MACHINE TRANSLATION: An Introductory Guide

Next: Preface

MACHINE TRANSLATION: An
Introductory Guide
Doug Arnold
Lorna Balkan
Siety Meijer
R.Lee Humphreys
Louisa Sadler

Here you will find the text of the book in a Web Browsable form. A printable (PostScript)
version of the book can be downloaded by following this link.

This book is copyright the authors. If you want to do more than read it or copy it for personal
use, you need permission. You can get this by mailing [email protected]

● Preface
● Contents
1. Introduction and Overview
2. Machine Translation in Practice
3. Representation and Processing
4. Machine Translation Engines
5. Dictionaries
6. Translation Problems
7. Representation and Processing Revisited: Meaning
8. Input
9. Evaluating MT Systems
10. New Directions in MT
● Useful Addresses
● Glossary
● References
● Index
● About this document ...

http://privatewww.essex.ac.uk/~doug/book/book.html (1 de 2) [20/09/2000 18:02:17]

MACHINE TRANSLATION: An Introductory Guide

Arnold D J
Thu Dec 21 10:52:49 GMT 1995 [accessed times]

http://privatewww.essex.ac.uk/~doug/book/book.html (2 de 2) [20/09/2000 18:02:17]

Preface

Next: Contents Up: Machine Translation Previous: Machine Translation

Preface
Automatic translation between human languages (`Machine Translation') is a
Science Fiction staple, and a long-term scientific dream of enormous social, political, and
scientific importance. It was one of the earliest applications suggested for digital computers, but
turning this dream into reality has turned out to be a much harder, and in many ways a much
more interesting task than at first appeared. Nevertheless, though there remain many
outstanding problems, some degree of automatic translation is now a daily reality, and it is
likely that during the next decade the bulk of routine technical and business translation will be
done with some kind of automatic translation tool, from humble databases containing canned
translations of technical terms to genuine Machine Translation Systems that can produce
reasonable draft translations (provided the input observes certain restrictions on subject matter,
style, and vocabulary).
Unfortunately, how this is possible or what it really means is hard to appreciate for those
without the time, patience, or training to read the relevant academic research papers, which in
any case do not give a very good picture of what is involved in practice. It was for this reason
that we decided to try to write a book which would be genuinely introductory (in the sense of
not presupposing a background in any relevant discipline), but which would look at all aspects
of Machine Translation: covering questions of what it is like to use a modern Machine
Translation system, through questions about how it is done, to questions of evaluating systems,
and what developments can be foreseen in the near to medium future.
We would like to express our thanks to various people. First, we would like to thank each other.
The process of writing this book has been slower than we originally hoped (five authors is five
pairs of hands, but also five sets of opinions). However, we think that our extensive discussions
and revisions have in the end produced a better book in terms of content, style, presentation,
and so on. We think we deserve no little credit for maintaining a pleasant working atmosphere
while expending this level of effort and commitment while under pressure caused by other
academic responsibilities.
We would also like to thank our colleagues at the Computational Linguistics and Machine
Translation (CL/MT) group at the University of Essex for suggestions and practical support,
especially Lisa Hamilton, Kerry Maxwell, Dave Moffat, Tim Nicholas, Melissa Parker, Martin
Rondell and Andy Way.
For proofreading and constructive criticism we would like to thank John Roberts of the
Department of Language and Linguistics at the University of Essex, and John Roberts and
Karen Woods of NCC Blackwell. We are also grateful to those people who have helped us by
checking the examples which are in languages other than English and Dutch, especially
Laurence Danlos (French), and Nicola Jörn (German).
Of course, none of them is responsible for the errors of content, style or presentation that

http://privatewww.essex.ac.uk/~doug/book/node1.html (1 de 2) [20/09/2000 18:02:19]

Preface

remain.
D.J. Arnold
L. Balkan
R. Lee Humphreys
S. Meijer
L. Sadler
Colchester, August 1993.

Next: Contents Up: Machine Translation Previous:

MACHINE TRANSLATION An

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node1.html (2 de 2) [20/09/2000 18:02:19]

Contents

Next: Introduction and Overview Up: Machine Translation Previous: Preface

Contents
● Preface
● Introduction and Overview
❍ Introduction
❍ Why MT Matters
❍ Popular Conceptions and Misconceptions
❍ A Bit of History
❍ Summary
❍ Further Reading
● Machine Translation in Practice
❍ Introduction
❍ The Scenario
❍ Document Preparation: Authoring and Pre-Editing
❍ The Translation Process
■ Dictionary-Based Translation Support Tools
■ Interaction in Translation
❍ Document Revision
❍ Summary
❍ Further Reading
● Representation and Processing
❍ Introduction
❍ Representing Linguistic Knowledge
■ Grammars and Constituent Structure
■ Further Analysis: Grammatical Relations
■ Meaning
❍ Processing
■ Parsing
■ Generation
❍ Summary
❍ Further Reading

http://privatewww.essex.ac.uk/~doug/book/node2.html (1 de 4) [20/09/2000 18:02:20]

Contents

● Machine Translation Engines

❍ Introduction
❍ Transformer Architectures
❍ Linguistic Knowledge Architectures
■ Comparative Grammar and Transfer
■ Interlinguas
■ LK Engines Summarised
❍ Summary
❍ Further Reading
● Dictionaries
❍ Introduction
❍ Paper Dictionaries
❍ Types of Word Information
❍ Dictionaries and Morphology
■ Inflection
■ Derivation
■ Compounds
❍ Terminology
❍ Summary
❍ Further Reading
● Translation Problems
❍ Introduction
❍ Ambiguity
❍ Lexical and Structural Mismatches
❍ Multiword units: Idioms and Collocations
❍ Summary
❍ Further Reading
● Representation and Processing Revisited: Meaning
❍ Introduction
❍ Semantics
❍ Pragmatics
❍ Real World Knowledge
❍ Summary
❍ Further Reading
● Input
❍ Introduction

http://privatewww.essex.ac.uk/~doug/book/node2.html (2 de 4) [20/09/2000 18:02:20]

Contents

❍ The Electronic Document

■ Basic Ideas
■ SGML Markup and MT Input
❍ Controlled Languages
❍ Sublanguage MT
❍ Summary
❍ Further Reading
● Evaluating MT Systems
❍ Introduction
❍ Some Central Issues
❍ Evaluation of Engine Performance
■ Intelligibility
■ Accuracy
■ Error Analysis
■ The Test Suite
❍ Operational Evaluation
❍ Summary
❍ Further Reading
● New Directions in MT
❍ Introduction
❍ Rule-Based MT
■ Flexible or Multi-level MT
■ Knowledge-Based MT
■ Feasibility of General Purpose Rule-Based MT Systems
❍ Resources for MT
❍ Empirical Approaches to MT
■ Example-Based Translation
■ Statistical MT
❍ Summary
❍ Further Reading
● Useful Addresses
● Glossary
● References
● Index
● About this document ...

http://privatewww.essex.ac.uk/~doug/book/node2.html (3 de 4) [20/09/2000 18:02:20]

Contents

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node2.html (4 de 4) [20/09/2000 18:02:20]

Introduction and Overview

Next: Introduction Up: Machine Translation Previous: Contents

Introduction and Overview

● Introduction
● Why MT Matters
● Popular Conceptions and Misconceptions
● A Bit of History
● Summary
● Further Reading

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node3.html [20/09/2000 18:02:21]

Introduction

Next: Why MT Matters Up: Introduction and Overview Previous: Introduction and Overview

Introduction
The topic of the book is the art or science of Automatic Translation, or Machine Translation
(MT) as it is generally known --- the attempt to automate all, or part of the process of
translating from one human language to another. The aim of the book is to introduce this topic
to the general reader --- anyone interested in human language, translation, or computers. The
idea is to give the reader a clear basic understanding of the state of the art, both in terms of what
is currently possible, and how it is achieved, and of what developments are on the horizon. This
should be especially interesting to anyone who is associated with what are sometimes called
``the language industries''; particularly translators, those training to be translators, and those
who commission or use translations extensively. But the topics the book deals with are of
general and lasting interest, as we hope the book will demonstrate, and no specialist knowledge
is presupposed --- no background in Computer Science, Artificial Intelligence (AI), Linguistics,
or Translation Studies.
Though the purpose of this book is introductory, it is not just introductory. For one thing, we
will, in Chapter , bring the reader up to date with the most recent developments. For another,
as well as giving an accurate picture of the state of the art, both practically and theoretically, we
have taken a position on some of what seem to us to be the key issues in MT today --- the fact is
that we have some axes to grind.
From the earliest days, MT has been bedevilled by grandiose claims and exaggerated
expectations. MT researchers and developers should stop over-selling. The general public
should stop over-expecting. One of the main aims of this book is that the reader comes to
appreciate where we are today in terms of actual achievement, reasonable expectation, and
unreasonable hype. This is not the kind of thing that one can sum up in a catchy headline (``No
Prospect for MT'' or ``MT Removes the Language Barrier''), but it is something one can absorb,
and which one can thereafter use to distill the essence of truth that will lie behind reports of
products and research.
With all this in mind, we begin (after some introductory remarks in this chapter) with a
description of what it might be like to work with a hypothetical state of the art MT system. This
should allow the reader to get an overall picture of what is involved, and a realistic notion of
what is actually possible. The context we have chosen for this description is that of a large
organization where relatively sophisticated tools are used in the preparation of documents, and
where translation is integrated into document preparation. This is partly because we think this
context shows MT at its most useful. In any case, the reader unfamiliar with this situation
should have no trouble understanding what is involved.
The aim of the following chapters is to `lift the lid' on the core component of an MT system to
give an idea of what goes on inside --- or rather, since there are several different basic designs
for MT system --- to give an idea of what the main approaches are, and to point out their

http://privatewww.essex.ac.uk/~doug/book/node4.html (1 de 3) [20/09/2000 18:02:22]

Introduction

strengths and weaknesses.

Unfortunately, even a basic understanding of what goes on inside an MT system requires a
grasp of some relatively simple ideas and terminology, mainly from Linguistics and
Computational Linguistics, and this has to be given `up front'. This is the purpose of Chapter
. In this chapter, we describe some fundamental ideas about how the most basic sort of
knowledge that is required for translation can be represented in, and used by, a computer.

In Chapter we look at how the main kinds of MT system actually translate, by describing
the operation of the `Translation Engine'. We begin by describing the simplest design, which
we call the transformer architecture . Though now somewhat old hat as regards the research
community, this is still the design used in most commercial MT systems. In the second part of
the chapter, we describe approaches which involve more extensive and sophisticated kinds of
linguistic knowledge. We call these Linguistic Knowledge (LK) systems . They include the
two approaches that have dominated MT research over most of the past twenty years. The first
is the so-called interlingual approach , where translation proceeds in two stages, by analyzing
input sentences into some abstract and ideally language independent meaning representation,
from which translations in several different languages can potentially be produced. The second
is the so-called transfer approach , where translation proceeds in three stages, analyzing input
sentences into a representation which still retains characteristics of the original, source language
text. This is then input to a special component (called a transfer component) which produces a
representation which has characteristics of the target (output) language, and from which a target
sentence can be produced.
The still somewhat schematic picture that this provides will be amplified in the two following
chapters. In Chapter , we focus on what is probably the single most important component in
an MT system, the dictionary, and describe the sorts of issue that arise in designing,
constructing, or modifying the sort of dictionary one is likely to find in an MT system.

Chapter will go into more detail about some of the problems that arise in designing and
building MT systems, and, where possible, describe how they are, or could be solved. This
chapter will give an idea of why MT is `hard', of the limitations of current technology. It also
begins to introduce some of the open questions for MT research that are the topic of the final
chapter.

Such questions are also introduced in Chapter . Here we return to questions of representation
and processing, which we began to look at in Chapter , but whereas we focused previously
on morphological, syntactic, and relatively superficial semantic issues, in this chapter we turn to
more abstract, `deeper' representations --- representations of various kinds of representation of
meaning.

One of the features of the scenario we imagine in Chapter is that texts are mainly created,
stored, and manipulated electronically (for example, by word processors). In Chapter we
look in more detail at what this involves (or ideally would involve), and how it can be exploited
to yield further benefits from MT. In particular, we will describe how standardization of
electronic document formats and the general notion of standardized markup (which separates
the content of a document from details of its realization, so that a writer, for example, specifies
that a word is to be emphasised, but need not specify which typeface must be used for this) can

http://privatewww.essex.ac.uk/~doug/book/node4.html (2 de 3) [20/09/2000 18:02:22]

Introduction

be exploited when one is dealing with documents and their translations. This will go beyond
what some readers will immediately need to know. However, we consider its inclusion
important since the integration of MT into the document processing environment is an
important step towards the successful use of MT. In this chapter we will also look at the
benefits and practicalities of using controlled languages --- specially simplified versions of, for
example, English, and sublanguages --- specialized languages of sub-domains. Although
these notions are not central to a proper understanding of the principles of MT, they are widely
thought to be critical for the successful application of MT in practice.
Continuing the orientation towards matters of more practical than theoretical importance,
Chapter addresses the issue of the evaluation of MT systems --- of how to tell if an MT
system is `good'. We will go into some detail about this, partly because it is such an obvious
and important question to ask, and partly because there is no other accessible discussion of the
standard methods for evaluating MT systems that an interested reader can refer to.
By this time, the reader should have a reasonably good idea of what the `state of the art' of MT
is. The aim of the final chapter (Chapter ) is to try to give the reader an idea of what the
future holds by describing where MT research is going and what are currently thought to be the
most promising lines of research.
Throughout the book, the reader may encounter terms and concepts with which she is
unfamiliar. If necessary the reader can refer to the Glossary at the back of the book, where such
terms are defined.

Next: Why MT Matters Up: Introduction and Overview Previous: Introduction and Overview

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node4.html (3 de 3) [20/09/2000 18:02:22]

Why MT Matters

Next: Popular Conceptions and Up: Introduction and Overview Previous: Introduction

Why MT Matters
The topic of MT is one that we have found sufficiently interesting to spend most of our
professional lives investigating, and we hope the reader will come to share, or at least
understand, this interest. But whatever one may think about its intrinsic interest, it is
undoubtedly an important topic --- socially, politically, commercially, scientifically, and
intellectually or philosophically --- and one whose importance is likely to increase as the 20th
Century ends, and the 21st begins.
The social or political importance of MT arises from the socio-political importance of
translation in communities where more than one language is generally spoken. Here the only
viable alternative to rather widespread use of translation is the adoption of a single common
`lingua franca', which (despite what one might first think) is not a particularly attractive
alternative, because it involves the dominance of the chosen language, to the disadvantage of
speakers of the other languages, and raises the prospect of the other languages becoming
second-class, and ultimately disappearing. Since the loss of a language often involves the
disappearance of a distinctive culture, and a way of thinking, this is a loss that should matter to
everyone. So translation is necessary for communication --- for ordinary human interaction, and
for gathering the information one needs to play a full part in society. Being allowed to express
yourself in your own language, and to receive information that directly affects you in the same
medium, seems to be an important, if often violated, right. And it is one that depends on the
availability of translation. The problem is that the demand for translation in the modern world
far outstrips any possible supply. Part of the problem is that there are too few human translators,
and that there is a limit on how far their productivity can be increased without automation. In
short, it seems as though automation of translation is a social and political necessity for modern
societies which do not wish to impose a common language on their members.
This is a point that is often missed by people who live in communities where one language is
dominant, and who speak the dominant language. Speakers of English in places like Britain,
and the Northern USA are examples. However, even they rapidly come to appreciate it when
they visit an area where English is not dominant (for example, Welsh speaking areas of Britain,
parts of the USA where the majority language is Spanish, not to mention most other countries
in the world). For countries like Canada and Switzerland, and organizations like the European
Community and the UN, for whom multilingualism is both a basic principle and a fact of every
day life, the point is obvious.
The commercial importance of MT is a result of related factors. First, translation itself is
commercially important: faced with a choice between a product with an instruction manual in
English, and one whose manual is written in Japanese, most English speakers will buy the
former --- and in the case of a repair manual for a piece of manufacturing machinery or the
manual for a safety critical system, this is not just a matter of taste. Secondly, translation is
expensive. Translation is a highly skilled job, requiring much more than mere knowledge of a

http://privatewww.essex.ac.uk/~doug/book/node5.html (1 de 3) [20/09/2000 18:02:23]

Why MT Matters

number of languages, and in some countries at least, translators' salaries are comparable to
other highly trained professionals. Moreover, delays in translation are costly. Estimates vary,
but producing high quality translations of difficult material, a professional translator may
average no more than about 4-6 pages of translation (perhaps 2000 words) per day, and it is
quite easy for delays in translating product documentation to erode the market lead time of a
new product. It has been estimated that some 40-45% of the running costs of European
Community institutions are `language costs', of which translation and interpreting are the main
element. This would give a cost of something like million per annum. This figure relates to
translations actually done, and is a tiny fraction of the cost that would be involved in doing all
the translations that could, or should be done.
Scientifically, MT is interesting, because it is an obvious application and testing ground for
many ideas in Computer Science, Artificial Intelligence, and Linguistics, and some of the most
important developments in these fields have begun in MT. To illustrate this: the origins of
Prolog , the first widely available logic programming language, which formed a key part of the
Japanese `Fifth Generation' programme of research in the late 1980s, can be found in the
`Q-Systems' language, originally developed for MT.
Philosophically, MT is interesting, because it represents an attempt to automate an activity that
can require the full range of human knowledge --- that is, for any piece of human knowledge, it
is possible to think of a context where the knowledge is required. For example, getting the
correct translation of negatively charged electrons and protons into French depends on
knowing that protons are positively charged, so the interpretation cannot be something like
``negatively charged electrons and negatively charged protons''. In this sense, the extent to
which one can automate translation is an indication of the extent to which one can automate
`thinking'.
Despite this, very few people, even those who are involved in producing or commissioning
translations, have much idea of what is involved in MT today, either at the practical level of
what it means to have and use an MT system, or at the level of what is technically feasible, and
what is science fiction. In the whole of the UK there are perhaps five companies who use MT
for making commercial translations on a day-to-day basis. In continental Europe, where the
need for commercial translation is for historical reasons greater, the number is larger, but it still
represents an extremely small proportion of the overall translation effort that is actually
undertaken. In Japan, where there is an enormous need for translation of Japanese into English,
MT is just beginning to become established on a commercial scale, and some familiarity with
MT is becoming a standard part of the training of a professional translator.
Of course, theorists, developers, and sellers of MT systems must be mainly responsible for this
level of ignorance and lack of uptake, and we hope this book will help here --- one motivation
for writing this book was our belief that an understanding of MT is an essential part of the
equipment of a professional translator, and the knowledge that no other book provided this in
accessible form.
We are reminded of this scale of ignorance every time we admit to working in the field of MT.
After initial explanations of what MT is, the typical reaction is one of two contradictory
responses (sometimes one gets both together). One is ``But that's impossible --- no machine
could ever translate Shakespeare.'' The other is ``Yes, I saw one of those in the Duty Free Shop
when I went on holiday last summer.'' These reactions are based on a number of misconceptions
that are worth exposing. We will look at these, as well as some correct conceptions, in the next

http://privatewww.essex.ac.uk/~doug/book/node5.html (2 de 3) [20/09/2000 18:02:23]

Why MT Matters

section.

