of
word grammar’, in K.Schubert (ed.) Interlinguistics. Aspects of the Science of
Planned Languages, Berlin/New York: Mouton de Gruyter, 249–74.
——(1990) ‘Kunskap om världen eller kunskap om texten? En metod for
korpusstödd maskinöversättning’, in J.Pind and E.Rögnvaldsson (eds) Papers
from the Seventh Scandinavian Conference of Computational Linguistics (Reykjavik
1989), Reykjavik: Institute of Lexicography/Institute of Linguistics, 218–28.
Scott, B.E. (1989) ‘The Logos System’, paper presented at MT Summit II,
Munich.
Shanker, S.G. (ed.) (1987) Godel’s Theorum in Focus, London: Croom Helm.
Sherwood, B.A. (1978) ‘Fast text-to-speech algorithms for Esperanto,
Spanish, Italian, Russian and English’, International Journal of Man-Machine
Studies 10:669–92.
——(1982) ‘Raporto pri sintezo de Esperanta parolado’, in I.Koutny (ed.)
Homa lingvo kaj komputilo, Budapest: NJSzT//Societo Neumann, 64–75.
——(1985) ‘Komputila tradukado de esperanta teksto’, in I.Koutny (ed.)
Perkomputila tekstoprilaboro, Budapest: Scienca Eldona Centre, 153–60.
Sherwood, J.N. and Sherwood, B.A. (1982) ‘Computer voices and ears
furnish novel teaching options’, Speech Technology [1982]: 46–51.
Shreve, G.M. and Vinciquerra, K.J. (1990) ‘Hypertext knowledge-bases for
computer-assisted translation: organization and structure’, in A.L. Wilson
(ed.) Proceedings of the 31st Annual Conference of the American Translators
Association, New Orleans, October 1990: Learned Information, Inc.
Sjögren, S. (1970) En syntax for datamaskinell analys av esperanto, FOA P rapport
C 8264–11 (64), Stockholm: Försvarets forskningsanstalt.
Skuce, D. and Meyer, I. (1990) ‘Computer-assisted concept analysis: an
essential component of a terminologist’s workstation’, in H.Czap and
W.Nedobity (eds) TKE ’90: Terminology and Knowledge Engineering,
Frankfurt: INDEKS Verlag, 187–99.
Slocum, J. (1984) ‘Machine translation: its history, current status, and future
prospects’, in 10th International Conference on Computational Linguistics,
Proceedings of COLING-84, Stanford, California, 546–61.
——(1985) ‘A survey of machine translation: its history, current status, and
future prospects’, Computational Linguistics 11:1–17.
——(ed.) (1988) Machine Translation Systems, Cambridge: Cambridge
University Press.
Smith, D. and Tyldesley, D. (1986) ‘Translation Practices Report’, Reading:
Digital Equipment Corporation.
Snell, B.M. (ed.) (1979) Translating and the Computer, Amsterdam: North-
Holland.
Snell, B.M. (ed.) (1983) Term Banks for Tomorrow’s World, London: Aslib.
Snell-Hornby, M. (1988) Translation Studies: An Integrated Approach,
Amsterdam: John Benjamins.
Somers, H.L. (1986) ‘Some thoughts on interface structure(s)’, in W. Wilss
and K.-D.Schmitz (eds) Maschinelle Übersetzung—Methoden und Werkzeuge,
Tübingen: Max Niemeyer Verlag, 81–9.
 Bibliography 219

Somers, H.L., Hirakawa, H., Miike, S. and Amano, S. (1988) ‘The

treatment of complex English nominalizations in machine translation’,
Computers and Translation 3:3–21.
Somers, H.L., Tsujii, J. and Jones, D. (1990) ‘Machine translation without a
source text’, in H.Karlgren (ed.) COLING-90: Papers Presented to the 13th
International Conference on Computational Linguistics, Helsinki: Yliopistopaino,
vol. 3:271–6.
Steer, M.G. and Stentiford, F.W.M. (1989) ‘Speech language translation’, in
J.Peckham (ed.) Recent Developments and Applications of Natural Language
Processing, London: Kogan Page, 129–40.
Steiner, E. (ed.) (1991) Special issue on Eurotra, Machine Translation, 6:2–3.
Sumita, E., Iida, H. and Kohyama, H. (1990) ‘Translating with examples: a
new approach to machine translation’, Third International Conference on
Theoretical and Methodological Issues in Machine Translation of Natural
Languages, Austin, Texas.
Tarski, A. (1943–4) ‘The semantic conception of truth’, Philosophy and
Phenomenological Research IV: 347.
Taylor, P.J. and Cronin, B. (eds) (1983) Information Management Research
inEurope, London: Aslib.
‘Terminological Data Banks’ (1989) TermNet News 24, Infoterm, Vienna: 25–
30.
Tesnière, L. (1959) Elements de syntaxe structural, 2nd edn, 4th print. 1982,
Paris: Klincksieck.
Thomas, P. (1988) ‘Analysis of an English and French LS P: some
comparisons with English general text corpora’, ALSED-UNESCO
Newsletter 11, 1 (26), 2–10.
Thurmair, G. (1990) ‘Complex lexical transfer in M ETAL’, Third
International Conference on Theoretical and Methodological Issues in Machine
Translation of Natural Languages, Austin, Texas, 91–107.
Toma, P. (1976) ‘An operational machine translation system’, in R.W.
Brislin (ed.) Translation: Applications and Research, New York: Gardner,
247–59.
Trabulsi, S. (1989) ‘Le système Systran’, in A.Abbou (ed.) Traduction Assistée
par Ordinateur: Perspectives technologiques, industrielles et économiques
envisageables à l’horizon 1990: l’offre, la demande, les marchés et les évolutions en
cours, Actes du Séminaire international (Paris, March 1988), Paris:
DAICADIF, 15–27.
Tsujii, J. (1989) ‘Machine translation with Japan’s neighboring countries’, in
M.Nagao (ed.) Machine Translation Summit, Tokyo: Ohmsha, 50–3.
Tsujii, J. and Nagao, M. (1988) ‘Dialogue translation vs text translation—
Interpretation based approach’, in O.Vargha (ed.) COLING Budapest:
Proceedings of the 12th International Conference on Computational Linguistics,
Budapest: John von Neumann Society for Computing Sciences, 688–93.
Uchida, H. and Kakizaki, T. (1989) ‘Electronic dictionary project’, in M.
Nagao (ed.) Machine Translation Summit, Tokyo: Ohmsha, 83–7.
Valentine, T. (1990) ‘Status of machine translation (MT) technology:
Hearing before the Subcommittee on Science, Research and Technology
of the Committee on Science, Space and Technology, US House of
Representatives, 101st Congress, Second Session, September 11, 1990
220 Bibliography

[no. 153], Chairman: Rep. T.Valentine. Washington: US Government

Printing Office.
van der Korst, B. (1989) ‘Functional grammar and machine translation’, in
J.H.Connolly and S.C.Dik (eds) Functional Grammar and the Computer,
Dordrecht: Foris, 289–316.
van Eynde, F. (1988) ‘The analysis of tense and aspect in Eurotra’, in D.
Vargha (ed.) COLING Budapest: Proceedings of the 12th International Conference
on Computational Linguistics, Budapest: John von Neumann Society for
Computing Sciences, 699–704.
van Noord, G., Dorrepaal, J., van der Eijk, P., Florenza, M. and des Tombe,
L. (1990) ‘The MiMo2 research system’, Third International Conference on
Theoretical and Methodological Issues in Machine Translation of Natural
Languages, Austin, Texas.
Vargha, D. (ed.) (1988) COLING Budapest: Proceedings of the 12th International
Conference on Computational Linguistics, Budapest: John von Neumann
Society for Computing Sciences.
Vasconcellos, M. (1984) ‘Machine translation at the Pan American Health
Organization: a review of highlights and insights’, Newsletter, British
Computer Society Natural Language Translation Specialist Group.
——(1985) ‘Theme and focus: cross-language comparison vis translations
from extended discourse’, unpublished PhD dissertation, Georgetown
University, Washington, DC.
——(1986) ‘Functional considerations in the postediting of MT output:
dealing with V(S)O versus SVO’, Computers and Translation 1, 1:21–38.
——(1987a) ‘Postediting on-screen: machine translation from Spanish into
English’, in C.Picken (ed.) A Profession on the Move: Proceedings of Translating
and the Computer 8, London: Aslib.
——(1987b) ‘A comparison of MT postediting and traditional revision’, in
K.Kummer (ed.) Proceedings of the 28th Annual Conference of the American
Translators Association, Medford, New Jersey: Learned Information.
——(1988) ‘Factors in the evaluation of MT: formal vs functional
approaches’, in M.Vasconcellos (ed.) Technology and Translation Strategy
(American Translators’ Association Scholarly Monograph Series, vol. II),
Binghamton, New York: State University of New York.
——(ed.) (1988) Technology as Translation Strategy (American Translators
Association Scholarly Monograph Series, vol. II), Binghamton, New
York: State University of New York.
——(1989a) ‘Cohesion and coherence in the presentation of machine
translation products’, in Georgetown University Round Table on Languages and
Linguistics 1989, Washington, DC: Georgetown University Press.
——(1989b) ‘Long-term data for an MT policy’, Literary and Linguistic
Computing 4, 3:203–13.
Vasconcellos, M. and León, M. (1988) ‘SPANAM and ENGSPAN: machine
translation at the Pan American Health Organization’, in J. Slocum (ed.)
Machine Translation Systems, Cambridge: Cambridge University Press.
Vauquois, B. (1985) ‘The approach of Geta to automatic translation:
comparison with some other methods’, paper presented at International
Symposium on Machine Translation, Riyadh; in Ch. Boitet (ed.) (1988)
 Bibliography 221

Bernard Vauquois et la TAO: vingt-cinq ans de traduction automatique— Analectes,

Grenoble: Association Champollion, 631–86.
Vauquois, B. and Boitet, Ch. (1985) ‘Automated translation at Grenoble
University’, Computational Linguistics 11:28–36.
Vinay, J.-P. and Darbelnet, J. (1977) Stylistique comparée du français et de
l’anglais, Paris: Didier.
Weaver, W. (1955) ‘Translation’, in W.N. Locke and A.D. Booth (eds)
Machine Translation of Languages, New York: Wiley, 15–23.
Wehrli, E. (1990) ‘STS: an experimental sentence translation system’, in
H.Karlgren (ed.) COLING-90: Papers Presented to the 13th International
Conference on Computational Linguistics, Helsinki: Yliopistopaino, vol. 1:76–
8.
Whitelock, P.J., Wood, M.M., Chandler, B.J., Holden, N. and Horsfall, H.J.
(1986) ‘Strategies for interactive machine translation: the experience and
implications of the UMIST Japanese project’, 11th International Conference
on Computational Linguistics, Proceedings of COLING-86, Bonn, 329–34.
Whitfield, F. (1969) ‘Glossematics’, in A.A.Hill (ed.) Linguistics, Voice of
America Forum Lectures.
Whorf, B.L. (1956) Language, Thought and Reality (collected papers),
Cambridge, Mass.: MIT Press.
Wilks, Y. (1972) Grammar, Meaning and the Machine Analysis of Language,
London: Routledge & Keegan Paul.
——(1984a) ‘Machine translation and the artificial intelligence paradigm of
language processes’, Computers in Language Research 2.
——(1984b) ‘Artificial intelligence and machine translation’, in S. and W.
Sedelon (eds) Current Trends in the Language Sciences, Amsterdam: North
Holland.
——(1990) ‘Form and content in semantics’, in Synthèse 82:329–51.
Wilks, Y., Carbonell, J., Farwell, D., Hovy, E. and Nirenburg, S. (1990)
‘Machine translation again?’, in Proceedings of the DARPA Speech and Natural
Language Workshop, Monterey, California.
Witkam, A.P.M. (1983) Distributed Language Translation. Feasibility study of a
multilingual facility for videotex information networks, Utrecht: BSO.
Wood, M.M. and Chandler, B.J. (1988) ‘Machine translation for
monolinguals’, in D.Vargha (ed.) COLING Budapest: Proceedings of the 12th
International Conference on Computational Linguistics, Budapest: John von
Neumann Society for Computing Sciences, 760–3.
Zajac, R. (1990) ‘A relational approach to translation’, Third International
Conference on Theoretical and Methodological Issues in Machine Translation of
Natural Languages, Austin, Texas.
Zeidenberg, M. (1987) ‘Modeling the Brain’, Byte, vol. 12, no. 14:237–46,
New York: McGraw-Hill.
Glossary of terms