Next: Popular Conceptions and Up: Introduction and Overview Previous: Introduction

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node5.html (3 de 3) [20/09/2000 18:02:23]

Popular Conceptions and Misconceptions

Next: A Bit of Up: Introduction and Overview Previous: Why MT Matters

Popular Conceptions and

Misconceptions

Some popular misconceptions about MT are listed on page . We will discuss them in turn.
``MT is a waste of time because you will never make a machine that can translate
Shakespeare''.
The criticism that MT systems cannot, and will never, produce translations of great literature of
any great merit is probably correct, but quite beside the point. It certainly does not show that
MT is impossible. First, translating literature requires special literary skill --- it is not the kind
of thing that the average professional translator normally attempts. So accepting the criticism
does not show that automatic translation of non-literary texts is impossible. Second, literary
translation is a small proportion of the translation that has to be done, so accepting the criticism
does not mean that MT is useless. Finally, one may wonder who would ever want to translate
Shakespeare by machine --- it is a job that human translators find challenging and rewarding,
and it is not a job that MT systems have been designed for. The criticism that MT systems
cannot translate Shakespeare is a bit like criticism of industrial robots for not being able to
dance Swan Lake.
``There was/is an MT system which translated The spirit is willing, but the flesh is weak into
the Russian equivalent of The vodka is good, but the steak is lousy, and hydraulic ram into the
French equivalent of water goat. MT is useless.''
The `spirit is willing' story is amusing, and it really is a pity that it is not true. However, like
most MT `howlers' it is a fabrication. In fact, for the most part, they were in circulation long
before any MT system could have produced them (variants of the `spirit is willing' example can
be found in the American press as early as 1956, but sadly, there does not seem to have been an
MT system in America which could translate from English into Russian until much more
recently --- for sound strategic reasons, work in the USA had concentrated on the translation of
Russian into English, not the other way round). Of course, there are real MT howlers. Two of
the nicest are the translation of French avocat (`advocate', `lawyer' or `barrister') as avocado,
and the translation of Les soldats sont dans le café as The soldiers are in the coffee. However,
they are not as easy to find as the reader might think, and they certainly do not show that MT is
useless.

http://privatewww.essex.ac.uk/~doug/book/node6.html (1 de 7) [20/09/2000 18:02:26]

Popular Conceptions and Misconceptions

``Generally, the quality of translation you can get from an MT system is very low. This
makes them useless in practice.''
Far from being useless, there are several MT systems in day-to-day use around the world.
Examples include METEO (in daily since 1977 use at the Canadian Meteorological Center in
Dorval, Montreal), SYSTRAN (in use at the CEC , and elsewhere), LOGOS , ALPS ,
ENGSPAN (and SPANAM ), METAL , GLOBALINK . It is true that the number of
organizations that use MT on a daily basis is relatively small, but those that do use it benefit
considerably. For example, as of 1990, METEO was regularly translating around 45 000 words
of weather bulletins every day, from English into French for transmission to press, radio, and
television. In the 1980s, the diesel engine manufacturers Perkins Engines was saving around £
4 000 on each diesel engine manual translated (using a PC version of WEIDNER system).
Moreover, overall translation time per manual was more than halved from around 26 weeks to
9-12 weeks --- this time saving can be very significant commercially, because a product like an
engine cannot easily be marketed without user manuals.
Of course, it is true that the quality of many MT systems is low, and probably no existing
system can produce really perfect translations. However, this does not make MT useless.
First, not every translation has to be perfect. Imagine you have in front of you a Chinese
newspaper which you suspect may contain some information of crucial importance to you or
your company. Even a very rough translation would help you. Apart from anything else, you
would be able to work out which, if any, parts of the paper would be worth getting translated
properly. Second, a human translator normally does not immediately produce a perfect
translation. It is normal to divide the job of translating a document into two stages. The first
stage is to produce a draft translation, i.e. a piece of running text in the target language, which
has the most obvious translation problems solved (e.g. choice of terminology, etc.), but which is

http://privatewww.essex.ac.uk/~doug/book/node6.html (2 de 7) [20/09/2000 18:02:26]

Popular Conceptions and Misconceptions

not necessarily perfect. This is then revised --- either by the same translator, or in some large
organizations by another translator --- with a view to producing something that is up to standard
for the job in hand. This might involve no more than checking, or it might involve quite radical
revision aimed at producing something that reads as though written originally in the target
language. For the most part, the aim of MT is only to automate the first, draft translation
process.
``MT threatens the jobs of translators.''
The quality of translation that is currently possible with MT is one reason why it is wrong to
think of MT systems as dehumanizing monsters which will eliminate human translators, or
enslave them. It will not eliminate them, simply because the volume of translation to be
performed is so huge, and constantly growing, and because of the limitations of current and
forseeable MT systems. While not an immediate prospect, it could, of course, turn out that MT
enslaves human translators, by controlling the translation process, and forcing them to work on
the problems it throws up, at its speed. There are no doubt examples of this happening to other
professions. However, there are not many such examples, and it is not likely to happen with
MT. What is more likely is that the process of producing draft translations, along with the often
tedious business of looking up unknown words in dictionaries, and ensuring terminological
consistency, will become automated, leaving human translators free to spend time on increasing
clarity and improving style, and to translate more important and interesting documents ---
editorials rather than weather reports, for example. This idea borne out in practice: the job
satisfaction of the human translators in the Canadian Meteorological Center improved when
METEO was installed, and their job became one of checking and trying to find ways to
improve the system output, rather than translating the weather bulletins by hand (the concrete
effect of this was a greatly reduced turnover in translation staff at the Center).
``The Japanese have developed a system that you can talk to on the phone. It translates what
you say into Japanese, and translates the other speaker's replies into English.''
The claim that the Japanese have a speech to speech translation system, of the kind described
above, is pure science fiction. It is true that speech-to-speech translation is a topic of current
research, and there are laboratory prototypes that can deal with a very restricted range of
questions. But this research is mainly aimed at investigating how the various technologies
involved in speech and language processing can be integrated, and is limited to very restricted
domains (hotel bookings, for example), and messages (offering little more than a phrase book in
these domains). It will be several years before even this sort of system will be in any sort of real
use. This is partly because of the limitations of speech systems, which are currently fine for
recognizing isolated words, uttered by a single speaker, for which the system has been specially
trained, in quiet conditions, but which do not go far beyond this. However, it is also because of
the limitations of the MT system (see later chapters).
``There is an amazing South American Indian language with a structure of such logical
perfection that it solves the problem of designing MT systems.''
The South American Indian language story is among the most irritating for MT researchers.
First, the point about having a `perfectly logical structure' is almost certainly completely false.
Such perfection is mainly in the eye of the beholder --- Diderot was convinced that the word
order of French exactly reflected the order of thought, a suggestion that non-French speakers do
not find very convincing. What people generally mean by this is that a language is very simple

http://privatewww.essex.ac.uk/~doug/book/node6.html (3 de 7) [20/09/2000 18:02:26]

Popular Conceptions and Misconceptions

to describe. Now, as far as anyone can tell all human languages are pretty much as complicated
as each other. It's hard to be definite, since the idea of simplicity is difficult to pin down, but the
general impression is that if a language has a very simple syntax, for example, it will
compensate by having a more complicated morphology (word structure), or phonology (sound
structure). However, even if one had a very neat logical language, it is hard to see that this
would solve the MT problem, since one would still have to perform automatic translation into,
and out of, this language.
``MT systems are machines, and buying an MT system should be very much like buying a
car.''
There are really two parts to this misconception. The first relates to the sense in which MT
systems are machines. They are, of course, but only in the sense that modern word processors
are machines. It is more accurate to think of MT systems as programs that run on computers
(which really are machines). Thus, when one talks about buying, modifying, or repairing an MT
system, one is talking about buying, modifying or repairing a piece of software. It was not
always so --- the earliest MT systems were dedicated machines, and even very recently, there
were some MT vendors who tried to sell their systems with specific hardware, but this is
becoming a thing of the past. Recent systems can be installed on different types of computers.
The second part of the misconception is the idea that one would take an MT system and `drive
it away', as one would a car. In fact, this is unlikely to be possible, and a better analogy is with
buying a house --- what one buys may be immediately habitable, but there is a considerable
amount of work involved in adapting it to one's own special needs. In the case of a house this
might involve changes to the decor and plumbing. In the case of an MT system this will involve
additions to the dictionaries to deal with the vocabulary of the subject area and possibly the type
of text to be translated. There will also be some work involved in integrating the system into the
rest of one's document processing environment. More of this in Chapters and . The
importance of customization, and the fact that changes to the dictionary form a major part of the
process is one reason why we have given a whole chapter to discussion of the dictionary
(Chapter ).
Against these misconceptions, we should place the genuine facts about MT. These are listed on
page .

http://privatewww.essex.ac.uk/~doug/book/node6.html (4 de 7) [20/09/2000 18:02:26]

Popular Conceptions and Misconceptions

The correct conclusion is that MT, although imperfect, is not only a possibility, but an actuality.
But it is important to see the product in a proper perspective, to be aware of its strong points
and shortcomings.
Machine Translation started out with the hope and expectation that most of the work of
translation could be handled by a system which contained all the information we find in a
standard paper bilingual dictionary. Source language words would be replaced with their target
language translational equivalents, as determined by the built-in dictionary, and where
necessary the order of the words in the input sentences would be rearranged by special rules
into something more characteristic of the target language. In effect, correct translations suitable
for immediate use would be manufactured in two simple steps. This corresponds to the view
that translation is nothing more than word substitution (determined by the dictionary) and

http://privatewww.essex.ac.uk/~doug/book/node6.html (5 de 7) [20/09/2000 18:02:26]

Popular Conceptions and Misconceptions

reordering (determined by reordering rules).

Reason and experience show that `good' MT cannot be produced by such delightfully simple
means. As all translators know, word for word translation doesn't produce a satisfying target
language text, not even when some local reordering rules (e.g. for the position of the adjective
with regard to the noun which it modifies) have been included in the system. Translating a text
requires not only a good knowledge of the vocabulary of both source and target language, but
also of their grammar --- the system of rules which specifies which sentences are well-formed
in a particular language and which are not. Additionally it requires some element of real world
knowledge --- knowledge of the nature of things out in the world and how they work together
--- and technical knowledge of the text's subject area. Researchers certainly believe that much
can be done to satisfy these requirements, but producing systems which actually do so is far
from easy. Most effort in the past 10 years or so has gone into increasing the subtlety, breadth
and depth of the linguistic or grammatical knowledge available to systems. We shall take a
more detailed look at these developments in due course.
In growing into some sort of maturity, the MT world has also come to realize that the `text in
translation out' assumption --- the assumption that MT is solely a matter of switching on the
machine and watching a faultless translation come flying out --- was rather too naive. A
translation process starts with providing the MT system with usable input. It is quite common
that texts which are submitted for translation need to be adapted (for example, typographically,
or in terms of format) before the system can deal with them. And when a text can actually be
submitted to an MT system, and the system produces a translation, the output is almost
invariably deemed to be grammatically and translationally imperfect. Despite the increased
complexity of MT systems they will never --- within the forseeable future --- be able to handle
all types of text reliably and accurately. This normally means that the translation will have to
be corrected (post-edited) and usually the person best equipped to do this is a translator.
This means that MT will only be profitable in environments that can exploit the strong points to
the full. As a consequence, we see that the main impact of MT in the immediate future will be
in large corporate environments where substantial amounts of translation are performed. The
implication of this is that MT is not (yet) for the individual self-employed translator working
from home, or the untrained lay-person who has the occasional letter to write in French. This is
not a matter of cost: MT systems sell at anywhere between a few hundred pounds and over . It
is a matter of effective use. The aim of MT is to achieve faster, and thus cheaper, translation.
The lay-person or self-employed translator would probably have to spend so much time on
dictionary updating and/or post-editing that MT would not be worthwhile. There is also the
problem of getting input texts in machine readable form, otherwise the effort of typing will
outweigh any gains of automation. The real gains come from integrating the MT system into the
whole document processing environment (see Chapter ), and they are greatest when several
users can share, for example, the effort of updating dictionaries, efficiencies of avoiding
unnecessary retranslation, and the benefits of terminological consistency.
Most of this book is about MT today, and to some extent tomorrow. But MT is a subject with
an interesting and dramatic past, and it is well worth a brief description.

Next: A Bit of Up: Introduction and Overview Previous: Why MT Matters

http://privatewww.essex.ac.uk/~doug/book/node6.html (6 de 7) [20/09/2000 18:02:26]

Popular Conceptions and Misconceptions

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node6.html (7 de 7) [20/09/2000 18:02:26]

A Bit of History

Next: Summary Up: Introduction and Overview Previous: Popular Conceptions and

A Bit of History

There is some dispute about who first had the idea of translating automatically between human
languages, but the actual development of MT can be traced to conversations and
correspondence between Andrew D. Booth, a British crystallographer, and Warren Weaver of
the Rockefeller Foundation in 1947, and more specifically to a memorandum written by
Weaver in 1949 to the Rockerfeller Foundation which included the following two sentences.
``I have a text in front of me which is written in Russian but I am going to pretend
that it is really written in English and that it has been coded in some strange
symbols. All I need to do is strip off the code in order to retrieve the information
contained in the text.''
The analogy of translation and decoding may strike the sophisticated reader as simplistic
(however complicated coding gets it is still basically a one-for-one substitution process where
there is only one right answer --- translation is a far more complex and subtle business), and
later in the memorandum Weaver proposed some other more sophisticated views, but it had
the virtue of turning an apparently difficult task into one that could be approached with the
emergent computer technology (there had been considerable success in using computers in
cryptography during the Second World War). This memorandum sparked a significant amount
of interest and research, and by the early 1950s there was a large number of research groups
working in Europe and the USA, representing a significant financial investment (equivalent to
around 000 000). But, despite some success, and the fact that many research questions were
raised that remain important to this day, there was widespread disappointment on the part of
funding authorities at the return on investment that this represented, and doubts about the
possibility of automating translation in general, or at least in the current state of knowledge.
The theoretical doubts were voiced most clearly by the philosopher Bar-Hillel in a 1959 report,
where he argued that fully automatic, high quality , MT (FAHQMT) was impossible, not just at
present, but in principle. The problem he raised was that of finding the right translation for pen
in a context like the following:
● John was looking for his toy box. Finally he found it. The box was in the pen. John
was very happy.
The argument was that (i) here pen could only have the interpretation play-pen, not the
alternative writing instrument interpretation, (ii) this could be critical in deciding the correct
translation for pen, (iii) discovering this depends on general knowledge about the world, and
(iv) there could be no way of building such knowledge into a computer. Some of these points
are well taken. Perhaps FAHQMT is impossible. But this does not mean that any form of MT
is impossible or useless, and in Chapter we will look at some of the ways one might go
about solving this problem. Nevertheless, historically, this was important in suggesting that

http://privatewww.essex.ac.uk/~doug/book/node7.html (1 de 3) [20/09/2000 18:02:27]

A Bit of History

research should focus on more fundamental issues in the processing and understanding of
human languages.
The doubts of funding authorities were voiced in the report which the US National Academy of
Sciences commissioned in 1964 when it set up the Automatic Language Processing Advisory
Committee (ALPAC) to report on the state of play with
respect to MT as regards quality , cost, and prospects, as against the existing cost of, and need
for translation. Its report, the so-called ALPAC Report , was damning, concluding that there was
no shortage of human translators, and that there was no immediate prospect of MT producing
useful translation of general scientific texts. This report led to the virtual end of Government
funding in the USA. Worse, it led to a general loss of morale in the field, as early hopes were
perceived to be groundless.
The spectre of the ALPAC report, with its threats of near complete withdrawal of funding, and
demoralization, still haunts workers in MT. Probably it should not, because the achievements of
MT are real, even if they fall short of the idea of FAHQMT all the time --- useful MT is neither
science fiction, nor merely a topic for scientific speculation. It is a daily reality in some places,
and for some purposes. However, the fear is understandable, because the conclusion of the
report was almost entirely mistaken. First, the idea that there was no need for machine
translation is one that should strike the reader as absurd, given what we said earlier. One can
only understand it in the anglo-centric context of cold-war America, where the main reason to
translate was to gain intelligence about Soviet activity. Similarly, the suggestion that there was
no prospect of successful MT seems to have been based on a narrow view of FAHQMT --- in
particular, on the idea that MT which required revision was not `real' MT. But, keeping in mind
the considerable time gain that can be achieved by automating the draft translation stage of the
process, this view is naive. Moreover, there were, even at the time the report was published,
three systems in regular, if not extensive, use (one at the Wright Patterson USAF base, one at
the Oak Ridge Laboratory of the US Atomic Energy Commission , and one the EURATOM
Centre at Ispra in Italy).
Nevertheless, the central conclusion that MT did not represent a useful goal for research or
development work had taken hold, and the number of groups and individuals involved in MT
research shrank dramatically. For the next ten years, MT research became the preserve of
groups funded by the Mormon Church, who had an interest in bible translation (the work that
was done at Brigham Young University in Provo, Utah ultimately led to the WEIDNER and
ALPS systems, two notable early commercial systems), and a handful of groups in Canada
(notably the TAUM group in Montreal, who developed the METEO system mentioned
earlier), the USSR (notably the groups led by Mel'cuk, and Apresian) , and Europe (notably the
GETA group in Grenoble, probably the single most influential group of this period, and the
SUSY group in Saarbrücken). A small fraction of the funding and effort that had been devoted
to MT was put into more fundamental research on Computational Linguistics, and Artificial
Intelligence, and some of this work took MT as a long term objective, even in the USA (Wilks'
work on AI is notable in this respect). It was not until the late 1970s that MT research
underwent something of a renaissance.