Anaphora The use of a word (e.g. a pronoun) to refer to a preceding word

or group of words (e.g. in Peter entered the room. He was wearing a blue
overcoat., The pronoun he is an example of anaphora).
Automatic look-up A function provided in some software packages that
allows the user direct access to a dictionary or glossary from a text file
(i.e. via a window).
Back translation A translation into the original source language of a text
which is itself a translation. This procedure may be used to verify the
accuracy of the original translation.
Batch processing A mode in which a number of files are processed
sequentially by the computer without intervention from the user.
Bit A binary digit, 0 or 1. The minimum element of computer coding
information.
Bottom up processing An approach in which meaning is computed
compositionally from the bottom up (i.e. by combining word meanings
to form larger semantic units). These word meanings are drawn from
lexica containing lists of predefined, distinct word senses. For each of
these senses, one literal meaning is held to be more basic, while
additional, metaphorical meanings are considered to be more
derivative.
Byte A group of bits (usually 8) used to represent individual letters and
symbols of a defined alphabet.
Central processing unit (CPU) The essential part of a computer that
performs all the arithmetical and logical operations.
Clone A microcomputer which is a near replica, in appearance,
configuration and function, of a machine produced by a major
manufacturer.
Coherence The underlying links that establish connectedness within the
communication act.

222
 Glossary 223

Cohesion The overt structural links by means of which the elements in

discourse can be decoded in relation to each other.
Computational linguistics The branch of computer science concerned
with natural-language processing.
Controlled language A circumscribed range of vocabulary and syntax.
May be used to optimize MT performance.
Corpus 1. A collection or body of texts on a specific topic or within a
specific domain. 2. The finite collection of grammatical sentences that is
used as a basis for the descriptive analysis of a language.
Data bank A large store of information in a form that can be processed by
a computer.
Data base, Database The data accessible to a computer.
Dictionary The component of an MT system or translation tools package
within which source and target language lexical items and data relating
thereto are stored. In MT, a dictionary will typically comprise an
indexed file of records corresponding to individual entries, usually each
consisting of a word to be matched against the source text, plus a series
of codes (syntactic, semantic, etc.) and a pointer to a translation gloss,
also accompanied by a series of codes.
Domain-specific language A sublanguage restricted to the vocabulary and
syntax of a specific domain.
Electronic mail The transmission of documents from one computer
terminal to another.
Ellipsis The omission of words that are understood in context but which
would need to be supplied to make a structure grammatically complete
(e.g. the tools that I use may be reduced by ellipsis to the tools I use).
Esperanto An artificial international language based largely on words
common to the main European languages, invented in 1887 by Dr L.L.
Zamenhof.
Flag 1. A character used to signal a condition or characteristic. 2. An
element of data used as an indicator.
General language Unrestricted and non-specialized language.
Gloss 1. A brief explanation of a word or expression. 2. A word or phrase used
as a translational equivalent, traditionally written between single quotes.
Grammar The branch of linguistics that deals with syntax and
morphology.
Homograph 1. A written form functional as two or more parts of speech
and/or representing two or more words. 2. Sometimes used in MT to
describe lexical items which, in their base or inflected forms, can
function as distinct parts of speech or as distinct words.
224 Glossary

Human translation (HT) Translation as performed by a human.

Idiolect A variety of a language peculiar to a specific user.
Idiom 1. A group of words the meaning of which cannot be deduced from
the meanings of the individual words. 2. Sometimes used in MT parlance
to indicate any dictionary entry that consists of more than one word.
Also referred to as a multiple-word entry.
Inflection The change of form that words undergo depending on their
function.
Inflection rules Rules applied in MT through flagging to determine how a
target language word will inflect.
Information-only translation Translation not intended for publication,
produced quickly and cheaply for a specified (and usually specialist)
readership. When generated using MT, minimal post-editing will
normally suffice, as style is not an important consideration.
Interactive system A system which allows the user to interact with the
computer by means of dialogue. This term is used to distinguish MT
systems which solicit information from the user during the translation process
from those which process translations in batch mode. Batch-processing
MT systems usually have interactive modes for functions such as
dictionary updating.
Interlingua Any artificial language used in an MT system to represent the
meaning of natural language.
Look-up see Automatic look-up.
Machine translation (MT) Translation as performed by a computer, with
or without human assistance.
Microcomputer A small computer, typically intended for a single user, in
which the central processing unit is contained in one or more silicon chips.
Microprocessor A single integrated circuit that functions as the central
processing unit of a microcomputer.
Morpheme A meaningful linguistic unit that cannot be subdivided into
further such units.
Morphology The study of the structure of words and their inflectional,
derivational and combinational patterns.
Natural language 1. A naturally evolved language that is the native speech
of a people. 2. Languages of this kind viewed collectively.
Natural language processing The branch of computer science concerned
with the machine processing of natural language(s).
Neural network A computing configuration in which a large number of
processing units operating in parallel (parallel processing units) are
 Glossary 225

interconnected in a complex manner thought to resemble the

connections between brain cells.
Output see Raw translation.
Page In the United States it is frequently considered that the translation page
is equivalent to 250 words.
Parse To assign constituent structure to a sentence and ascertain the
grammatical relationships that exist between its component words.
Parser A computer program which parses sentences. In the case of MT, the
parser is the engine that drives the analysis component.
Parsing see Parse.
Paste-in, Pasting-in The copying of data directly into an open text file from
a database via a window.
Personal computer (PC) see Microcomputer.
Platform A computer and operating system regarded as the fundamental
components of a more complex whole such as a translator workstation.
Postediting, Post-editing, Post editing Editing performed for the purpose
of improving the quality and acceptability of raw MT output.
Posteditor, Post-editor, Post editor A person who undertakes the task of
post-editing.
Post-modern philosophy Wittgenstein [later works], Heidegger, Husserl,
Levinas.
Pre-analysis The process in MT during which the computer breaks the
source text down into the smallest units it can analyse.
Pre-editing The editing of a source text prior to its submission to MT. The
pre-editor seeks to eliminate ambiguity and confusion through structural
and lexical modifications, thereby optimizing the system’s performance.
A pre-edited English-language source text will typically contain more
prepositions and conjunctions and have shorter sentences than the
unedited original; this is intended to assist the computer’s analysis.
Pre-editor A person who undertakes the task of pre-editing.
Raw translation The raw, unedited output of an MT system.
Register A form of a language associated with a particular situation, set of
circumstances, subject matter or purpose.
Restricted language see Controlled language.
Reverse translation see Back translation.
Semantic network A collection of points, called nodes, each of which
represents a concept. Each node has a name (e.g. fair-haired girl, black
226 Glossary

dog) which is connected by a directed arc (or arrow), called a relation, to

another node. The relation has a label.
Semantics The branch of linguistics concerned with meaning.
Source language The language in which a text is originally written and
from which a translation is to be or has been produced.
Source text The source language text from which a translation is to be or
has been produced.
Special language Special language has features in common with both
general language and sublanguage. It combines relative structural
freedom with specialized vocabulary. Scientific reports and technical
documentation fall into this category.
Split-screen editing A post-editing mode of MT that allows the user to view
and edit a segment of raw translation whilst viewing the corresponding
segment of the source text.
Staging Point of departure for the clause, sentence, paragraph, episode or
discourse.
Standard Generalized Markup Language (SGML) A highly flexible
mark-up language devised to make document structure descriptions
available in a universal format. Wide adoption of SGML document
descriptions will facilitate the transfer of texts between different word
processing formats. SGML became an international standard (ISO 8879)
in 1986.
Sublanguage A variety of a language having restricted grammar and
vocabulary, such as would facilitate MT analysis.
Syntax A set of rules governing sentence construction and the grammatical
arrangement of words.
Target language The language into which a translation is to be or has been
produced.
Target text The target language text which is to be or has been produced
through translation.
Term A word or expression used to designate a specific object, action or
process, particularly in a specialized field of knowledge.
Term bank see Terminology data base, Terminology database.
Terminology The body of specialized words and terms relating to a
particular domain; also the study of such words and terms.
Terminology data base, Terminology database A database dedicated to
the storage and retrieval of terminology.
Terminology management The act of maintaining and updating a
terminology database.
 Glossary 227

Text encoding initiative The goal of the TEI is to develop and disseminate
a clearly defined format for the interchange of machine-readable texts
among researchers so as to allow easier and more efficient sharing of
resources for textual computing and natural-language processing. The
interchange format is intended to specify how texts should be encoded or
marked so that they can be shared by different research projects for
different purposes. The use of specific delimiter characters and specific
tabs to express specific information are all prescribed by this interchange
format, based on the international standard I SO 8879, Standard
Generalized Markup Language (SGML).
Translation 1. A rendering of a text from one language into another. 2. The
product of such a rendering.
Translation tools Software packages which are designed to assist translators
in their work but which do not perform translation. The main
constituents of a translation tools package are a text analyser (parser) and
a user-updatable dictionary. The text analyser is used to identify lexical
items for dictionary entry. Such products usually allow translation
equivalents to be pasted directly into a text file via a window during
dictionary look-up.
Translator’s workbench see Translator workstation.
Translator workstation A custom-built, economically designed unit having
a computer as its main platform, incorporating translation tools and
supporting a potentially wide range of software utilities and peripheral
devices. Translator workstations offer various levels of computer
assistance up to and including MT. They can also incorporate fax,
electronic mail and desk-top publishing facilities, and offer remote access
to terminology databases.
Vocabulary 1. A listing, selective or exhaustive, of the lexical items of a
language or used by a group or individual or within a specialized field of
knowledge, with definitions or translation equivalents. 2. The aggregate
of lexical items used or understood by a specified group, class,
profession, etc.
Vocabulary search A function of MT and translation tools packages in
which the system compares the words in a text with those that figure in
a specified dictionary or sequence of dictionaries and copies unfound
words into a file which can later be used as a basis for dictionary
updating.
Window A feature of certain software packages that allows the user to
access utilities from within a text file.
Wordage The number of words contained in a text.
Word count, Wordcount see Wordage.
Index

Note: Numbers in brackets refer to notes. Arrow Technical Translations 141

artificial intelligence (AI):
accuracy, claims for MT 98–9, improvements 118–19;in
100, 103 Japanese MT 141–2;and MT
ACL (Association for 127–8 (n21), 192–3;semantic
Computational Linguistics) representation 143;and
157, 202 SYSTRAN 168, 169
Advanced Telephony Research artificial symbols, and language
(ATR), Japan 192, 198, 200 83–4
affixes 201 ASCII (American Standard Code
Ahmad, K. 140 for Information Interchange) 40,
AI see artificial intelligence 138
AIRSPEAK 143 Aslib (Association for Information
Alam, Y.S. 194 Management) 132
ALPAC report Automatic Aslib Conference 202
Language Processing Advisory Association for Computational
Committee) 16–17, 19, 24, 171 Linguistics (ACL) 157, 202
alphabets, non-Roman 25, 140, Association for Computing in the
141 Humanities 157
ALPNET 22, 163 ATLAS 142
Alvey programme 20 ATR see Advanced Telephony
ambiguities:semantic 90–1; Research
syntactic 88–89 automatic grammar updating 196–7
American Translators Association automatic look-up 2, 162–3
164
Amores Carredano, J.G. 194 Bar-Hillel, Y 79
analogy-based machine translation batch processing, MT 5, 48, 54, 97,
88 98
Ananiadou, S. 196 Battelle Memorial Institute report
anaphora 53, 192, 193 170, 172–3
Apple Macintosh 154, 156 Ben-Avi, S. 79
AQUILA 141 Bennett, W.S. 24
ARIANE 20, 125 (n6), 189 bilingual corpora 166, 167