There were several signs of this renaissance. The Commission of the European Communities
(CEC) purchased the English-French version of the SYSTRAN system, a greatly improved
descendent of the earliest systems developed at Georgetown University (in Washington, DC), a
Russian-English system whose development had continued throughout the lean years after

http://privatewww.essex.ac.uk/~doug/book/node7.html (2 de 3) [20/09/2000 18:02:27]

A Bit of History

ALPAC , and which had been used by both the USAF and NASA . The CEC also
commissioned the development of a French-English version , and Italian-English version. At
about the same time, there was a rapid expansion of MT activity in Japan, and the CEC also
began to set up what was to become the EUROTRA project, building on the work of the
GETA and SUSY groups. This was perhaps the largest, and certainly among the most
ambitious research and development projects in Natural Language Processing. The aim was to
produce a `pre-industrial' MT system of advanced design (what we call a Linguistic Knowledge
system) for the EC languages. Also in the late 1970s the Pan American Health Organization
(PAHO) began development of a Spanish-English MT system (SPANAM) , the United States
Air Force funded work on the METAL system at the Linguistics Research Center, at the
University of Texas in Austin, and the results of work at the TAUM group led to the
installation of the METEO system. For the most part, the history of the 1980s in MT is the
history of these initiatives, and the exploitation of results in neighbouring disciplines.
As one moves nearer to the present, views of history are less clear and more subjective. Chapter
will describe what we think are the most interesting and important technical innovations. As
regards the practical and commercial application of MT systems. The systems that were on the
market in the late 1970s have had their ups and downs, but for commercial and marketing
reasons, rather than scientific or technical reasons, and a number of the research projects which
were started in the 1970s and 1980s have led to working, commercially available systems. This
should mean that MT is firmly established, both as an area of legitimate research, and a useful
application of technology. But researching and developing MT systems is a difficult task both
technically, and in terms of management, organization and infrastructure, and it is an expensive
task, in terms of time, personnel, and money. From a technical point of view, there are still
fundamental problems to address. However, all of this is the topic of the remainder of this book.

Next: Summary Up: Introduction and Overview Previous: Popular Conceptions and

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node7.html (3 de 3) [20/09/2000 18:02:27]

Summary

Next: Further Reading Up: Introduction and Overview Previous: A Bit of

Summary
This chapter has given an outline of the rest of the book, and given a potted history of MT. It
has also tried to lay a few ghosts, in the form of misconceptions which haunt the enterprise.
Above all we hope to convince the reader that MT is possible and potentially useful, despite
current limitations.

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node8.html [20/09/2000 18:02:27]

Further Reading

Next: Machine Translation in Up: Introduction and Overview Previous: Summary

Further Reading
A broad, practically oriented view of the field of current MT by a variety of authors can be
found in [Newton1992a]. Generally speaking, the best source of material that takes an MT
user's viewpoint is the series of books titled Translating and the Computer, with various editors
and publishers, including [Lawson1982a], [Snell1979], [Snell1982], [Lawson1982b],
[Picken1985], [Picken1986], [Picken1987], [Picken1988], [Mayorcas1990], [Picken1990], and
[MayorcasForthcoming]. These are the published proceedings of the annual Conference on
Translating and the Computer, sponsored by Aslib (The Association for Information
Management), and the Institute for Translation and Interpreting.
By far the best technical introduction to MT is [Hutchins and Somers1992]. This would be
appropriate for readers who want to know more technical and scientific details about MT, and
we will often refer to it in later chapters. This book contains useful discussions of some of the
main MT systems, but for descriptions of these systems by their actual designers the reader
should look at [Slocum1988] , and [King1987] . Slocum's introduction to the former,
[Slocum1986], is particularly recommended as an overview of the key issues in MT. These
books all contain detailed descriptions of the research of the TAUM group which developed
the
METEO system referred to in section . The METEO system is discussed further in Chapter
.
A short assessment of the current state of MT in terms of availability and use of systems in
Europe, North America, and Japan and East Asia can be found in [Pugh1992]. An up-to-date
picture of the state of MT as regards both commercial and scientific points of view is provided
every two years by the Machine Translation Summits. A report of one of these can be found in
[Nagao1989]. There is a description of the successful use of MT in a corporate setting in
[Newton1992b].

On the history of MT (which we have outlined here, but which will not be discussed again), the
most comprehensive discussion can be found in [Hutchins1986], though there are also useful
discussions in [Warwick1987] , and [Buchmann1987] . [Nagao1986] also provides a useful
insight into the history of MT, together with a general introduction to MT. The ALPAC report
is [Pierce and Carroll1966]. The work of Wilks' that is referred to in section is
[Wilks1973].

For general descriptions and discussion of the activity of translation (both human and machine)
[Picken1989] is a useful and up-to-date source. This contains references to (for example)
works on translation theory, and gives a great deal of practical information of value to
translators (such as lists national translators' and interpreters' organizations, and bibliographies

http://privatewww.essex.ac.uk/~doug/book/node9.html (1 de 2) [20/09/2000 18:02:28]

Further Reading

of translations).
For up-to-date information about the state of MT, there is the newsletter of the International
Association for Machine Translation MT News International. See the list of addresses on page
.

Next: Machine Translation in Up: Introduction and Overview Previous: Summary

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node9.html (2 de 2) [20/09/2000 18:02:28]

Machine Translation in Practice

Next: Introduction Up: Machine Translation Previous: Further Reading

Machine Translation in Practice

● Introduction
● The Scenario
● Document Preparation: Authoring and Pre-Editing
● The Translation Process
❍ Dictionary-Based Translation Support Tools
❍ Interaction in Translation
● Document Revision
● Summary
● Further Reading

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node10.html [20/09/2000 18:02:29]

Introduction

Next: The Scenario Up: Machine Translation in Previous: Machine Translation in

Introduction
At the time of writing, the use of MT --- or indeed, any sort of computerised tool for translation
support --- is completely unknown to the vast majority of individuals and organizations in the
world, even those involved in the so called `language industries', like translators,
terminologists, technical writers, etc.
Given this, one of the first things a reader is likely to want to know about MT is what it might
be like to work with an MT system and how it fits in with the day-to-day business of
translation. The purpose of the present chapter is to provide just such information --- a view of
MT at the user level, and from the outside. In later chapters we shall in effect lift off the covers
of an MT system and take a look at what goes on inside. For the moment, however, the central
components of an MT system are treated as a black box.
We introduce the business of MT in terms of a scenario describing the usage of MT inside a
fairly large multinational corporation. The scenario is not based exactly on any one existing
corporation. Our description is somewhat idealised in that we assume methods of working
which are only just starting to come into use. However, there is nothing idly futuristic in our
description: it is based on a consensus view of commercial MT experts and envisages tools
which we know to be either already available or in an advanced state of development in Europe
or elsewhere. The commercialisation of MT is not awaiting a `miracle breakthrough' in the
science of MT; it is not necessary, nor do we expect it to occur. What will happen over the next
ten years are progressive improvements in functionality and performance which, taken in
conjunction with the continuously falling costs of basic computing power, will ensure that MT
becomes more and more cost effective. In short, we have no doubt that in general outline, if not
in every detail, we are sketching the professional life of the machine translator in the 90s, and
of most translators in the early part of the next century.

Next: The Scenario Up: Machine Translation in Previous: Machine Translation in

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node11.html [20/09/2000 18:02:29]

The Scenario

Next: Document Preparation: Authoring Up: Machine Translation in Previous: Introduction

The Scenario
Let us suppose that you are a native English speaker engaged as a professional German-English translator in the Language
Centre for a multinational manufacturing company. One of the products this company supplies is computer products. In this
organization the Language Centre is principally responsible for the translation of documents created within the company into
a variety of European and Oriental languages. The Language Centre is also charged with exercising control over the content
and presentation of company documentation in general. To this end, it attempts to specify standards for the final appearance
of documents in distributed form, including style, terminology, and content in general. The overall policy is enshrined in the
form of a corporate Document Design and Content Guide which the Centre periodically updates and revises.
The material for which MT is to be used consists of technical documentation such as User and Repair manuals for software
and hardware products manufactured or sourced by the company. Some classes of highly routine internal business
correspondence are also submitted for MT. Legal and marketing material, and much external business correspondence, is
normally translated by hand, although some translators in the organization prefer to use MT here as well.
All material for translation is available in electronic form on a computer network which supports the company's
documentation system. Although most documents will be printed out at some point as standard paper User Manuals and so
forth, the system also supports the preparation of multi-media hypertext documents. These are documents which exist
primarily in electronic form with a sophisticated cross-reference system; they contain both text and pictures (and perhaps
speech and other sounds). These documents are usually distributed to their final users as CD-ROMs, although they can be
distributed in other electronic forms, including electronic mail. Printed versions of these documents can also be made.
Everyone in the language department has a workstation --- an individual computer. These are linked together by the network.
The documentation system which runs on this network allows users to create and modify documents by typing in text; in
other words, it provides very sophisticated word processing facilities. It also provides sophisticated means for storing and
retrieving electronic documents , and for passing them around the network inside the company or via external networks to
external organizations. As is usual with current computer systems, everything is done with the help of a friendly interface
based on windows, icons and menus, selections being made with a mouse.
The MT system which you use is called ETRANS and forms part of the overall documentation system. (ETRANS is just a
name we have invented for a prototypical MT system.) Parts of an electronic document on the system can be sent to the MT
system in the same way that they can be sent to a printer or to another device or facility on the network. ETRANS is
simultaneously available from any workstation and, for each person using it, behaves as if it is his or her own personal MT
system.
Earlier this morning, one of the technical authors had completed (two days after the deadline) a User Manual for a printer the
company is about to launch. The text is in German . Although this author works in a building 50 kilometres away, the
network ensures that the document is fully accessible from your workstation. What follows is a fragment of the text which
you are viewing in a window on the workstation screen and which you are going to translate:

http://privatewww.essex.ac.uk/~doug/book/node12.html (1 de 5) [20/09/2000 18:02:32]

The Scenario

As with all the technical documents submitted to ETRANS, all the sentences are relatively short and rather plain. Indeed, it
was written in accordance with the Language Centre document specification and with MT very much in mind. There are no
obvious idioms or complicated linguistic constructions. Many or all of the technical terms relating to printers (e.g.
Druckdichte `print density') are in regular use in the company and are stored and defined in paper or electronic dictionaries
available to the company's technical authors and translators.
To start up ETRANS, you click on the icon bearing an ETRANS logo, and this pops up a menu giving various translation
options. ETRANS handles six languages: English, German, French, Italian, Spanish and Japanese. The printer document
needs to be translated into English, so you select English as the target language option. Another menu shows the source
language to be used. In this case, there is no need to select German because ETRANS has already had a very quick look at
your printer document and decided, given rather superficial criteria such as the presence of umlauts and other characteristics
of German orthography , that it is probably German text. If ETRANS had guessed wrongly --- as it sometimes does --- then
you could select the correct source language from the menu yourself. By clicking on an additional menu of ETRANS
options, you start it translating in batch or full-text mode; that is, the whole text will be translated automatically without any
intervention on your part. The translation starts appearing in a separate screen window more or less immediately. However,
since the full source text is quite long, it will take some time to translate it in its entirety. Rather than sit around, you decide
to continue with the revision of another translation in another window. You will look at the output as soon as it has finished
translating the first chapter.

The output of ETRANS can be found on page . The quality of this raw output is pretty much as you expect from
ETRANS. Most sentences are more or less intelligible even if you don't go back to the German source. (Sometimes some
sentences may be completely unintelligible.) The translation is relatively accurate in the sense that it is not misleading --- it
doesn't lead you to think that the source text says one thing when it really says something quite the opposite. However, the
translation is very far from being a good specimen of English. For one thing, ETRANS clearly had difficulties with choosing
the correct translation of the German word ein which has three possible English equivalents: a/ an, on and one.

● ❍ Turn the button or two positions in direction of the dark indicator.

❍ Switch off the printer for a moment and then again

http://privatewww.essex.ac.uk/~doug/book/node12.html (2 de 5) [20/09/2000 18:02:32]

The Scenario

,
so that the test page is printed.
Apart from these details, it has also made quite a mess of a whole phrase:
● ❍ ....., similarly like at easily unclean copies of a photocopier.

In order to post-edit such phrases it will be necessary to refer back to the German source text.

Leaving ETRANS to continue translating later chapters of the document, you start post-editing the first chapter by opening
up a post-edit window, which interleaves a copy of the raw ETRANS output with the corresponding source sentences (e.g.
so that each source sentence appears next to its proposed translation). Your workstation screen probably now looks
something like the Figure on page .
Icons and menus give access to large scale on-line multilingual dictionaries --- either the ones used by the ETRANS itself or
others specifically intended for human users. You post-edit the raw MT using the range of word-processing functions
provided by the document processing system. Using search facilities, you skip through the document looking for all
instances of a, an or one, since you know that these are often wrong and may need replacement. (Discussions are in progress
with the supplier of ETRANS who has promised to look into this problem and make improvements.) After two or three other
global searches for known problem areas, you start to go through the document making corrections sentence by sentence.
The result of this is automatically separated from the source text, and can be displayed in yet another window. Page
shows what your workstation screen might now look like.

http://privatewww.essex.ac.uk/~doug/book/node12.html (3 de 5) [20/09/2000 18:02:32]

The Scenario

Figure: Translators' Workstation while Post-Editing a Translation

Note that ETRANS has left the document format completely unaltered. It may be that the translation is actually slightly
longer (or shorter) than the source text; any necessary adjustment to the pagination of the translation compared to the source
is a matter for the document processing system.
After post-editing the remaining text, you have almost completed the entire translation process. Since it is not uncommon
for translators to miss some small translation errors introduced by the MT system, you observe company policy by sending
your post-edited electronic text to a colleague to have it double-checked. The result is similar to that on page .
The only thing left to be done is to update the term dictionary , by adding any technical terms that have appeared in the
document with their translation terms which other translators should in future translate in the same way, and report any new
errors the MT system has committed (with a view to the system being improved in the future).

Figure: Translators' Workstation Previewing Output

So that, in outline, is how MT fits into the commercial translation process. Let us review the individuals, entities and
processes involved. Proceeding logically, we have as individuals:
● Documentation managers, who specify company policy on documentation.

● Authors of texts who (ideally) write with MT in mind, following certain established guidelines.

● Translators who manage the translation system in all respects pertaining to its day to day operation and its linguistic
performance.
In many cases the document management role will be fulfilled by translators or technical authors. For obvious reasons, there
will be fairly few individuals who are both technical authors and translators.

The important entities in the process are:

http://privatewww.essex.ac.uk/~doug/book/node12.html (4 de 5) [20/09/2000 18:02:32]

The Scenario

● Multi-Lingual Electronic Documents which contain text for translation.

● The Document Preparation system which helps to create, revise, distribute and archive electronic documents .
● The Translation System which operates on source text in a document to produce a translated text of that document.
Clearly any translation system is likely to be a very complex and sophisticated piece of software; its design at the linguistic
level is discussed in detail in other chapters in this book. A detailed discussion of Electronic Documents can be found in
Chapter .
Finally, the various processes or steps in the whole business are:
● Document Preparation (which includes authoring and pre-editing ).