228
 Index 229

Blanke, Detlev 92–3 Centre Canadien de Recherche en

Bloomfield, Leonard 105–8, 125 Informatisation du Travail
(n8) (CCRIT) 24
Boitet, C. 141, 189 Centre for Computational
Bostad, D. 9 Linguistics (CCL) UMIST 20–
Bourbeau, L. 190 1, 195, 196, 198, 200
Briem, S. 80 CGIAR (Consultative Group on
Brigham Young University International Agricultural
Translation Research Group 164 Research) 71–2
British Telecom Research Chandioux, John 7, 8, 23, 33–5,
Laboratories 21, 132, 133, 198 194
Brown, P.P. 166, 200, 201 Chandler, B.J. 193
BSO 21, 80, 87 Chevalier, M. 189
Buro voor Systeemontwikkeling Chinese characters 141
(BSO) 21, 80, 87 Chisholm, S. 126 (n12)
Chomsky, Noam 107–9, 126 (n10)
Canada:CCRIT 24;Digital CIAT (International Center for
Equipment of Canada Ltd 42– Tropical Agriculture) 71
4;METEO 33–9, 194;MT 23– CICC (Center for International
4;TERMIUM 132, 133; Cultural Cooperation) (Japan)
Waterloo University 159 27
Canadian National Railways 44 cloze technique 172
Carbonell, J.G. 20, 25, 124 (n3), CMU (Centre for Machine
192 Translation at Carnegie Mellon
Carroll, J.B. 171 University) 24, 192
Carroll, J.J. 200 code-breaking, and MT 16, 96–7,
case-based techniques 169 131
CAT (computer-assisted COLING (International
translation) 97 Conference on Computational
categorial grammar 193, 194 Linguistics) 202
Caterpillar Fundamental English Collins Electronic Dictionaries
(CFE) 47 140–1
CCL (Centre for Computational Colmerauer, Alain 34
Linguistics) UMIST 20–1, 195, combinatorial explosion 81–3
196, 198, 200 Commission of the European
CCRIT (Centre Canadien de Communities (CEC) 17, 132–3,
Recherche en Informatisation 134
du Travail) 24 compositionality 90, 190, 199–
CD-ROM 133;databases 71; 200;and Esperanto 89, 90, 91;in
dictionaries 141 translation 190, 199–200
CEC (Commission of the computational linguistics:
European Communities) 17, Chomsky 107–9;and corpora
132–3, 134 87–8;and Esperanto 79–81;
Celt, S. 127 (n15) EUROTRA 142;Hays, David
Center for International Cultural 152;PAHO environment
Cooperation (CICC) (Japan) 27 64;progress 25–6;theories 193–4
Center for Machine Translation, computer-assisted translation
Carnegie Mellon University (CAT) 97
(CMU) 24, 192
230 Index

computerised translation tools 2, deterministic labelled-tree

12–13, 18 transducer 35
concept-based terminology files 161 dialogue-based MT 20, 197–9
concordancing 140, 143 dictionaries:and accuracy 100;
conditional match holes 50 back-up 131;bilingual document
Confederation Life Insurance 2;building 48–53; deeply-coded
Company 39–42 59;electronic 27, 131, 132, 140–
conferences on MT 202 1; ENGSPAN 70–1;SPANAM
CONJUGUONS! 45 70–1;and translator 70–1;
Conseil Supérieur de la Langue updating 12, 54, 173
Française 45 dictionary sequencing 52–3
consistency:Esperanto 80; human dictionary tools 140
translation 116; terminological Dietze, J. 79
5–6, 12, 47, 155; vocabulary Digital Equipment of Canada Ltd
100–1 42–4
constrained domains 194 DIGITRAD 42–4
contrastive linguistic knowledge disambiguation 85, 116–17, 191,
143, 190, 196 198;see also ambiguities
controlled language 46, 55, 98 discontinuous phrasal verbs 71
core grammar, priming 201 discourse pragmatics 192
corpora 87–8;bilingual 166, disk storage 155
167;structured 90 DLT (distributed
corpus-based MT 199–201 languagetranslation) :Esperanto
costs, MT 101–2, 104 80–1, 124 (n3);feasibility study
CP/M 154 88–9; knowledge banks 91;and
Cromemco, and GramR 35 monolingual user 85–6;and
cross-cultural difficulties 111–13 post-editing 85–6; sponsorship
cryptography, MT 16;see also code- 21
breaking documentalists, and phraseology
69
DANTERM 134–5, 136–7 domain knowledge 192–3
Danzin, A. 142 domain-specific networks 150–1
Darbelnet, J. 117 domains:classified 132; constrained
DARPA 25 194;delimited 134;specific 133–
databases 132;on CD-ROM 71; 4, 150–1
ENGSPAN 71;indexing Drexler, E.K. 128 (n24)
programs 160;relational 138–9, DTP (desk-top publishing) 141
160;for terminology of previous Durand, J. 193
translations 163 Dymetman, M. 194
datacommunications networks,
international 80 Eastern Europe, MT 23
Datta, J. 124 (n5) ECHO (European Commission
declarative linguistic formalisms Host Organisation) 133
190 EDITSYS 73–6
Denisov, P.N. 79 Electronic Dictionary Research
dependency syntax 79 (EDR) Project 27
dependency trees 199 Elliston, J.S.G. 194, 195
desk-top publishing (DTP) 141 ENGSPAN 58, 59, 62–3, 64, 76
(n3)
 Index 231

ENGSPAN dictionaries 70–1 Fodor, J.A. 126 (n12)