● The Translation Process, mediated by the translation system, perhaps in conjunction with the translator.

● Document Revision (which is principally a matter of post-editing by the translator).

The scenario gave a brief flavour of all three steps. We shall now examine each of them in rather more detail.

Next: Document Preparation: Authoring Up: Machine Translation in Previous: Introduction

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node12.html (5 de 5) [20/09/2000 18:02:32]

Document Preparation: Authoring and Pre-Editing

Next: The Translation Process Up: Machine Translation in Previous: The Scenario

Document Preparation: Authoring

and Pre-Editing
The corporate language policy as described in the scenario tries to ensure that text which is
submitted to an MT system is written in a way which helps to achieve the best possible raw MT
output. A human translator will often be able to turn a badly written text into a well written
translation; an MT system certainly will not. Bad input means bad output. Exactly what
constitutes good input will vary a little from system to system. However, it is easy to identify
some simple writing rules and strategies that can improve the performance of almost any
general-purpose MT system. Here are some example rules:

Our example rules indicate sentences should be short. This is because MT systems find it
difficult to analyse long sentences quickly or --- more importantly --- reliably. Lacking a human
perspective, the system is always uncertain about the correct way to analyse a sentence; as the
sentence gets longer, the number of uncertainties increases rather dramatically.
Sentences should also be grammatical, and at the same time not contain very complicated
grammatical constructions. Whether or not an MT system uses explicit grammatical rules in
order to parse the input, correct, uncomplicated sentences are always easier to translate
Some MT systems use linguistic knowledge to analyse the input sentences, others do not. In
both cases correct, uncomplicated input sentences will enhance the translation performance
because unnecessary translation problems are avoided. For example, the second piece of text
below is more likely to be successfully translated than the first:
●

❍ New toner units are held level during installation and, since they do not as supplied
contain toner, must be filled prior to installation from a toner cartridge.
❍ Fill the new toner unit with toner from a toner cartridge. Hold the new toner unit

http://privatewww.essex.ac.uk/~doug/book/node13.html (1 de 3) [20/09/2000 18:02:33]

Document Preparation: Authoring and Pre-Editing

level while you put it in the printer.

The subclauses in the first sentence have been separated out as independent sentences in the
second piece of text. The latter gives the instructions as a simple series of imperatives, ordered
in the same way as the operations themselves.
The two final points in the list of writing rules prevent mistranslations by reducing potential
sources of ambiguity . Many MT systems can do a reasonable job of selecting a correct
interpretation of an ambiguous word in some circumstances, but they are unlikely to do this
successfully in all cases. (For example, ETRANS failed to get the correct interpretation of the
two different occurrences of Seite (i.e. `side' or `page') in the passage above.) Problems of
ambiguity are extensively discussed in later chapters.
Restricting MT input according to simple writing rules like the ones given above can greatly
enhance the performance of an MT system. But this is not the only advantage: it can also
improve the understandability of a text for human readers. This is a desirable feature in, for
example, technical texts and instruction manuals. As a consequence, several large companies
have developed and extended the idea of writing rules, including limited vocabulary, in order to
produce restricted forms of English suitable for technical texts. These restricted forms are
known as controlled languages . We will discuss controlled languages in detail in Chapter .
In the past few years special tools have become available for supporting the production of text
according to certain writing rules. There are spelling checkers and grammar checkers which
can highlight words that are spelled incorrectly, or grammatical errors. There are also
critiquing systems which analyse the text produced by an author and indicate where it
deviates from the norms of the language. For example, given the example above of an
over-complex sentence in a printer manual, such a tool might produce the following output:

This is a rather sophisticated analysis of various violations found in the sentence. The
controlled language this critiquing system is designed for only sanctions the word installation
if it refers to some concrete object, as in Remove the forward wheel hydraulic installation; in
this particular case installation is being used to denote the process of installing something. For
the time being, this type of analysis is too advanced for most critiquing systems, which would
find the sentence too difficult to analyse and would simply note that it is too long, not
analysable, and contains the unknown word prior.
Critiquing systems ensure that texts are written according to a set of writing rules or the rules of
a controlled language and thus help to catch errors which might upset an MT system. As a
consequence they reduce the amount of time necessary for post-editing machine translated
texts. They also reduce the time that someone else would normally have to spend on checking
and revising the input text.
There is no theoretical reason why a controlled language critiquing system could not be
completely integrated with an MT system designed to handle the controlled language --- so that
the translation system itself produces the critique while analysing the text for the purpose of
translation. In fact, if the MT system and the critiquing system are completely separate, then the
same piece of text will always have to be analysed twice --- once by the critiquing system and a
second time by the MT system. Moreover, the separation means that the same controlled
language rules and electronic dictionary entries are repeated twice --- once for each

http://privatewww.essex.ac.uk/~doug/book/node13.html (2 de 3) [20/09/2000 18:02:33]

Document Preparation: Authoring and Pre-Editing

component. This makes it more expensive to revise or alter the controlled language . For these
reasons, we can expect that MT system suppliers will seek to integrate controlled language
critiquing and controlled language MT as closely as possible.
Of course, in practice not all text submitted to MT systems is (or can be, or should be) written
according to a set of writing rules. Although this is not necessarily problematic it should be
borne in mind that the less a text conforms to the rules mentioned above, the worse the raw
translation output is likely to be. There will be a cutoff point where the input text is so badly
written or so complicated that the raw output requires an uneconomically large amount of
post-editing effort. In this case it may be possible to rewrite the problematic sentences in the
input text or it may prove simplest to do the whole thing by hand.

Next: The Translation Process Up: Machine Translation in Previous: The Scenario

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node13.html (3 de 3) [20/09/2000 18:02:33]

The Translation Process

Next: Dictionary-Based Translation Support Up: Machine Translation in Previous: Document

Preparation: Authoring

The Translation Process

In the scenario we sketched above, the source text or some selected portion thereof was passed
to the translation system which then produced raw translated output without any further human
intervention. In fact, this is merely one of many ways the translation step can proceed.

● Dictionary-Based Translation Support Tools

● Interaction in Translation

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node14.html [20/09/2000 18:02:34]

Dictionary-Based Translation Support Tools

Next: Interaction in Translation Up: The Translation Process Previous: The Translation
Process

Dictionary-Based Translation Support Tools

One point to bear in mind is that translation support can be given without actually providing full
automatic translation. All MT systems are linked to electronic dictionaries which, for the
present discussion, we can regard as sophisticated variants of their paper cousins. Such
electronic dictionaries can be of immense help even if they are supplied or used without
automatic translation of text. Here is one possible scenario:
You are translating a text by hand. Using a mouse or the keyboard, you click on a
word in the source text and a list of its possible translations is shown on screen.
You click on the possible translation which seems most appropriate in the context
and it is inserted directly into the target language text. Since you usually do this
before you start typing in the translation of the sentence which contains the
unknown work, the inserted word is inserted in the middle of an otherwise blank
target language sentence. You then type in the rest of the translation around this
inserted word.
Since technical texts typically contain contain large number of terms, and their preferred
translations are not always remembered by the translator, this simple form of support can save
a lot of time. It also helps to ensure that terms are consistently translated.
This click to see, click to insert facility is useful in dealing with low-frequency words in the
source text. In technical text, technical terms --- which can be complex multi-word units such as
faceplate delivery hose clip --- will usually have only one translation in the target language. If
the electronic dictionary has a list of terms and their translations, those translations can be
directly inserted into the target text. This gives the following scenario:
You are translating a technical text by hand. You click on the icon Term Support
and all the source language terms in the current text unit which are recognised as
being in the electronic term dictionary are highlighted. A second click causes all
the translations of those terms to be inserted in otherwise empty target language
sentences. You then type in the rest of the translation around each inserted term.
Dictionary-based translation support tools of this sort depend on two things:
1. The required terms and words must be available in the electronic dictionary . This may
well require that they were put there in the first place by translators in the organization
using the tool.
2. There must be some simple means for dealing with the inflections on the ends of words
since the form of a word or term in the text may not be the same as the cited form in the
dictionary. As a simple example, the text may contain the plural form faceplate delivery
hose clips rather than the singular form kept in the dictionary. The problem is more
complex with verb inflections and in languages other than English.

http://privatewww.essex.ac.uk/~doug/book/node15.html (1 de 2) [20/09/2000 18:02:35]

Dictionary-Based Translation Support Tools

These and other issues concerning the MT dictionary will be discussed in Chapter .

Next: Interaction in Translation Up: The Translation Process Previous: The Translation
Process

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node15.html (2 de 2) [20/09/2000 18:02:35]

Interaction in Translation

Next: Document Revision Up: The Translation Process Previous: Dictionary-Based

Translation Support

Interaction in Translation
MT systems analyse text and must decide what its structure is. In most MT systems, where
there are doubts and uncertainties about the structure, or about the correct choice of word for a
translation, they are resolved by appeal to in-built rules-of-thumb --- which may well be wrong
for a particular case. It has often been suggested that MT systems could usefully interact with
translators by pausing from time to time to ask simple questions about translation problems.
Another sort of interaction could occur when the system has problems in choosing a correct
source language analysis ; a good analysis is needed to ensure good translation. For example,
suppose that a printer manual being translated from English contains the following sentence:
● the printer to the PC with a parallel interface cable.

The question is: are we talking about a particular type of PC (personal computer) which comes
with a parallel interface cable (whatever that is) or any old PC which can be connected to the
printer by means of an independent parallel interface cable? In the first case, the with, in the
phrase with a parallel interface cable means having or fitted with and modifies the noun PC,
whilst in the second it means using and modifies the verb attach. One good reason for worrying
about the choice is because in many languages with will be translated differently for the two
cases. Faced with such an example, an MT system might ask on screen exactly the same
question:

● modify or does it modify ?

Another sort of analysis question arises with pronouns. Consider translating the following:
● the paper in the paper tray and replace the cover. Ensure that is completely
closed.
Does it in the second sentence refer to the paper, the paper tray, or the cover? The decision
matters because the translation of it in many languages will vary depending on the gender of the
expression it refers back to. Making such a decision depends on rather subtle knowledge, such
as the fact that covers, but not trays or paper are typical things to be closed, which is hard
perhaps impossible to build into an MT system. However, it is the sort of question that a human
translator may be able to answer.
The following is a possible scenario:
You are translating a text interactively with an MT system. The system displays
the source text in one window, while displaying the target text as it is produced in
another. On encountering the word it, the system parses, highlights the words
paper, paper tray, and cover in the first sentence, and asks you to click on the one
which is the antecedent (i.e. the one it refers back to). It is then able to choose the

http://privatewww.essex.ac.uk/~doug/book/node16.html (1 de 2) [20/09/2000 18:02:36]

Interaction in Translation

appropriate form of the translation, and it proceeds with the rest of the sentence.
It is hardly surprising that a machine may need to ask such questions because the answers may
not be at all clear, in some cases even for a human translator. With poorly written technical
texts, it may even be the case that only the author knows.

Next: Document Revision Up: The Translation Process Previous: Dictionary-Based

Translation Support

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node16.html (2 de 2) [20/09/2000 18:02:36]

Document Revision

Next: Summary Up: Machine Translation in Previous: Interaction in Translation

Document Revision
The main factor which decides the amount of post-editing that needs to be done on a
translation produced by machine is of course the quality of the output. But this itself depends
on the requirements of the client, in particular (a) the translation aim and (b) the time available.
In the case of the printer manual in the scenario above the translation aim was to provide a
printer manual in English for export purposes. The fact that the translation was going to be
widely distributed outside the organization required it to be of high quality --- a correct,
well-written and clear piece of English text, which means thorough and conscientious
post-editing.
The opposite situation occurs when a rough and ready translation is needed out of some
language for personal or internal use, perhaps only to get the gist of some incoming text to see
if it looks interesting enough for proper translation. (If it is not, little time or money or effort
has been wasted finding out). Here is the sort of scenario in which it might work:
You are an English-speaking agronomist monitoring a stream of information on
cereal crop diseases coming in over global computer networks in four different
languages. You have a fast MT system which is hooked into the network and
translates --- extremely badly --- from three of the languages into English. Looking
at the output and using your experience of the sort of things that reports contain,
you should be able to get enough of an idea to know whether to ignore it or pass it
on to your specialist translators.
Of course, in this situation it is the speed of the MT system, not its quality that matters --- a
very simple system that does no more than transliterate and translate a few of the words may
even be enough.
We've now looked at two cases: one in which full post-editing needed to be done, one in which
no post-editing whatsoever was required. Another option could be to do some post-editing on a
translation in order to make it easy to read and understand, but without having the perfection
of a published text in mind. Most post-editors are also translators and are used to producing
high quality texts. They are likely to apply the same sort of output standards to their
translations produced automatically. Though this policy is very desirable for, for instance,
business correspondence and manuals, it is not at all necessary to reach the same sort of
standard for internal electronic mail. Some MT output could be subject to a rough and ready
post-edit --- where the post-editor tries to remove or adjust only the grossest errors and
incomprehensibilities --- rather than the usual thorough and painstaking job. The main
advantage of this option is that translator time is saved. Even if documents are occasionally
sent back for re-translation or re-editing, the rough and ready post-edit policy might still save
money overall. Again, the factors of translation aim and time available play an important role.
MT systems make the same sorts of translation mistake time and time again. Sometimes these

http://privatewww.essex.ac.uk/~doug/book/node17.html (1 de 2) [20/09/2000 18:02:37]

Document Revision

errors can be eliminated by modifying the information in the dictionary. Other sorts of errors
may stem from subtle problems in the system's grammars or linguistic processing strategies
which cannot ordinarily be resolved without specialist knowledge. Once an error pattern has
been recognised, a translator can scan text looking for just such errors. If the error is just a
matter of consistently mistranslating one word or string of words, then --- as in the scenario ---
the ordinary search-and-replace tools familiar from word processors will be of some help. In
general, since the errors one will find in machine translated texts are different from those one
finds in other texts, specialized word processor commands may be helpful. For example,
commands which transpose words, or at a more sophisticated level, ones which change the
form of a single word, or all the words in a certain region from masculine to feminine, or
singular to plural, might be useful post-editing tools .
The imaginary company that we have been discussing in the previous sections deals with large
volumes of similar, technical text. This text similarity allows the MT system to be tuned in
various ways, so as to achieve the best possible performance on one particular type of text on
one particular topic. An illustration of this can be found in the section heading of our example
text Einstellung der Druckdichte. The German word Einstellung can have several translations:
employment, discontinuation, adjustment and attitude. Since we are dealing here with technical
texts we can discard the first and last possible translations. Of the two translations left,
adjustment, is the most common one in this text type, and the computer dictionaries as
originally supplied have been updated accordingly. The tuning of a system takes time and
effort, but will in the long run save post-editing time.
Obviously enough, the difficulty of post-editing and the time required for it correlates with the
quality of the raw MT output: the worse the output, the greater the post-edit effort. For one
thing, the post-editor will need to refer more and more to the source language text when the
output gets less intelligible . Even though this seems to be a major drawback at the beginning,
bear in mind that post-editors will get used to the typical error patterns of the MT system; MT
output that may seem unintelligible at the beginning will require less reference to the source
language text after some time. Familiarity with the pattern of errors produced by a particular
MT system is thus an important factor in reducing post-editing time. More generally, familiarity
with the document processing environment used for post-editing and its particular facilities is
an important time saver.

Next: Summary Up: Machine Translation in Previous: Interaction in Translation

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node17.html (2 de 2) [20/09/2000 18:02:37]

Summary

Next: Further Reading Up: Machine Translation in Previous: Document Revision

Summary
This chapter has given a picture of how MT might be used in an imaginary company, and
looked in outline at the typical stages of translation: document preparation, translation
(including various kinds of human involvement and interaction) , and document revision , and
at the various skills and tools required. In doing this we have tried also to give an idea of some
of the different situations in which MT can be useful. In particular, the case of `gist' translation,
where speed is important, and quality less important, compared to the case where a translation
is intended for widespread publication, and the quality of the finished (post-edited) product is
paramount. These are all matters we will return to in the following chapters.

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node18.html [20/09/2000 18:02:38]

Further Reading

Next: Representation and Processing Up: Machine Translation in Previous: Summary

Further Reading
Descriptions of how MT is actually used in corporate settings can be found in the Proceedings
of the Aslib Conferences (normally subtitled Translating and the Computer) which we
mentioned in the Further Reading section of Chapter .
For readers interested in finding out more about the practicalities of pre- and post-editing ,
there are several relevant contribution in [Vasconcellos1988] , in [Lawson1982a] . There is a
useful discussion of issues in pre-editing and text preparation, in [Pym1990], and we will say
more about some related issues in Chapter .
An issue that we have not addressed specifically in this chapter is that of machine aids to
(human) translation , such as on-line and automatic dictionaries and terminological databases ,
multilingual word processors, and so on. We will say more about terminological databases in
Chapter . Relevant discussion of interaction between machine (and machine aided)
translation systems and human users can be found in
[Vasconcellos1988] ,[Stoll1988],[Knowles1990] and various papers by Alan Melby ,
including [Melby1987,Melby1992], who discusses the idea of a `translator's workbench'. In
fact, it should be clear that there is no really hard and fast line that can be drawn between such
things and the sort of MT system we have described here. For one thing, an adequate MT
system should clearly include such aids in addition to anything else. In any case, in the kind of
setting we have described, there is a sense in which even an MT system which produces very
high quality output is really serving as a translators' aid, since it is helping improve their
productivity by producing draft translations. What are sometimes called distinction between
`Machine Aided Human Translation' , `Human Aided Machine Translation' , and `Machine
Translation' per se actually form a continuum.

Next: Representation and Processing Up: Machine Translation in Previous: Summary

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node19.html [20/09/2000 18:02:38]

Representation and Processing

Next: Introduction Up: Machine Translation Previous: Further Reading

Representation and Processing

● Introduction
● Representing Linguistic Knowledge
❍ Grammars and Constituent Structure
❍ Further Analysis: Grammatical Relations
❍ Meaning
● Processing
❍ Parsing
❍ Generation
● Summary
● Further Reading

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node20.html [20/09/2000 18:02:39]

Introduction

Next: Representing Linguistic Knowledge Up: Representation and Processing Previous:

Representation and Processing

Introduction
In this chapter we will introduce some of the techniques that can be used to represent the kind
of information that is needed for translation in such a way that it can be processed
automatically. This will provide some necessary background for Chapter , where we
describe how MT systems actually work.
Human Translators actually deploy at least five distinct kinds of knowledge:
● Knowledge of the source language.

● Knowledge of the target language. This allows them to produce texts that are acceptable
in the target language.
● Knowledge of various correspondences between source language and target language (at
the simplest level, this is knowledge of how individual words can be translated).
● Knowledge of the subject matter, including ordinary general knowledge and `common
sense' . This, along with knowledge of the source language, allows them to understand
what the text to be translated means.
● Knowledge of the culture, social conventions, customs, and expectations, etc. of the
speakers of the source and target languages.
This last kind of knowledge is what allows translators to act as genuine mediators, ensuring that
the target text genuinely communicates the same sort of message, and has the same sort of
impact on the reader, as the source text. Since no one has the remotest idea how to represent
or manipulate this sort of knowledge, we will not pursue it here --- except to note that it is the
lack of this sort of knowledge that makes us think that the proper role of MT is the production
of draft or `literal' translations.
Knowledge of the target language is important because without it, what a human or automatic
translator produces will be ungrammatical, or otherwise unacceptable. Knowledge of the source
language is important because the first task of the human translator is to figure out what the
words of the source text mean (without knowing what they mean it is not generally possible to
find their equivalent in the target language).
It is usual to distinguish several kinds of linguistic knowledge:
● Phonological knowledge : knowledge about the sound system of a language, knowledge
which, for example, allows one to work out the likely pronunciation of novel words.
When dealing with written texts, such knowledge is not particularly useful. However,
there is related knowledge about orthography which can be useful. Knowledge about
spelling is an obvious example.
● Morphological knowledge : knowledge about how words can be constructed: that printer

http://privatewww.essex.ac.uk/~doug/book/node21.html (1 de 2) [20/09/2000 18:02:39]

Introduction

is made up of print + er.

● Syntactic knowledge : knowledge about how sentences, and other sorts of phrases can be
made up out of words.
● Semantic knowledge : knowledge about what words and phrases mean, about how the
meaning of a phrase is related to the meaning of its component words.
Some of this knowledge is knowledge about individual words, and is represented in
dictionaries . For example, the fact that the word print is spelled the way it is, that it is not
made up of other words, that it is a verb, that it has a meaning related to that of the verb write,
and so on. This, along with issues relating to the nature and use of morphological knowledge,
will be discussed in Chapter .
However, some of the knowledge is about whole classes or categories of word. In this chapter,
we will focus on this sort of knowledge about syntax and semantics . Sections , and
discuss syntax, issues relating to semantics are considered in Section . We will look first on
how syntactic knowledge of the source and target languages can be expressed so that a
machine can use it. In the second part of the chapter, we will look at how this knowledge can be
used in automatic processing of human language.

Next: Representing Linguistic Knowledge Up: Representation and Processing Previous:

Representation and Processing

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node21.html (2 de 2) [20/09/2000 18:02:39]

Representing Linguistic Knowledge

Next: Grammars and Constituent Up: Representation and Processing Previous: Introduction

Representing Linguistic Knowledge

In general, syntax is concerned with two slightly different sorts of analysis of sentences. The
first is constituent or phrase structure analysis --- the division of sentences into their
constituent parts and the categorization of these parts as nominal, verbal, and so on. The
second is to do with grammatical relations; the assignment of grammatical relations such as
SUBJECT, OBJECT, HEAD and so on to various parts of the sentence. We will discuss these
in turn.

● Grammars and Constituent Structure

● Further Analysis: Grammatical Relations
● Meaning

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node22.html [20/09/2000 18:02:40]

Grammars and Constituent Structure

Next: Further Analysis: Grammatical Up: Representing Linguistic Knowledge Previous: Representing
Linguistic Knowledge

Grammars and Constituent Structure

Sentences are made up of words, traditionally categorised into parts of speech or categories including nouns,
verbs, adjectives, adverbs and prepositions (normally abbreviated to N, V, A, ADV, and P). A grammar of a
language is a set of rules which says how these parts of speech can be put together to make grammatical, or
`well-formed' sentences.
For English, these rules should indicate that (

a) is grammatical, but that (

b) is not (we indicate this by marking it with a `*').
● ❍ Put some paper in the printer.

❍ *Printer some put the in paper.

Here are some simple rules for English grammar, with examples. A sentence consists of a noun phrase, such
as the user followed by a modal or an auxiliary verb, such as should, followed by a verb phrase, such as
clean the printer:
● ❍ The user should clean the printer.
A noun phrase can consist of a determiner, or article, such as the, or a, and a noun, such as printer ( a). In
some circumstances, the determiner can be omitted ( b).
● ❍ the printer

❍ printers

`Sentence', is often abbreviated to S, `noun phrase' to NP, `verb phrase' to VP, `auxiliary' to AUX, and
`determiner' to DET. This information is easily visualized by means of a labelled bracketing of a string of
words , as follows, or as a tree diagram, as in Figure .
● ❍ Users should clean the printer.

http://privatewww.essex.ac.uk/~doug/book/node23.html (1 de 6) [20/09/2000 18:02:49]

Grammars and Constituent Structure

users

should

clean

the printer

Figure: A Tree Structure for a Simple Sentence

The auxiliary verb is optional, as can be seen from ( ), and the verb phrase can consist of just a verb (such as
stopped):
● ❍ The printer should stop.

❍ The printer stopped.

NP and VP can contain prepositional phrases (PPs), made up of prepositions ( on, in, with, etc.) and NPs:
● ❍ The printer stops .
❍ Put the cover .
❍ Clean the printer .
The reader may recall that traditional grammar distinguishes between phrases and clauses. The phrases in the

http://privatewww.essex.ac.uk/~doug/book/node23.html (2 de 6) [20/09/2000 18:02:49]

Grammars and Constituent Structure

examples above are parts of the sentence which cannot be used by themselves to form independent sentences.
Taking The printer stopped, neither its NP nor its VP can be used as independent sentences:
● ❍ *The printer

❍ *Stopped

By contrast, many types of clause can stand as independent sentences. For example, ( a) is a sentence which
consists of a single clause --- The printer stopped. As the bracketing indicates, ( b) consists of two clauses
co-ordinated by and. The sentence ( c) also consists of two clauses, one ( that the printer stops) embedded in
the other, as a sentential complement of the verb.
● ❍ The printer stopped

❍ The printer stopped and the warning light went on

❍ You will observe that the printer stops .

There is a wide range of criteria that linguists use for deciding whether something is a phrase, and if it is, what
sort of phrase it is, what category it belongs to. As regards the first issue, the leading idea is that phrases consist
of classes of words which normally group together. If we consider example ( ) again ( The user should clean
the printer), one can see that there are good reasons for grouping the and user together as a phrase, rather than
grouping user and should. The point is the and user can be found together in many other contexts, while user
and should cannot.
● ❍ A full set of instructions are supplied to .
❍ must clean the printer with care.
❍ It is who is responsible for day-to-day maintenance.
❍ *User should clean the printer.
As regards what category a phrase like the user belongs to, one can observe that it contains a noun as its
`chief' element (one can omit the determiner more easily than the noun), and the positions it occurs in are also
the positions where one gets proper nouns (e.g. names such as Sam). This is not to say that questions about
constituency and category are all clear cut. For example, we have supposed that auxiliary verbs are part of the
sentence, but not part of the VP. One could easily find arguments to show that this is wrong, and that should
clean the printer should be a VP, just like clean the printer, giving a structure like the following, and Figure
:
● ❍

users

should

http://privatewww.essex.ac.uk/~doug/book/node23.html (3 de 6) [20/09/2000 18:02:49]

Grammars and Constituent Structure

clean

the printer

Moreover, from a practical point of view, making the right assumptions about constituency can be important,
since making wrong ones can lead to having to write grammars that are much more complex than otherwise.
For example, suppose that we decided that determiners and nouns did not, in fact, form constituents. Instead of
being able to say that a sentence is an NP followed by an auxiliary, followed by a VP, we would have to say
that it was a determiner followed by an noun, followed by a VP. This may not seem like much, but notice that
we would have to complicate the rules we gave for VP and for PP in the same way. Not only this, but our rule
for NP is rather simplified, since we have not allowed for adjectives before the noun, or PPs after the noun. So
everywhere we could have written `NP', we would have to write something very much longer. In practice, we
would quickly see that our grammar was unnecessarily complex, and simplify it by introducing something like
an NP constituent.

Figure: An Alternative Analysis

For convenience linguists often use a special notation to write out grammar rules. In this notation, a rule
consists of a `left-hand-side' (LHS) and a `right-hand-side' (RHS) connected by an arrow ( ):
S NP (AUX) VP VP V (NP) PP* NP (DET) (ADJ) N PP* PP P
NP N user N

http://privatewww.essex.ac.uk/~doug/book/node23.html (4 de 6) [20/09/2000 18:02:49]

Grammars and Constituent Structure

users N printer N printers V clean V cleans AUX

should DET the DET a P with
The first rule says that a Sentence can be rewritten as (or decomposes into, or consists of) an NP followed by
an optional AUX, followed by VP (optionality is indicated by brackets). Another rule says that a PP can consist
of a P and an NP. Looked at the other way, the first rule can be interpreted as saying that an NP, and AUX and
a VP make up a sentence. Items marked with a star (`*') can appear any number of times (including zero) --- so
the second rule allows there to be any number of PPs in a VP. The rules with `real words' like user on their
RHS serve as a sort of primitive dictionary. Thus the first one says that user is a noun, the fifth one that clean is
a verb. Since the NP rule says that an N by itself can make up an NP, we can also infer that printers is an NP,
and since (by the VP rule) a V and an NP make up a VP, clean printers is a VP. Thus, a grammar such as this
gives information about what the constituents of a sentence are, and what categories they belong to, in the
same way as our informal rules at the start of the section.

Returning to the tree representation in Figure , each node in the tree (and each bracketed part of the string
representation) corresponds to the LHS of a particular rule, while the daughters of each node correspond to the
RHS of that rule. If the RHS has two constituent s, as in NP DET N, there will be two branches and two
daughters; if there are three constitituents, there will be three branches and three daughters, and so on.

It is worthwhile to have some terminology for talking about trees. Looking from the top, the trees above
start from (or `are rooted in') a sentence node --- the LHS of our sentence rule. Near the bottom of the trees, we
have a series of nodes corresponding to the LHS's of dictionary rules and, immediately below them at the very
bottom of the trees, actual words from the corresponding RHS's of the dictionary rules. These are called the
`leaves' or terminal nodes of the tree. It is normal to speak of `mother' nodes and `daughter' nodes (e.g. the S
node is the mother of the NP, AUX, and VP nodes), and of mothers `dominating' daughters. In practice most
sentences are longer and more complicated than our example. If we add adjectives and prepositional phrases,
and some more words, more complex tree s can be produced, as shown in Figure , where the NP which is
the left daughter of the S node contains an adjective and a noun but no determiner (the NP rule in our grammar
above allows for noun phrases of this form), the NP in VP contains a determiner and a PP.

http://privatewww.essex.ac.uk/~doug/book/node23.html (5 de 6) [20/09/2000 18:02:49]

Grammars and Constituent Structure

A large collection of such rules will constitute a formal grammar for a language --- formal, because it attempts
to give a mathematically precise account of what it is for a sentence to be grammatical. As well as being more
concise than the informal descriptions at the beginning of the section, the precision of formal grammars is an
advantage when it comes to providing computational treatments.

Figure: A More Complex Tree Structure

We should emphasise that the little grammar we have given is not the only possible grammar for the fragment
of English it is supposed to describe. The question of which grammar is `best' is a matter for investigation. One
question is that of completeness -- does the grammar describe all sentences of the language? In this respect, one
can see that our example above is woefully inadequate. Another issue is whether a grammar is correct in the
sense of allowing only sentences that are in fact grammatical: our example grammar falls down in this respect,
since it allows the examples in ( ), among many others.
● ❍ *User cleans.

❍ *Users cleans printers.

❍ *Users should cleans printers.

A grammar may also be incorrect in associating constituents with the wrong categories. For example, as we
noted above, one would probably prefer a grammar which recognizes that determiners and nouns make up NPs,
and that the NP that occur in S (i.e. subject NPs) and those that appear in VP (object NPs) are the same (as our
grammar does) to a grammar which treats them as belonging to different categories --- this would suggest
(wrongly) that there are things that can appear as subjects, but not as objects, and vice versa. This is obviously
not true, except for some pronouns that can appear as subjects but not as objects: I, he, she, etc. A worse defect
of this kind is the treatment of words -- the grammar gives far too little information about them, and completely
misses the fact that clean, and cleans are actually different forms of the same verb. We will show how this
problem can be overcome in Chapter .
In a practical context, a further issue is how easy it is to understand the grammar, and to modify it (by
extending it, or fixing mistakes), and how easy it is to use it for automatic processing (an issue to which we will
return). Of course, all these matters are often related.

Next: Further Analysis: Grammatical Up: Representing Linguistic Knowledge Previous: Representing
Linguistic Knowledge

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node23.html (6 de 6) [20/09/2000 18:02:49]

Further Analysis: Grammatical Relations

Next: Meaning Up: Representing Linguistic Knowledge Previous: Grammars and Constituent

Further Analysis: Grammatical Relations

So far we have talked about the kind of grammatical knowledge that can be expressed in terms
of a constituent structure tree --- information about the constituent units, and the parts of
speech. But there are other kinds of information implicit in these representations which it is
useful to make explicit. In particular, information about which phrases fulfil which grammatical
relations or grammatical functions such as SUBJECT, OBJECT and SENTENTIAL
COMPLEMENT. English SUBJECTs are normally the NPs which come before the verb, and
OBJECTs normally occur immediately after the verb. In other languages these relations may be
realised differently with respect to the verb. For example, in Japanese the normal word order is
SUBJECT OBJECT VERB, and in Irish and Welsh it is VERB SUBJECT OBJECT. In many
languages, such as Russian , the VERB, SUBJECT and OBJECT can appear in essentially any
order. (In such languages the different grammatical relations can often be recognized by
different forms of the noun -- usually called cases. In English, this only occurs with pronouns
--- he, she, etc., are only possible as SUBJECTs). What this suggests, of course, is that while
the constituent structures of languages differ greatly, they may appear more similar when
described in terms of grammatical relations .
Phrases which serve as SUBJECT, OBJECT, etc., should also be distinguished from those
which serve as MODIFIERs, or ADJUNCTs, of various sorts. For example, in the sentence (
) You is the SUBJECT of the verb clean, the printer casing is its OBJECT, whilst the
prepositional phrases with a non-abrasive compound and at any time are ADJUNCTs.
● can clean the printer casing with a non-abrasive compound at any time.

ADJUNCTs are prototypically optional --- unlike SUBJECTs. For example, a sentence which
omits them is still perfectly well formed: there is nothing wrong with ( a), but omitting the
SUBJECT, as illustrated in ( b) produces an ungrammatical result.
● ❍ You can clean the printer casing.

❍ *Can clean the printer casing.

There are various ways of representing sentences in terms of grammatical relations , but it is
essentially not very different from that of constituent structure tree representation, which we
have seen earlier in this chapter. The basic idea is to represent sentences in terms of their
constituent parts (so a tree representation is convenient), but since one wants to represent the
grammatical relation which the parts have to the whole, it is common to mark either the
branches or the nodes with the appropriate relation. Figure gives a representation of ( ).
This can be compared with a constituent structure representation for the same sentence in
Figure .

http://privatewww.essex.ac.uk/~doug/book/node24.html (1 de 4) [20/09/2000 18:02:53]

Further Analysis: Grammatical Relations

● temperature has affected the printer.