Ericsson Vertriebs-Partner 140 font generation 140
error-calculating transmission Foreign Technology Division
protocols 158 (FTD), US Air Force 72–6; MT
errors:avoidance 115;detection systems, assessment 166–
158;levels 98–9;MT 9–10, 12, 87;post-editing 73–4
115;repeated 196–7 format codes 157
Essex, University of 20, 174 Forster, P.G. 78
Esperanto:agglutination 89; frame-based conceptual interlingua
development 91–2;and MT 79– 192–3
81, 86–93;prejudices 92–3; and France:GETA 20, 33;MT 19–
system design 85–6 20;SITE 141; standardisation of
Esperantology 89–90 terminology 134;term banks
ESPRIT II 140 133
EURATOM 17 FRANTEXT 133
EURODICAUTOM 132, 133, free syntax 59
134–5 Fulford, H. 134
Europe, commercial companies in fully automated high quality
MT 19–23 translation (FAHQT) 17–18, 98,
European Commission, databanks 117
132–3, 134 function words 89
European Commission Host
Organisation (ECHO) 133 General TAO 39–42
EUROTRA 142;launch 17, 19; and generative grammar 151
linguistic models 190, 193; Germany:LEXIS 132;MT 20;
transition programme 22–3 standardisation of terminology
EUSIDIC (European Association 134;SUSY 20, 79;TEAM 140
of Scientific Information GETA (Groupe d’études pour la
Dissemination Centres) 132 traduction automatique) 20, 33
evaluators (SYSTRAN) 170–2, 174, gisting 74–5
175;bilingual 171, 182–3; Globalink 22
monolingual 171, 182–3; glossaries 100–1, 118, 133–4, 140
variance 180, 181–2, 185 Glossematics 113, 127 (n19)
example-based machine translation glosses 70
199, 200 Godel, K. 127 (n20)
expert systems 196 Gordos, G. 80
experts, roster 162 grammar:categorial 193, 194; core
expressiveness, in language 84, 85, 201;generative 151; head-driven
86 phrase structure 194;lexical
extrapolation 88 functional 193; logic 194;
Montague 143, 193;tree-
FAHQT (fully automated high adjoining 194; universal 108
quality translation) 17–18, 98, grammar checkers 6, 120
117 grammar updating 196–7
Farwell, D. 24, 194 GramR 35;other developments of
fax machine 1, 158 39–44;transformational rules 36
fidelity 171–3, 176–8 graphical user interfaces (GUI)
filters 160 155–6
flags, flagging 49, 70, 73 Greenberg, J. 127 (n14)
232 Index

Grimaila, Annette 8 idioms 50, 54;and distinctions 51–2

Grimes, J.E. 69 Iljon, A. 132
Grosjean, F. 125–6 (n9) ILSAM (International Language
Gross, A. 7, 9, 10 for Service and Maintenance) 47
GUI (graphical user interfaces) indexing programs, databases 160
155–6 indicative MT 148
Guzmán de Rojas, I. 126–7 (n13) inflection rules 49
Infoterm 138, 164
Halliday, M.A.K. 69 INK Text Tools 140
HAMT (human-assisted machine input text:see source text
translation) 18, 19 intelligibility 178–9
Harrap’s on CD-ROM interactive machine translation 74–
Multilingual Dictionary 5, 97, 98, 191
Database 141 INTERDOC 140
Hartley, A. 56 interlingual approach 97–8, 151,
Hauenschild, C. 194 192–3
Hays, David G. 152 intermediate language 97–8;and
head-driven phrase structure MT 81–5;requirements 83–5; see
grammar 194 also Esperanto
Hidden Markov Models 119, 121, international communications 91
128 (n23) International Conference on
highlighting, for post-editor 73–4 Computational Linguistics
Hjelmslev, Louis 84, 87, 127 (n19) (COLING) 202
Hockett, C.F. 105, 107, 125 (n8), International Rice Research
126 (n11) Institute 71
Hodges, A. 124 (n1) international standard terms 159
homographs 48, 50, 51–2 International Standards
HT see human translation Organization Technical
Huang, X. 198 Committee (ISO TC) 138
Hukuda, M. 80 IPSS (International Packet Switch
human-assisted machine translation Stream) 133
(HAMT) 18, 19 Isabelle, P. 24, 194
human translation:macros 118; ISSCO 21
meaning inference 5;and MT,
compared 3–7, 49–51, 64–5, Jacobs, P. 169
102;omissions 6–7; Janot-Giorgetti, M.T. 80
perfectionism 116, 117;see also Japan:and AI 141–2;ATR 192, 198,
HAMT 200;JICST 27;MITI 27;MT 17,
Hunn, E.S. 127 (n17) 25–8, 141–2, 143
Hutchins, W.J.:and AI 192; future Japanese characters 141
of MT 29;generative grammar JEIDA (Japanese Electronic
151;history of MT 15, 124 Industry Development
(n1);language and syntax 143 Association) 27
hypercorrections 70 JICST (Japan Information Centre
hypertext cross-reference links 159– of Science and Technology) 27
60 Johnson, R.L. 192
Jones, D. 198
IBM 154, 166, 167, 200–1 journals, MT 202
idiom holes 50
 Index 233

Kakizaki, T. 27 LIDIA 141

Kaplan, R.M. 194 linguistic expert system 196
Kasuya, H. 80 linguistic machine editor 142
Kat, J.O. de 80 linguistic rule-writing formalisms
Katumori, H. 79–80 190
Kay, M. 122, 147, 201 linguistic theories 104, 193–4
Kelly, I.D.K. 79 linguistics:transformational 126
Kermit 158 (n11);see also computational
keyword searches 162 linguistics
King, M. 21 LISP 34
Kingscott, G. 125 (n6) literal translations 190
Kittredge, R. 195 literary text 149
Klein, F. 125 (n5) logic grammar 194
knowledge bases 141–2, 160–1, Logos 22
164 look-up, automatic 2, 162–3
knowledge-based MT 192 LT/Electric Word 22, 145
Knowles, F. 131
Kosaka, M. 195 Maas, H.D. 20, 79
Koutny, I. 80 Mac II 154, 156
Kudo, I. 192, 194 machine-assisted human translation
Kyoto University 27, 142 see MAHT
machine-assisted translation (MAT)
Lakoff, G. 151 139–42
Landsbergen, J. 21, 193 machine translation:see MT
language:and artificial symbols, Machine Translation 202
compared 83;combinatorial magnetic tape exchange format
capacity 89;deep structure 108, (MATER) 138
109;developing 92; functions MAHT (machine-assisted human
110;general 149; and neutrality translation) 18, 19
110; philosophy 150;and reality Makino, S. 80
5, 7, 121, 122;relativistic 107;see Malmberg, B. 127 (n19)
also DLT;sublanguage MAPTRAN 197
language barriers, removal 96–7, mark-up language 157
109 Markov Models, Hidden 119, 121,
language-independent-systems 143 128(n23)
languages, non-Roman-alphabet 25, Massachusetts Institute of
140, 141 Technology (MIT) 16
Lehrberger, J. 189, 195 MAT (machine-assisted translation)
León, M. 59 139–42
Levin, L. 192 MATER (magnetic tape exchange
lexical functional grammar (LFG) format) 138
193–4 McClelland, J.L. 128 (n22)
lexical gaps 193 McCord, M.C. 194
LEXICOM 133 McElhaney, T. 69
lexicon modification 169 McNaught, J. 140
LEXIS 132 meaning inference 5
LFG (lexical functional grammar) Mehta, V. 126 (n11)
193–4 Melby, A. 2, 7, 138, 147, 150
Li Wei 80
234 Index