Figure: A Representation of Grammatical Relations

Figure: A Constituent Structure Representation

In Figure , the relations are marked on the nodes, and a new relation HEAD has been
introduced. The HEAD element is, intuitively, the most important element from the point of
view of the grammar of the whole phrase --- the element which makes the phrase what it is.
This is the noun in an NP, the verb in a VP or sentence, the preposition in a PP.
There are three important differences between this tree representing grammatical relations , and
those representing constituent structure . First, instead of consisting of an NP, and a VP
(containing a V and an NP), the representation of grammatical relations consists of a V and two
NPs -- the VP node has disappeared. Second, in this grammatical relations representation, the
order of the branches is unimportant. This is possible, of course, because the grammatical
relations have been indicated and this gives information about word order implicitly. Figure
could be redrawn with the branches in any order, and it would still be a representation of The
temperature affects the printer, since this is the only sentence that has these items with these
relations. By contrast, reordering the branches in a constituent structure tree might produce a
representation of a quite different sentence, or no sentence at all.

The third difference is that some of the words have been missed out from Figure , and have
been replaced by features , that is, pairs that consist of an attribute, such as def, tense, and
aspect, and a value, such as +, pres, and perfective. The features
aspect=perfective and tense=pres indicate that the sentence as a whole is in the
present perfect tense. It is called perfect because it is used to describe events or actions that
have been `perfected' or completed, unlike, for example, a sentence such as The temperature
was affecting the printer, where the `affecting' is still going on at the time the writer is referring
to. It is called present perfect because the auxiliary verb is in a present tense form ( has not
had). The feature def=+ on the NPs means these NPs are definite. This definiteness indicates
that the writer and reader have some particular object of the appropriate kind in mind. Compare,
for example, The printer has stopped where one particular printer which is in some sense
known to both writer and reader is being discussed, with A printer has stopped, where this is
not the case.
These three differences are all intended to represent what is expressed by the sentence,
abstracting away from the way it is expressed: we abstract away from the division into NP and
VP, from the particular word order, and from the way in which the definiteness of the NPs and
the tense and aspect of the sentence are realized (in English it is by the determiners, and the
auxiliary verb respectively; in other languages it might be realized differently).
When it comes to describing the relationship between constituent structure , and what we might
call relational structures, such as Figure , there are basically two approaches. One is simply
to add information about grammatical relations to the grammar rules.
S NP{SUBJECT} AUX VP{HEAD} VP V{HEAD} NP{OBJECT}
PP{ADJUNCT}* AUX has{aspect=perfective, tense=pres}

http://privatewww.essex.ac.uk/~doug/book/node24.html (2 de 4) [20/09/2000 18:02:53]

Further Analysis: Grammatical Relations

The idea is that these annotations can be interpreted in such a way that a representation like
Figure can be constructed at the same time as the constituent structure tree . To do this
requires a convention to `flatten' the constituent structure tree `merging' a structure (e.g. the
structure of S) that is associated with the LHS of a rule with that of the HEAD daughter on the
RHS, and a convention which simply merges in information that comes from items which do
not have a grammatical relation , such as the AUX.
A second approach is to have special rules which relate the constituent structure representation
to the representation of grammatical relations. One such rule might look like this:
NP:$1, AUX:$2, V:$3, NP:$4

HEAD:$3, SUBJ:$1, OBJ:$4

In this rule, $1, $2, etc. are variables, or temporary names for pieces of structure. The idea is
that such a rule matches a constituent structure such as that in Figure , and assigns (or
`binds') the variables to various pieces of structure. For example the NP containing temperature
becomes bound to the variable $1. The rule can then be interpreted as an instruction to
transform the constituent structure tree into a tree like Figure . This involves making this NP
into the SUBJECT, making the V into the HEAD, and missing out the AUX entirely, among
other things. The rule is rather simplified, of course, since it does not mention putting the
information about perfective aspect into the grammatical relation representation, and ignores
the problem of dealing with PPs, but it should give some idea.
The reader may also notice that the arrow used in this rule is bidirectional . This is intended to
suggest that the rule simply states a correspondence between constituent structure , and
grammatical relation representations, without suggesting that one is prior to the other. Thus, the
idea is that one could equally well use the rule to transform Figure into Figure and vice
versa. Similarly, the annotation approach is not supposed to be directional (though this may be
somewhat harder to appreciate).
Many verbs have what are called active and passive forms, as in the following.

http://privatewww.essex.ac.uk/~doug/book/node24.html (3 de 4) [20/09/2000 18:02:53]

Further Analysis: Grammatical Relations

● ❍ Temperature affects . ( Active)

❍ are affected by temperature. ( Passive)
Notice that the object in the active sentence corresponds to the subject in the passive. This
raises the question of what the grammatical relations SUBJECT and OBJECT should mean.
One possibility is to use the the terms in the sense of the `surface' grammatical relations. The
SUBJECTs of actives and the corresponding passives would be different, then. In particular,
temperature would be the SUBJECT of ( a), and printers would be the SUBJECT of ( b).
The alternative is to adopt a notion of a deep relation which picks out the same elements in both
active and passive sentence. We would then say that (in English) the D-OBJECT (`deep'
OBJECT) corresponds to the noun phrase after the verb in active sentences and to the noun
phrase that precedes the verb in the corresponding passive. In active sentences, the surface and
deep relations are the same, but they are different in passives, as can be seen from the following
(in the passive sentence there is no surface OBJECT, and the D-SUBJECT has become a sort of
ADJUNCT, in a PP with the preposition by).
● ❍ Temperature affects printers. ( Active)
SUBJECT = temperature, OBJECT = printers
D-SUBJECT = temperature, D-OBJECT = printers
❍ Printers are affected by temperature. ( Passive )
SUBJECT = printers, OBJECT = 0,
D-SUBJECT = temperature D-OBJECT = printers
Interpreting SUBJECT as deep subject is clearly consistent with the general idea of abstracting
away from surface characteristics in the grammatical relational representation. But it is not
obviously the right move to make. For example, English verbs often vary their form depending
on the nature of their subject (this is called agreement -- as the following makes clear, there is
also agreement of demonstratives like this/ these with their head noun).
● ❍ These factors affect printers.

❍ This factor affects printers.

❍ *These factors affects printers.

❍ *This factor affect printers.

However, the point to notice is that the agreement is with the surface subject, not the deep
subject. Thus, if one wants to use a representation of grammatical relations to describe the
phenomenon of agreement , the notion of SUBJECT had better be surface subject. This is not,
in itself, a critical point here. The point we are making is simply that there is a range of options,
and that the option chosen can make a difference for the overall description.

Next: Meaning Up: Representing Linguistic Knowledge Previous: Grammars and Constituent

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node24.html (4 de 4) [20/09/2000 18:02:53]

Meaning

Next: Processing Up: Representing Linguistic Knowledge Previous: Further Analysis:

Grammatical

Meaning

Representing information about grammar in the form of grammar rules is useful in two ways in
MT. First, as will become clear in the Chapter , it is possible to use the sort of linguistic
representation that the rules provide to get simpler, and better descriptions of what is involved
in translation, by abstracting away from some superficial differences between languages -- as
we have noted the abstract representations of sentences in different languages are often more
similar than the sentences themselves. But one can also use such representations as the basis for
still more abstract representations of meaning. Working out the meaning of sentences is an
important part of the translation process for human translators, and the ability to work out the
meaning --- to `understand' (in some sense) the source text would allow an MT system to
produce much better translations. This may sound an impossible task, and perhaps at some level
it is. However, there is another, less ambitious, level where automatic `understanding' is
possible. In this section we will look at what this involves in a preliminary way (we will say
more about it in Chapter ).
It is useful to think of `understanding' as involving three kinds of knowledge:
1. Semantic knowledge. This is knowledge of what expressions (individual words and
sentences) mean, independent of the context they appear in.
2. Pragmatic knowledge. This is knowledge of what expressions mean in situations and
particular occasions of use.
3. Real world , or common sense knowledge .
Consider the following example:
● ❍ The user may prefer to clean the printer every week with a non-corrosive fluid.
Do not use abrasive or corrosive solvents, as this may harm its appearance.
One thing that is involved in understanding the meaning of this is working out the different
semantic relations that the different NPs have to the predicates. For example, a non-corrosive
fluid is understood as an instrument to be used in cleaning, every week indicates the time period
in which the cleaning should be repeated, the printer denotes the thing to be cleaned, and the
user denotes both the entity that has a preference, and which performs the cleaning. This is
semantic information, because it is information that this sentence would convey on any
occasion of use. However, recovering this information is not enough to `understand' the
example. One must also be able to work out that these sentences --- or at least the second
sentence --- is to be understood as a warning not to do something. In this case, the form of the
sentence is a fairly clear guide to this, but this is not always so. For example, sentences that are
interrogative in form are often requests for information, but it is quite possible for such

http://privatewww.essex.ac.uk/~doug/book/node25.html (1 de 3) [20/09/2000 18:02:54]

Meaning

sentences to be interpreted as offers, requests for action, warnings, or as assertions (i.e. as

giving information). This last case is what is called a rhetorical question; the following
interrogatives might be interpreted in some of the other ways, depending on the context.
● ❍ Would you like some cake?

❍ Don't you think it is cold in here?

❍ Can't you see what you are doing to that printer?

Of course, the key words here are `depending on the context'. Working out, for example, that (
b) is interpreted as a request for the speaker to close a window depends on many things in the
context where it is uttered (it might also, for example, be a comment on the social atmosphere).
The sort of knowledge of social and linguistic conventions involved here is part of what is
normally thought of as pragmatic knowledge.
But even this is not enough to understand the example completely. For example, there are the
pronouns this, and it in the second sentence. It is obvious (to the human reader) that this should
be interpreted as cleaning with an abrasive or corrosive solvent, and that it should be interpreted
as referring to the printer (i.e. the sense is: `cleaning ... may harm the printer's appearance'). But
this is not the only semantically and pragmatically possible interpretation. One could imagine
the same sentence being uttered in a context where it is the appearance of the fluid that will be
affected (imagine one is dealing with a precious fluid of some kind):
● not place the fluid in sunlight, as this may harm its appearance.

What is involved here is real world , or common sense knowledge , perhaps the knowledge that
if a corrosive fluid comes into contact with a printer (or something similar), it is the printer's
appearance that is damaged. This is not knowledge about the meanings of the words, or about
how language is used in different social contexts.

Similarly, consider the meaning of a word like printers. Semantic knowledge should supply the
information that one interpretation of this refers to a collection of machines which perform the
activity of printing, or perhaps to such things in general (as in printers are expensive and
unreliable). Real world knowledge will indicate that the members of this collection are
typically of a certain size (bigger than pencils, but smaller than houses, say), and have certain
parts, and characteristic flaws. When someone utters the word Printers!, in an exasperated tone,
with a piece of chewed up paper in their hand, you may realize that what they intend to convey
is some quite complicated attitude, including annoyance. It is pragmatic knowledge that allows
you to work out that this is their intention, and that they do not, for example, want you to go
and buy them a number of printers.
Of course, the distinctions between these different kinds of knowledge are not always clear, and
they interact in complex ways in determining how an utterance is actually understood.
Nevertheless, the basic idea of the distinction should be clear.
How can this sort of information about sentences be represented? The representation of
pragmatic and common sense or real world knowledge raises many difficult problems, and is
not really necessary for understanding the discussion in the following chapters, so we will
postpone discussion until Chapter . However, we will say something about semantic
representations here.
One kind of semantic representation would provide different relation names , and indicate

http://privatewww.essex.ac.uk/~doug/book/node25.html (2 de 3) [20/09/2000 18:02:54]

Meaning

which NP had which relation. In the following example, which is a simplified part of ( ) , one
might have relations like INSTRUMENT, AGENT (for the user), and THEME or PATIENT
(for the printer), giving a representation like Figure These relations are sometimes called
semantic roles, (deep) cases, or thematic roles.
● user cleans the printer with a non-abrasive solvent.

Figure: A Representation of Semantic Relations

Such a representation looks very much like Figure , except that the labels SUBJECT,
OBJECT, etc. have been replaced by these thematic roles, and syntactic information about
tense and aspect has been replaced by information about time reference. The rules that relate
these roles to grammatical relations would say things like ``The AGENT will normally
correspond to the SUBJECT of an active sentence, and an NP in a by-phrase in a passive
sentence''; ``The INSTRUMENT normally appears in a PP with the preposition with'', ``The
PATIENT is very often the OBJECT of active sentences'' However, there are some verbs which
violate these general patterns. For example, they are very different with like and please -- the
PATIENT ( bright colours in the following example) is normally the OBJECT of like, but the
SUBJECT of please.
● ❍ Children like bright colours.

❍ Bright colours please children.

The usefulness of a semantic representation is further explored in Chapter .

Next: Processing Up: Representing Linguistic Knowledge Previous: Further Analysis:

Grammatical

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node25.html (3 de 3) [20/09/2000 18:02:54]

Processing

Next: Parsing Up: Representation and Processing Previous: Meaning

Processing
In the previous sections, we have tried to give an outline of some of the different kinds of
knowledge that are needed in text understanding (and hence, translation), and how they can be
represented. We will now give an idea of how this knowledge can be manipulated
automatically. We will do this in two stages. First, we will look at what is called analysis, or
parsing . This is the process of taking an input string of expressions, and producing
representations of the kind we have seen in the previous section. Second, we will look at
synthesis, or generation , which is the reverse process -- taking a representation, and producing
the corresponding sentence.
It may be helpful to point out at the beginning that though the representations we have given are
generally graphic objects --- tree s or networks drawn with lines --- these are not themselves the
representations that the computer deals with. For example, the standard internal representation
of a tree is as a list, containing sublists, with any labels on a node being represented as the first
element of the list. If we write lists between `(' and `)', and separate elements with commas,
then the tree representation given in Figure would look as follows (in fact, we have already
shown this sort of representation for linguistic trees).
(S, (NP, (N, users)), (AUX, should), (VP, (V, clean), (NP, (DET, the), (N, printer))))
Lists are one of the datastructures that can be represented and manipulated very easily within a
computer.

● Parsing
● Generation

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node26.html [20/09/2000 18:02:55]

Parsing

Next: Generation Up: Processing Previous: Processing

Parsing

The task of an automatic parser is to take a formal grammar and a sentence and apply the grammar to the sentence in
order to (a) check that it is indeed grammatical and (b) given that it is grammatical, show how the words are
combined into phrases and how the phrases are put together to form larger phrases (including sentences). So, for
example, a parser would use the rules we gave above to check that the sentence The temperature has affected the
printer consists of a noun phrase, consisting of the noun Temperature followed by an auxiliary verb, followed by a
verb phrase, and that the verb phrase affected the printer consists of the verb affect and a noun phrase consisting of
the noun printers. In effect, this gives the same information as the sorts of tree structure we have given above, for
example in Figure . Thus, one can think of a parser as taking sentences, and producing such representations
(assuming the sentences are in fact well-formed according to the grammar).
How can this be done? There are many ways to apply the rules to the input to produce an output tree -- many different
procedures, or parsing algorithms by which an input string can be assigned a structure. Here is one method:
1. For each word in the sentence, find a rule whose right hand side matches it. This means that every word would
then be labelled with its part of speech (shown on the left hand side of the rule that matched it). This step is
exactly equivalent to looking up the words in an English dictionary. Given rules of the type N user, N
printer, and V clean, this will produce a partial structure as we can see at the top left corner
(Stage 0) of Figure .

Figure: Parsing Using a Bottom-Up Algorithm

2. Starting from the left hand end of the sentence, find every rule whose right-hand side will match one or more of
the parts of speech (Stage 1 of Figure ).
3. Keep on doing step 2, matching larger and larger bits of phrase structure until no more rules can be applied. (In
our example, this will be when the sentence rule finally matches up with a noun phrase and a verb phrase which
have already been identified). The sentence is now parsed (Stage 2-4 of Figure ).
It is generally possible to find more than one algorithm to produce a given result. As already mentioned, this is
certainly true of parsing: the algorithm given here is just one of many possible variants which differ in their ability to
cope efficiently with different types of grammar. The one we gave started out with the words of the sentence, and
built the tree `bottom up'. However, we could also have used an algorithm that built the tree `top-down' , starting
with the S node. Essentially, what this algorithm would do is guess that it is looking at a sentence, and then guess that
the sentence starts with a noun phrase, and then guess that the noun phrase consists of a noun, and then check to see
whether there really is a noun at the start of the sentence. Each time there is a choice of possibilities (maybe the noun
phrase starts with a determiner) it makes the first choice and, if that proves incorrect, backs up and tries the next
alternative. During the course of parsing a sentence with a complicated grammar it would eventually get the right
answer -- perhaps only after many wrong guesses. (The algorithms that MT and other NLP systems use are more
sophisticated and efficient than this, of course). The first few stages in a top-down parse are illustrated in Figure .

http://privatewww.essex.ac.uk/~doug/book/node27.html (1 de 3) [20/09/2000 18:02:58]

Parsing

Figure: Parsing Using a Top-Down Algorithm

This description applies only to building the surface, constituent structure tree , of course. As regards other levels of
representation (representations of grammatical relations , and semantic representations) , there are two basic
approaches, as we noted above. If information about other levels of representation is represented as annotations on the
constituent structure rules, then it should be possible to construct these other representations at the same time as the
constituent structure representation. This is slightly harder if the relationships between levels is stated in a separate
collection of rules. In this case, the natural thing to do is to first build the constituent structure representation, and
apply these rules to that representation.
The simplest procedure for this operates `recursively' down the surface (constituent) structure tree , dealing with each
node in turn. Beginning at the root node, the algorithm looks for a rule whose lhs matches this node, and its daughters.
In the case of the following rule (which we gave above, but repeat here for convenience), this means the root node
must be labelled with an S, and there must be three daughters, labelled NP, AUX, and VP, and the VP must in turn
contain a daughter labelled V, and a daughter labelled NP.

http://privatewww.essex.ac.uk/~doug/book/node27.html (2 de 3) [20/09/2000 18:02:58]

Parsing

NP:$1, AUX:$2, V:$3, NP:$4

HEAD:$3, SUBJECT:$1, OBJECT:$4

One interpretation of such a rule leaves the constituent structure tree untouched, and creates a new structure
representing the grammatical relations . This requires the algorithm to create a temporary structure corresponding to
the rhs of the rule. This will be labelled S, and will contain three daughters, one the HEAD, one the SUBJECT, and
one the OBJECT. Of course, this structure cannot be complete yet, because it is not yet known what these daughters
should contain. However, the algorithm now deals with the daughter nodes of the surface structure tree in exactly the
same way as it dealt with the root node (hence the process is called recursive). That is, it tries to find rules to match
each of NP, AUX, V, and NP, and produce the corresponding structures. When it has done this, it will be able to fill
in the parts of the temporary structure it created originally, and a representation of the grammatical relations will
have been produced. This can be seen in Figure .
A similar procedure can be used to interpret the rules that relate grammatical relation structures to semantic
structures. There are a number of details and refinements which should really be described, such as how we ensure
that all possible grammatical relation structures are produced, what we do about nodes that are mentioned on the LHS
but not on the RHS, and so on. But these are refinements, and do not matter here, so long as this basic picture is
clear.

Next: Generation Up: Processing Previous: Processing

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node27.html (3 de 3) [20/09/2000 18:02:58]

Generation

Next: Summary Up: Processing Previous: Parsing

Generation
So far, we have described how to take an input string, and produce a representation. But,
obviously, for most applications, the reverse process is also necessary. Equally obviously, how
hard this is depends on where you start from. Generating a string from a constituent structure
representation like those above is almost trivial. At worst one needs to do something to the
words to get the correct form (e.g. to get clean, not cleans in The user should clean the printer
regularly). For the rest, it is simply a matter of `forgetting' what structure there is (and perhaps
the not-so-trivial matter of arranging punctuation).

Figure: Building a Representation of Grammatical Relations

Starting from a representation of grammatical relations , or a semantic representation is harder.
If the relations between syntactic , grammatical relation structures, and semantic structures are
described by means of explicit rules, then one approach is to use those rules in the same way as
we described for parsing , but `in reverse' --- that is with the part of the rule written after the
interpreted as the lhs. Things are not quite so straightforward when information about
grammatical relation s, and/or semantics is packed into the constituent structure rules.
One possibility is to have a completely separate set of procedures for producing sentences from
semantic or grammatical relation structures, without going through the constituent structure
stage (for example, one would need a rule that puts HEAD, SUBJECT, and OBJECT into the
normal word order for English, depending on whether the sentence was active or passive,
interrogative or declarative). This has attractions, in particular, it may be that one does not want
to be able to generate exactly the sentences one can parse (one may want one's parser to accept
stylistically rather bad sentences, which one would not want to produce, for example).
However, the disadvantage is that one will end up describing again most, if not all, of the
knowledge that is contained in the grammar which is used for parsing .
A naive (and utterly impractical) approach would be to simply apply constituent structure rules
at random, until a structure was produced that matched the grammatical relation structure that is
input to generation . A useful variation of this is to start with the whole input structure, and take
all the rules for the category S (assuming one expects the structure to represent a sentence), and
to compare the grammatical relation structure each of these rules produces with the input
structure. If the structure produced by a particular rule matches the input structure, then build a

http://privatewww.essex.ac.uk/~doug/book/node28.html (1 de 2) [20/09/2000 18:03:00]

Generation

partial tree with this rule, and mark each of these parts as belonging to that tree. For example,
given the rule for S above, one could take the grammatical relation structure of a sentence like
The user has cleaned the printer and begin to make a phrase structure tree , as is illustrated in
Figure .

Figure: Generation from a Grammatical Relation Structure 1

One can see that a partial constituent structure tree has been created, whose nodes are linked to
parts of the grammatical relation structure (a convention is assumed here whereby everything
not explicitly mentioned in the rule is associated with the HEAD element). Now all that is
necessary is to do the same thing to all the parts of the Grammatical relation structure, attaching
the partial trees that have been constructed in the appropriate places. This is illustrated in Figure
. Again, there are many refinements and details missed out here, but again, all that matters is
the basic picture.

Figure: Generation from a Grammatical Relation Structure 2

Next: Summary Up: Processing Previous: Parsing

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node28.html (2 de 2) [20/09/2000 18:03:00]

Summary

Next: Further Reading Up: Representation and Processing Previous: Generation

Summary
This chapter has introduced the different kinds of knowledge needed to do translation, namely
grammatical or syntactic knowledge , semantic, pragmatic and real world knowledge .
Focussing on syntactic and semantic knowledge, we then looked at how this knowledge can be
represented and described. Finally, again concentrating on syntax and semantics , we looked
briefly at how this knowledge can be used for processing by means of parsing and generation
algorithms.

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node29.html [20/09/2000 18:03:00]

Further Reading

Next: Machine Translation Engines Up: Representation and Processing Previous: Summary

Further Reading
A somewhat more detailed discussion of many of the issues touched on in this Chapter can be
found in [Hutchins and Somers1992], especially Chapters 1, 3, 5, and 7.

The issue of how linguistic knowledge should be represented and described is one of the key
concerns of Linguistic theory, and will be covered by most introductory books on Linguistics.
On syntax , [Brown and Miller1991] is an accessible introduction. An elementary introduction
to linguistic semantics can be found in [Hurford and Heasley1983], a somewhat more
advanced introduction can be found in [Kempson1977] .

It is by no means the case that linguists agree on the sorts of representation that are required,
though the use of some kind of constituent structure is almost universal. In particular, there is
disagreement about how one should think about more abstract levels of representation. Here
[Borsley1991] provides a useful comparative discussion at an introductory level. Discussion of
the special requirements that MT makes of linguistic representation and description can be
found in [Van Eynde1993b] .

The issue of how linguistic representations and descriptions can be used for processing is the
topic of the fields of Computational Linguistics and Natural Language Processing (NLP). Here
[Allen1987,Grishman1986,Gazdar and Mellish1989] and [Winograd1983] provide excellent
introductions, though all go well beyond what is required for a basic understanding. Parts of
[Charniak and Wilks1976] are more elementary, though now somewhat out of date.

Much work in NLP focusses on analysis rather than synthesis or generation . For an
introduction to issues in generation , see [McDonald1987].

NLP is also a key area of interest in the field of Artificial Intelligence (AI), and many
introductions to AI contain some useful introductory material on NLP, examples are
[Rich1983,Charniak and McDermott1985,Tennant1981,Barr and Fiegenbaum1981]. Many of
the entries in [Shapiro1987] will also be useful.

Next: Machine Translation Engines Up: Representation and Processing Previous: Summary

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node30.html [20/09/2000 18:03:01]

Machine Translation Engines

Next: Introduction Up: Machine Translation Previous: Further Reading

Machine Translation Engines

● Introduction
● Transformer Architectures
● Linguistic Knowledge Architectures
❍ Comparative Grammar and Transfer
❍ Interlinguas
❍ LK Engines Summarised
● Summary
● Further Reading

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node31.html [20/09/2000 18:03:02]

Introduction

Next: Transformer Architectures Up: Machine Translation Engines Previous: Machine

Translation Engines

Introduction
In Chapter , we gave an overview of the environment in which a typical MT system might
operate, and outlined the various processes and parts involved. In Chapter , we discussed
how basic linguistic knowledge can be represented and used for automatic analysis and
synthesis. It is now time to look inside the most important non-human component in MT --- the
component that actually performs automatic translation --- what we will call the translation
engine.
MT engines can be classified by their architecture --- the overall processing organisation, or the
abstract arrangement of its various processing modules. Traditionally, MT has been based on
direct or transformer architecture engines, and this is still the architecture found in many of
the more well-established commercial MT systems. We shall therefore look at this architecture
in detail in Section before moving on to consider the newer indirect or linguistic
knowledge architectures which, having dominated MT research for several years, are starting
to become available in commercial form (Section ).

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node32.html [20/09/2000 18:03:02]

Transformer Architectures

Next: Linguistic Knowledge Architectures Up: Machine Translation Engines Previous:

Introduction

Transformer Architectures

The main idea behind transformer engines is that input (source language) sentences can be
transformed into output (target language) sentences by carrying out the simplest possible parse,
replacing source words with their target language equivalents as specified in a bilingual
dictionary , and then roughly re-arranging their order to suit the rules of the target language.
The overall arrangement of such an Engine is shown in Figure .
The first stage of processing involves the parser , which does some preliminary analysis of the
source sentence. The result need not be a complete representation of the kind described in
Chapter , but might just be a list of words with their parts of speech. This is passed to a
package of rules which transform the sentence into a target sentence, using --- where necessary
--- information provided by the parsing process. The transformation rules include bilingual
dictionary rules and various rules to re-order words. They may also include rules to change the
form of target words, for example, to make sure verbs have the correct person, number, and
tense suffixes.

http://privatewww.essex.ac.uk/~doug/book/node33.html (1 de 6) [20/09/2000 18:03:05]

Transformer Architectures

Figure: A Transformer Architecture (German to English)

To get a more detailed idea of how it works, we shall examine the steps in the translation of a
sentence taken from the printer manual text in Chapter :
● Sie den Knopf eine Position zurück. `Turn you the button one position back.' (Turn
the button back one position.)
Step 1: The German words are looked up in a German electronic dictionary , and the
appropriate category (for example, noun, verb) is assigned. In this particular case the look-up is
easy: almost all the words in the sentence are present in their base form --- the form they
normally have as dictionary entries. The only exceptions to this are the determiners den and
eine, which are inflected forms of der and ein and have to be recognised as such. After all, an
electronic dictionary is likely to be similar to an ordinary paper dictionary in that regularly
inflected forms of verbs, nouns, adjectives and determiners are not given since they can be
deduced from general rules. This is why most MT systems make use of a morphological
component. This component contains specific rules that deal with the regularities of inflection.
Take for example a verb like drehen (`turn'), which has the 3rd person singular form dreht
(`turns'). This form is not shown in monolingual or bilingual paper dictionaries like Düden
because other verbs of the same general form have the same form for 3rd person singular. If the
input sentence contained dreht, the lookup system would first follow its general policy of
looking up directly. Assuming that fails, it would then refer to some built-in inflection rules to
see if they could be used to derive an infinitive or stem form. One rule might say (in effect) ``If
the word has t on the end, it might be a 3rd person singular verb. Try to confirm the hypothesis
by removing the t, adding infinitive/imperative en, then looking for the resultant drehen.'' A
detailed account of the type of rules that we can encounter in a morphological component is
described in Chapter .
Note that the generalizations of a morphological component can also help the system to deal
with words which are not in its dictionary in any form at all. In the past few years, German has
acquired the verbs faxen and mailen, which are derived from English to fax and to
(electronically) mail. Let us suppose they are not in the German dictionary. If mailt or faxt are
encountered in the input, our 3rd person singular rule could apply and, as a result of the verb
annotation on the RHS, it would `guess' that the input forms might be 3rd person singular
versions of the hypothesised verb mailen or faxen. Obviously this hypothesis cannot be
confirmed in the available dictionary, but it is certainly useful: the parser can now work on the
assumption that the unknown word is probably a verb --- this is much more helpful in the parse
process than having no idea at all what its category/part of speech might be.
One problem with which the system also has to deal is the fact that the two words drehen and
zurück together form the main verb of the sentence: zurückdrehen. The recognition may be
done by a rule which specifies that prepositions which stand alone (i.e. without a complement)
at the end of a sentence can form part of the main verb. This possibility is then checked in the
dictionary , which should contain an entry for the verb zurückdrehen.
Step 2: Some rules of a German grammar are used to try to parse the sentence. This parse
might result in the assumption that the NP den Knopf (`the button') is the object of
zurückdrehen and (possibly) that the next NP eine Position is a modifier of some sort. An
advanced parser might work out that it is in fact a measure modifier. However, it is quite
possible that the transformer Engine will not need any parse at all in this case (beyond
identification of the category of the words in the string). This is because the difference between

http://privatewww.essex.ac.uk/~doug/book/node33.html (2 de 6) [20/09/2000 18:03:05]

Transformer Architectures

the German and some possible English translations is not great.

Step 3: The Engine now applies some German to English transformation rules. The first step
here is to find translations of the German words in a German to English dictionary . Taking the
simple cases, der --- the nominative form of den --- goes to the, Knopf goes to button, ein to a,
Position to position. The rules might have the following form:
knopf{cat=n} button{cat=n} ein{cat=det} a{cat=det}
... ...
and so on. That is, when knopf is a noun ( cat=n) it is translated as button. Similarly, ein
translates as the determiner a --- in the present context, ein would be best translated as one, but
let us assume that it is routinely translated as a by the Engine.
Turning to zurückdrehen, there needs to be a rule which says ``If there is an imperative verb X,
followed by the NP Sie, the translation is the translation of X. In this case, we have an
imperative verb ( zurückdrehen) followed by the NP Sie, so we will get turn back as the
translation. This rule is intended to prevent the translation of the German NP Sie which
functions as the subject. English imperatives do not have an overt subject and therefore the
literal translation Turn back you the button one position is unacceptable. Our proposed rule
would give Turn back the button a position, which is better .
In practice, the imperative translation might be handled by a pair of rules. The first could look
like this:
X{cat=v,mood=imperative} Sie
X
The LHS matches cases where there is any imperative verb X followed by Sie. The RHS says
that the translation of such a structure simply consists of the translation of the imperative verb.
As we have stated it, this first rule has not done any translation. What it has done is to re-order
part of the German sentence prior to translation into English. The Engine can now simply
apply the lexical translation rules to the re-ordered sentence:
zurückdrehen turn_back
After applying all these rules, the Engine now has an internal representation of the form Turn
back the button a position.
Step 4: The Engine would now apply rules which turn the stem or dictionary forms of English
words to their inflected forms . As it happens, in the present example, the English stem forms
happen to be exactly what is wanted. For example, the stem form turn which the dictionary
supplied is identical to imperative turn. Moreover, all the nouns are singular, so it is
unnecessary to add any plural affixes (e.g. s or es).
This discussion is rather sketchy and we have ignored many details. For example, we have said
very little about how the various types of transformation rule should be ordered: how should
re-ordering rules be interleaved with the bilingual dictionary rules ? We have also not said
anything much here about how the system copes with ambiguities, or how rules are prevented
from applying in the wrong circumstances; for example, it will not always be the case that a
preposition at the end of a German clause `belongs' to an earlier imperative verb. However, this
should have given the reader an impression of what is involved in a transformer architecture.

http://privatewww.essex.ac.uk/~doug/book/node33.html (3 de 6) [20/09/2000 18:03:05]

Transformer Architectures

We can now summarize some of the distinctive design features of this sort of engine:
● Input sentences are automatically parsed only so far as it is necessary for the successful
operation of the various lexical (word-based) and phrasal transformation rules. The
transformer engine is often content to find out just a few incomplete pieces of
information about the structure of some of the phrases in a sentence, and where the main
verb might be, rather than worrying about getting a full and complete parse for the whole
thing. In other words, parsing may stop before an S rule of the kind described in Chapter
has been applied.
In practice, transformer systems tend not to have particularly large grammar s for the
language they translate from. Thus in the German to English transformer system
discussed above, we assumed that the grammar covered only some features of German.
As a consequence it would not be able to decide for many (or perhaps any) input
sentences whether it is grammatically acceptable.
● The use of limited grammars and incomplete parsing means that transformer systems do
not generally construct elaborate representations of input sentences --- in many cases, not
even the simplest surface constituent structure tree. As we will see, other types of MT
system construct much more abstract and deep representations.
● Most of the engine's translational competence lies in the rules which transform bits of
input sentence into bits of output sentence, including the bilingual dictionary rules. In a
sense a transformer system has some knowledge of the comparative grammar of the
two languages --- of what makes the one structurally different from the other.
● Inflection rules aside, transformers generally have no independent linguistic knowledge
of the target language because they have no independent grammar for that language. In
the German-English system, there would be few, if any, independently stated rules about
English --- although you could perhaps infer some aspects of English grammar from the
rules which transform bits of German into bits of `English'.
Given these general features, we can describe the translational behaviour that can be expected
from a system with a transformer engine.
Characteristic to the performance of such a system is the fact that the engine will not be
particularly troubled when faced with unusual, marginally acceptable or frankly unacceptable
source language sentences; it will rarely have sufficient source language grammatical
knowledge to recognise something as ungrammatical. If the grammatical structures in the input
sentence are not recognised by some transforming rule, that structure will pass through to the
output sentence without any re-arrangement. We have seen this in the example above, where all
the word order and structure of Drehen Sie den Knopf eine Position zurück apart from the
relationship between drehen and zurück was passed through into the English output. Something
similar is true for the words in the input sentence: if they are not found in the system's
dictionary then they are passed through into the English output and remain untranslated. As a
consequence of these features this type of architecture implies that, in the worst case, the whole
input sentence could survive unchanged as the output sentence. This would happen in the
highly unlikely case that none of the input words are found in the bilingual dictionary and none
of the input sentence grammatical structure is recognised.
With regard to the target language performance of the system we can say that since the system
has no detailed knowledge of target language grammar there is no guarantee that the
transformed input sentence is actually a grammatical sentence in the target language. Although

http://privatewww.essex.ac.uk/~doug/book/node33.html (4 de 6) [20/09/2000 18:03:05]