memory-based machine translation MT projects, current 189–202

199–200 MTX (TERMEX) 140
METAL 20 multilingual word-processing
metaphors 149 packages 18, 22, 23
METEO:background 23–4; Multilingual Wordworker’s
bulletin sample 37–9; Resource Guide 22
Chandioux, John 33–5; GramR multiple language, simultaneous
35, 36;sublanguage text 148–9, display 156
194, 197;success 34, 36–7 multiple translation equivalents 52
Meyer, I. 139 multiple-word entry 49–51, 54
MicroCAT:implementation 54– multiword phrases 70
5;dictionary 48–53;post-editing must match holes 50
54;sublanguage 194–5
MicroMATER 138, 164 Nagao, Makoto 25, 27, 198, 199,
MicroSoft Windows 154, 155 200
Minitel 133 Nagel, E. 127 (n20)
Minsky, Marvin 112, 127 (n16), nanotechnology 119, 128 (n24)
127–8 (n21) National Science Foundation
mistranslations 9–10, 12, 115, 131 (NSF), USA 24–5
MIT (Massachusetts Institute of natural language see language
Technology) 16 natural language processing (NLP)
MITI (Ministry of International 27–8, 87–8
Trade and Industry), Japan 27 Nedobejkine, N. 189
modems 132, 158 Netherlands:BSO 21, 80, 87; DLT
modularity 11, 190 21, 80–1, 85, 87; INK Text
Mohai, L. 80 Tools 140;MT 21; Philips
monolingual user 20–1, 141, 198 143;Tron 141
Montague grammar 143, 193 networks, local area 155
Montreal, University of 23 neural nets 119, 128 (n22), 151
morphological analysis 162–3, 201 New Mexico State University 24,
morphological search and replace 25
156–7 New York University 24, 25
MT (machine translation) 7–8, 9– Newman, P.E. 124 (n5), 127 (n20)
11, 15–17; accuracy 98–9, 100, Newton, J. 7
103;and AI 127–8 (n21); and Nirenburg, S. 24, 124 (n3), 192,
Chomsky 109; conferences 202; 202
corpus-based 199–201;costs 65, Nishida, F. 196
66, 101–2, 104;defence NLP (natural language processing)
applications 72;Europe 19– 27–8, 87–8
23;future 28–30, 198;HT, NoTe-Bergen 132
compared 3–7, 8–9, 64–5, 102; Nomura, H. 194
indicative 148;Japan 17, 25–8, non-Roman scripts 25, 140, 141
141–2;knowledge-based 192;and NORMATERM 133, 134
MAT prestige 78; raw 75;scope Norway, NoTe-Bergen 132
xvii-xviii, 7–8, 17–19, 99–104, NTRAN 20–1, 191, 193, 197–8
151;and term banks 142–5;USA
24–6 objectivism 151, 152
MT developers, and translators optical character reader/recognition
115–18 (OCR) 2–3, 66, 140, 142–3
 Index 235

ORTHOGRAPHE MODERNE Pym, Peter 46, 47, 56, 195

45
ORTOGRAF+ 45 Q-SYSTEMS 34, 35
Ottawa, University of 24
Raw, A. 19, 190
PACE (Perkins Approved Clear raw translation 4, 7, 8, 9, 48, 49,
English) 46–8, 53, 54–5 50, 53, 54, 69–70, 73, 117
PAHO see Pan American Health real-world knowledge 4, 7, 120,
Organization 143, 191
PAHO-specific terminology 70 record formats 134, 135, 138, 139
Pan American Health Organization record layouts, externally defined
(PAHO) 58–72 159, 160
Papegaaij, B.C. 21, 80 register 5, 69
parallel processing techniques 143 relational database 138–9
parsers 59, 89, 169 research, databases 158
partially edited machine translation retrieval:intelligent 139;
73 synchronized 163; term banks
PC-DOS 154, 155 132
PECOF (Post-Editor’s Correction Rohrer, C. 194
Feedback) 197 Röscheisen, M. 201
Perkins installation:dictionary Rosetta 193
updating 48–53;homographs rule-writing formalisms, linguistic
51–2;PACE 46–8, 53, 54–5; 190
Pym, Peter 46, 47, 56, 195 Rumelhart, D.E. 128 (n22)
PHENIX 141 Rupp, C.J. 194
Philips Research Laboratories 21, Russian-English 168–9, 184
143
phrase structure 201 Sadler, L. 194
phrase-building capability 71 Sadler, Victor 80, 87
Piron, C. 78 Saito, H. 198
POLICESPEAK 143–4 Santangelo, S. 69
political texts 168, 169 Sapir, Edward 105, 106, 125 (n7)
post-editing 7, 8, 68–70;and DLT Sato, S. 80, 199, 200
85–6;interactive 196–7; at Schubert, Klaus 3, 21, 79, 80, 84,
Perkins 53, 54;and pre-editing 87, 89
98 Schwartz 59
post-editors:dictionary updating searching:EURODICAUTOM
64;and FTD MT 73;outputs 133; keyword
67;recurrent structural problems 162;morphological 156–7
and 55, 71;role 58 SEASPEAK 143
pre-editing 18, 53, 98;PAHO 59 second-generation architecture 189–
pre-processing, text 162 90, 191
preferred usage, term banks 132 secondary responses to language
priming, expert systems 196 105–6, 107, 108, 117–18
PROLOG 34, 94 semantic atoms 90
pronouns, correct resolution 192 semantic net 140
prototypical texts, database 157 semantic networks 150–1
PSS Dialplus, BT 133 semantic primitives 90
Pugh, J. 3
236 Index

semantics:and AI 143; ambiguities structure-preserving translation 190

90–1; structured text 159
situation 194 sentence pattern rules style-checking 120
47 subject classification 144–5 (n3)
sets, theory 113–15 sublanguage 194–5;research
SGML (standard generalized mark- 196;text 148–9;see also METEO
up language) 157–8 sublanguage plus 196
Sherwood, B.A. and J.N. 80 suffixes 201
Shreve, Professor Gregory 157, 164 Sumita, E. 200
SITE, France 141 SUSY (Saarbrücker
situation semantics 194 ÜbersetzungsSYstem) 20; and
Sjögren, S. 79 Esperanto 79
Skuce, D. 139 Sweden, TERMDOK 132
Slocum, J. 24, 25, 192 Switzerland, MT 21
Smirnov-Trojanskij, Petr 79 synonyms/antonyms 138–9
Smith, D. 134 syntactic analysis 142–3, 201
Snell-Hornby, M. 149 syntactic errors 70
Somers, H.L. 3, 190 syntactic representation, universal
source text 68, 69–70; 79
electronically compatible 161– syntactic trees, partial 199
4;homogeneity 47, 55; syntagma 88–9
structured 159;styles 59;and syntax:ambiguity 88–9;
terminology file 162 dependency 79;free 59;
SPANAM 58, 59–63, 64, 76 (n3) METEO 37–9;rules 47;
SPANAM dictionaries 70–1 systemic functional grammar
speaker-assisted speech-processing 194
systems 119 SYSTRAN:and AI 168, 169;
special-language texts 149–50 constrained domains 194;in EC
speech synthesis 143–4 17, 22, 150, 167–8; and FTD
speech processing 80, 119 166–87; methodology 173–
spelling, reformed 45 82;success 166–7;as transfer
spelling checkers 6, 156 system 169
split-screen editing mode 8
staging 69 Takamatsu 196
Standard Generalized Mark-up Tandem communications node 35
Language (SGML) 157–8 TAUM group 33–4
standardization: international TEAM, Germany 140
159;terminology 134, 159; technical glossary 100–1
translation 100–1 technical terminology 69
statistics-based machine translation technical writers 161
200–1 TED, text editor 140
Steer, M.G. 21 TEI (text encoding initiative)
Steiner, E. 19, 190 157–8
Stentiford, F.W.M. 21, 144 telecommunications 158
stop words 162 tenses 193
stratificational approach 190–1 term acquisition 134–8
structural problems, recurrent 55, term banks 2, 131–44; compiling
71 133–4;exchange of data
structural transfer 190–1 138;functions 132; future 138–
 Index 237

9;hierarchical structures translation:analogy-based 88;

139;multilingual 131;preferred example-based 87–8, 200;full-
usage 132 dress 98, 100;history 1–3;
term record formats 134, 135, 138 information-only 72–6, 98, 100;
TERMDOK 132 literary 111;market 150;
TERMEX (MTX) 140 memory-based 199–200;for
terminological consistency 5–6, 12, publication 148–52;raw 69–70,
47, 155 73, 117;speed 75, 99–
terminological data-exchange 100;statistics-based 200–1
format, universal 164 translation equivalents 52, 145
terminologists 64, 92, 121, 163–4; (n4), 200–1
and semantic nets 140 translation tools 2, 12–13, 18
terminology:concept-based 161; translator workstations:see
graphic representation 139; workstations
management 149, 158–61; translators:and dictionaries 70–1;
relationships 143 as experts 114–15;and MT
terminology data banks:see term developers 115–18; and MT
banks output 68–70;and term banks
terminology file, for source text 134, 140;see also post-editors
162 trees:dependency 199;partial
TERMIUM 132, 133;domains 135 syntactic 199, 200;transducer 35
tertiary responses to language 106 Tron, Netherlands 141
Tesnière, Lucien 79 Tsujii, J. 27, 198
TEXAN 140 Turing, Alan 96, 124 (n1)
Texas, University of 24 Tyldesley, D. 134
text analysis 2, 139, 140, 161–3
text editing 140 Uchida, H. 27
text encoding initiative (TEI) 157– Uldall, H.J. 127 (n19)
8 UMIST (University of Manchester
text-oriented MT 20 Institute of Science and
text statistics, IBM 166 Technology) 20–1, 195, 196,
text-searching 162 198, 200;monolingual user
text-types 7–8, 12, 195; see also system 141
source text United Kingdom:British Telecom
third generation 143, 192 Research Laboratories 21, 132,
Thomas, P. 2 133, 198;MT 20–1;Perkins
Thurmair, G. 20 installation 46–57;UMIST
Toma, P. 166, 169 (CCL) 20–1, 141, 195, 196,
Tomita, M. 124 (n3), 192, 198 198, 200
totalizator program 174 universal grammar 108
Trabulsi, S. 22 universal terminological data-
transducer, deterministic labelled- exchange format 164
tree 35 Unix 155, 156
transfer approach 98, 151 updating:dictionaries 12, 54, 173;
transfer systems 81, 197 term banks 132
transformational linguists 126 (n11) USA:CMU 24, 192;FTD 72–6,
transformational rules 36 166–87;MT 16, 24–6;PAHO
translating machine, hand-held 10, 58–72
144 user-interface, standard 155
238 Index

USSR, standardization of Wilks, Y. 4, 24, 128 (n21), 167,

terminology 134 194
utilities, switching between word- Witkam, A.P.M. 80, 88
processors 157–8 Wood, M.M. 193
word class system 89
Valentine, T. 25 word equivalences 200–1
van der Korst, B. 194 word frequency programs 140; see
van Eynde, F. 193 also concordances
van Noord, G. 194 word processing 1, 153, 156–8
Vasconcellos, M. 9, 18, 65, 69, word senses, lexicography 152
124–5 (n5) word strings 114
Vauquois, Professor Bernard 33, word-based knowledge banks 90
34, 189 word-collocations 169
verbs, discontinuous phrasal 71 word-formation, productive 89–90
views, terminology 160 workstations 144;advanced
Vinay, J.P. 117 164;affordability 153; level-one
vocabulary search 48, 55 156–61;level-two 161–4;level-
voice writers 119, 120 three 164; operating system
153–5;Unix 155
Waterloo, University of 159 world knowledge 4, 7, 120, 143,
Weaver memorandum 15–16 191
Wehrli, E. 194
Weidner:see MicroCAT X-modem 158
Whitelock, P.J. 20–1, 193 X-windows 156
Whitfield, F. 127 (n19) Xerox Corporation
Whorf, B.L. 105, 106, 125 (n7)
Whorf-Sapir hypothesis 113, 125 Zajac, R. 194 154
(n7)