Transformer Architectures

in most cases output will resemble the target language (especially the use of target language
words), the result can sometimes be a completely unintelligible `word salad'. In such cases one
could say that the output does not belong to any known language --- natural or artificial.
The typical design features of a transformer system pose some restrictions on the development
of additional language modules. First, the engine will run in one direction only, for example,
from German to English. If the engine developer wants it to go in the other direction she more
or less has to completely rewrite the transformer rules. Since the transformer rules include
bilingual dictionary rules , this can mean that the Engine has to be supplied with two bilingual
dictionaries, for example, German-English and English-German. This is rather clumsy since,
apart from the differences in their directionality, the dictionaries contain much the same
information. Secondly, the engine links a single pair of languages only. If the developer wants it
to translate into another target language then again she more or less has to completely re-write
the transformer rules. Again, this amounts to rewriting most of the system. Grammatical
knowledge of English and of German which is built into a German-English system cannot then
be transferred to a English-French or a German-French system. Even in cases where a system
contains only a rather limited grammatical knowledge of the languages it involves reproducing
this knowledge for the development of other language pairs means an unnecessary time loss.
Drawing these various points together, we can summarise the situation of the transformer
engine architecture as follows:
● It is highly robust. That is, the Engine does not break down or stop in an `error
condition' when it encounters input which contains unknown words or unknown
grammatical constructions. Robustness is clearly important for general-purpose MT.
● In the worst case it can work rather badly, being prone to produce output that is simply
unacceptable in the target language (`word salad').
● The translation process involves many different rules interacting in many different ways.
This makes transformer systems rather hard to understand in practice --- which means
that they can be hard to extend or modify.
● The transformer approach is really designed with translation in one direction , between
one pair of languages in mind, it is not conducive to the development of genuinely
multi-lingual systems (as opposed to mere collections of independent one-pair,
one-direction engines).
To close this section, we give an example of a German Teletext Travel News broadcast and a
translation produced by an actual small transformer Engine (which is available commercially,
and rather cheaply for use on PCs). The source text and the raw (unedited) MT output are given
on page . The Engine is clearly struggling here with unfamiliar words and structures,
occasionally producing completely unintelligible output which would be unsuitable even for
gisting. This example represents the `bottom end' of transformer performance, but gives a good
idea of how useful even this quality of translation can be --- readers with no knowledge of
German will certainly get more information from the translation than they could from the
original. Note, however, that the quality of the output could be improved considerably if the
system were adapted to dealing with this particular text type and vocabulary. As we mentioned
in Chapter , tuning the system to a particular text type is worthwhile if the input consists of
many texts of that type.

http://privatewww.essex.ac.uk/~doug/book/node33.html (5 de 6) [20/09/2000 18:03:05]

Transformer Architectures

Next: Linguistic Knowledge Architectures Up: Machine Translation Engines Previous:

Introduction

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node33.html (6 de 6) [20/09/2000 18:03:05]

Linguistic Knowledge Architectures

Next: Comparative Grammar and Up: Machine Translation Engines Previous: Transformer
Architectures

Linguistic Knowledge Architectures

The second major architecture --- indirect or linguistic knowledge (LK) architecture --- has
dominated research in MT design during the past decade and is starting to appear in a number
of commercial systems. The idea behind LK engines is straightforward enough:
High quality MT requires linguistic knowledge of both the source and the target
languages as well as the differences between them.
We use the term `linguistic knowledge' to refer to extensive formal grammar s which permit
abstract/relatively deep analyses in the sense of Chapter . We shall see later on just how deep
the analysis can go.
With the Transformer architecture, the translation process relies on some knowledge of the
source language and some knowledge about how to transform partly analysed source sentences
into strings that look like target language sentences. With the LK architecture , on the other
hand, translation relies on extensive knowledge of both the source and the target languages and
of the relationships between analysed sentences in both languages. In short, LK architecture
typically accords the target language the same status as the source language. As can be seen
from Figure , the LK architecture requires two things:
● A substantial grammar of both the source language and the target language. These
grammars are used by parsers to analyse sentences in each language into representations
which show their underlying structure, and by generators to produce output sentences
from such representations.
● An additional comparative grammar which is used to relate every source sentence
representation to some corresponding target language representation --- a representation
which will form the basis for generating a target language translation.
The LK engine will have grammar s for each language it deals with: in a German-English
system, there would be one for German and one for English. Each of these grammars is an
independent entity, i.e. there will be a set of rules which is identifiably for German, and
another, separate set which is identifiably for English. In fact the physical and conceptual
separation between the two grammars is such that in the initial stages of developing an LK
engine, a group of English specialists might write the grammar for English entirely
independently of another group of German specialists who are writing the system's German
grammar. In such case both groups would have to aim though at a similar deep representation
of their language, otherwise structural discrepancies can be created that would require extra
transfer rules for mapping these different structures onto each other.

http://privatewww.essex.ac.uk/~doug/book/node34.html (1 de 2) [20/09/2000 18:03:06]

Linguistic Knowledge Architectures

Figure: The Components of a Transfer System

Looking at Figure , it is clear that if (say) the system is translating from German to
English, the first (analysis) step involves using the parser and the German grammar to analyse
the German input. The second ( transfer) step involves changing the underlying representation
of the German sentence into an underlying representation of an English sentence. The third
(synthesis) step and final major step involves changing the underlying English representation
into an English sentence, using a generator and the English grammar. The fact that a proper
English grammar is being used means that the output of the system --- the English sentences ---
are far more likely to be grammatically correct than those of a German-English Transformer
system (recall that the latter had no explicit English grammar to guide it). In fact, if ( per
impossibile) we had an LK German-English system with a `perfect' English grammar the only
sort of mistake it could make in the output would be errors in translational accuracy . That is, it
would always produce perfectly well-formed English sentences even when it did not produce
correct translations.
This also means that the whole Engine should be reversible , at least in theory. Taking the
German-English LK engine in Figure , we could run the translation from right to left. That
is, we could give it English sentences, which would then be analysed into underlying
representations. These representations would be changed into German underlying
representations and a German translation would then be synthesised from the result. The same
grammar s for each language are used regardless of the direction of the translation. In practice
few translation engines are reversible, since some rules that are necessary for correct translation
in one direction could cause problems if the process was reversed . This is especially true for
lexical transfer rules , as we will see later on in this chapter.
With this general picture in mind, the next subsection focusses on the so-called transfer
component, which embodies the comparative grammar that links the analysis and synthesis
components together --- the module in the centre of Figure .

● Comparative Grammar and Transfer

● Interlinguas
● LK Engines Summarised

Next: Comparative Grammar and Up: Machine Translation Engines Previous: Transformer
Architectures

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node34.html (2 de 2) [20/09/2000 18:03:06]

Comparative Grammar and Transfer

Next: Interlinguas Up: Linguistic Knowledge Architectures Previous: Linguistic Knowledge Architectures

Comparative Grammar and Transfer

We have said that parser s in LK engines typically analyse to relatively abstract, or deep underlying representations.
Of course individual systems differ radically in the precise sorts of representations they use, but suppose the Engine
uses the English grammar to produce the sort of deep syntactic representation we described in Chapter (this is far
from being the most abstract representation one can imagine, of course). If we are translating sentence ( ) into
German, the analysis component might produce a representation along the lines of Figure
● temperature has affected the print density.

We can look at how the comparative grammar relates such a representation to corresponding representations for
target language sentences. Just as each monolingual grammar has a `dictionary' of rules (e.g. N temperature)
so also the comparative grammar has bilingual dictionary rules. In the simplest case, these may just relate source
lexical items (`words') to target lexical items:
temperature temperatur print_density druckdichte affect beeinflu
en

Figure: Abstract Tree Representation

One difference between these bilingual dictionary rules and those shown for the Transformer engine is that the latter
were intended to be used in one direction only. The in the present rules indicates that they can (in principle) serve
as English-German or German-English rules.
These dictionary rules can be seen as relating leaves (the word nodes) on the source language tree to leaves on the
target language tree. The comparative grammar also contains some structural rules which relate other parts and nodes
of the two trees to each other.
One such structural rule might be read as follows: ``The translation of the whole sentence is normally made up of the
translation of the verb + the translation of the subject + the translation of the object.'' Note that `translation' in this
context has the restricted sense of translation into the corresponding target language representation --- this
representation has to be input to synthesis before a `full' translation is reached. The structural rule we need might be
written in the following way (where the LHS describes an English structure and the RHS describes the German, and
$H, $S, and $O are variables interpreted as standing for pieces of English structure on one side, and for their
translations on the other side).
HEAD:$HEAD, D-SUBJ:$SUBJECT, D-OBJ:$OBJECT

http://privatewww.essex.ac.uk/~doug/book/node35.html (1 de 6) [20/09/2000 18:03:11]

Comparative Grammar and Transfer

HEAD:$H,

D-SUBJ:$S, D-OBJ:$O
The left and right hand sides of the rule reflect the `canonical' order (HEAD, then DEEP SUBJECT, then DEEP
OBJECT) that one finds in the source (and target) representations. In some systems, the rule application procedure
might be set up so that rule would work regardless of the left-right order of the nodes in the source representation.
This rule says that in the translation of the sentence as a whole, the HEAD is whatever the HEAD in the source
language translates as. The HEAD is the verb affect, and its translation is given by a bilingual dictionary rule. The
DEEP SUBJECT and DEEP OBJECT just contain single content words ( temperature and print_density) and so they
too are translated by the appropriate dictionary rules.
The annotations on the nodes of the representations must also be translated in some way. The rules relevant to our
example are straightforward, indicating that the given values are simply carried over from source structure to target
structure:
{def=+} {def=+} {aspect=perfective} {aspect=perfective}
Of course, one could imagine that this `copying' of information without changes could occur by default, i.e. features
are copied unless a rule explicitly says otherwise (although specifying how this sort of system should actually work
turns out to be surprisingly difficult).

Applying these rules to the English representation in Figure will result in the construction of the corresponding
German representation in Figure .

http://privatewww.essex.ac.uk/~doug/book/node35.html (2 de 6) [20/09/2000 18:03:11]

Comparative Grammar and Transfer

Figure: Tree Representation after Translation

This representation serves as input for the German synthesis module, which applies the rules of the German
grammar to produce a German sentence. These rules will include one or more which require that the past participle
of a verb is realised at the end of the clause when there is an auxiliary ( hat, in this example). Thus, ( ) should be
produced as the translation.
● Temperatur hat die Druckdichte beeinfluß t

It should be clear that LK and Transformer architectures handle the word order problem rather differently. A
Transformer engine generally preserves the surface order of the source language and directly re-uses it --- with
modifications where appropriate --- to order the target language words. An LK engine , on the other hand, extracts all
the information it can from the source word order and recodes this information in a more or less abstract
representation. The generator for the target language will use the information in the representation and in the target
language grammar to construct a target language sentence with a word order that it is grammatically appropriate for
that language. In short, ordering information is not normally carried over directly.
The only differences between the English and the German representation in this example is in the words on the leaf
nodes; the geometry and annotations on the tree are the same. Ideally, this similarity will hold for most sentences, so
that most of the work in constructing the representation is done by the dictionary rules. However, it is important to
realise that the design of the comparative grammar anticipates the possibility that the structures could be very
different indeed if the differences between the source and its target language translation are very great. We will look
at some such examples in the following chapters (cf. especially Chapter ).
The similarity of the representations is related to the simplicity of the rules. For example, according to the rule, DEEP
SUBJECTS translate as DEEP SUBJECTS, and DEEP OBJECTS as DEEP OBJECTS, and the rules for translating
the words are stated without any conditions. But in general, one would only want to say that subjects and objects are
normally translated as subjects and objects, and it is easy to think of cases where one would want to put extra
conditions on such lexical rules. For example, English import translates as French importer when it is a verb, and
importation when it is a noun, and the verb effect translates réaliser or effet, depending on whether it is a noun or a
verb. Such examples can be multiplied at will. Similarly, one cannot always simply preserve the values of features
such as det, or aspect. For example, in translating from English to French , one cannot generally expect to
preserve the values of attributes indicating tense and aspect, if these are direct encodings of surface word forms (cf.
Chapter ).
A relatively straightforward example where a more complex rule is called for involves the translation of the English

http://privatewww.essex.ac.uk/~doug/book/node35.html (3 de 6) [20/09/2000 18:03:11]

Comparative Grammar and Transfer

verb like into French plaire, as in ( ), which shows the `switching' of arguments.
● ❍ Sam likes the new laser printer.

❍ La nouvelle imprimante à laser plaît à Sam.

Such a rule might look as follows:

HEAD:like, SUBJ:$1, OBJ:$2

HEAD:plaire, SUBJ:$2, OBJ:$1

Switching of arguments occurs because the variables $1, and $2 are associated with different grammatical relations
on the two sides of the rule ( $1 will be bound to the representation of Sam, and $2 will be bound to the
representation of the new laser printer (on the English side of the rule), and la nouvelle imprimante à laser (on the
French side of the rule)). The identity of the words that fill the HEAD relation has been given to prevent this rule
applying to examples involving `normal' verbs (one will also have to make sure that the `normal' rules do not apply in
translating like and plaire, of course). This process of argument switching is illustrated in Figure .
Special rules like the one given above have to be written for every case where there is some difference between the
output of the source language analysis and the input expected by the target language generator. In practice, one
would expect the contrastive grammar for an English-French , or English-German MT system whose most abstract
representations involve surface grammatical relations to be quite large.
In general, the size and complexity of a comparative grammar can be reduced by increasing the depth of the parsing
towards more abstract levels of representation. For example, the use of Semantic Relations (see Chapter ) would
remove the need for a special like-plaire rule, because both English and French sentences in ( ) would have
representations with Sam as EXPERIENCER, and the new laser printer/la nouvelle imprimante à laser as THEME.

http://privatewww.essex.ac.uk/~doug/book/node35.html (4 de 6) [20/09/2000 18:03:11]

Comparative Grammar and Transfer

Figure: Complex Transfer

The discussion so far may give the impression that there is a single transfer approach to MT. But this is far from
being the case. For one thing, different systems use different styles, and levels of representation. For another thing, we
have only given one view of the relation of the various components. That other views are possible is indicated below,
where we discuss some variable aspects of transfer systems.
Intermediate representations in transfer
As we have described transfer, the mapping between source and target structure is direct in the sense that
there are no intermediate structures. There are, for example, no structures which have target words, and source
geometry. Some systems, however, make a distinction between lexical transfer (which simply changes source
words to target words) and structural transfer (where rules actually change the shape of the tree ) with one set
of rules being applied before the other. Also, the rules we have given each deal with a structure in one step,
without using an intermediate representation. But it is possible to have a transfer rule which changes the
source tree in some way, producing an intermediate representation, that must have another rule applied to it
before a genuine target structure results. The problem with systems that allow this is that problems of complex
rule interaction can occur, in the way that they do with a transformer architecture. We have allowed for a
limited degree of collaboration between rules that deal with structure, and rules that deal with features , for
example. The advantage of this is that we do not have to state facts about the relation between, for example,
determination values in each rule. This seems both natural and economical in terms of effort involved. The
disadvantage of this is that it increases the number of rules that must be applied in order to translate each tree .
An alternative is to state the rules separately like this, but in some way compiling them together, to produce
rules that deal with entire subtree s. The problem with this is that the set of compiled rules tends to be very
large.
Symmetry
Throughout this chapter the picture of transfer that we have described is rather symmetric. That is, it assumes
the target structure is rather similar to the source structure in the sense of being of corresponding depth of
analysis or linguistic abstraction. This suggests analysis and synthesis are to a large extent `inverses' of each
other. But this is not a requirement. It is possible to imagine systems where the input to transfer was a deep
syntactic representation, and the output was a representation of surface syntactic structure. Moreover, in a
one-direction al system for one pair of languages, no real distinction might be drawn between transfer and
synthesis . Symmetry is however desirable as soon as one deals with more than one language or direction. In
such cases the advantages become obvious, having a separate synthesis component with a role broadly the
inverse of to that of analysis --- not only can the same synthesis component be used for all transfer pairs, but
one will avoid duplicating work by using the same (or similar) grammar s in analysis and synthesis .
Reversibility
We noted that transfer rules could be reversible in principle, and though this is natural, and attractive
(because it halves the number of transfer components one has to construct and makes testing easier, since, if a
rule works in one direction it should work in the other), it is not obvious that reversible transfer rules are
always possible, or desirable. This is because a system should be able to translate a wide variety of input
strings, some of them the type of string that one would normally not want to produce as output. As a simple
lexical example of the reversibility problem consider the slightly old-fashioned Dutch word aanvangen. One
would like to be able to translate this into English as begin, but one would normally not want to translate begin
into aanvangen. One would choose the more common verb beginnen instead. So the following translation rule
cannot be reversible :
aanvangen begin
Well-formedness
In order for transfer output to be useful for synthesis it is desirable that it is in some sense well-formed for the
target language. To produce well-formed target language structures transfer components can become rather
complex. Some systems supplement normal transfer with a set of adjustment rules which transform the output
of transfer to make it more suitable for input to the target synthesis .
Instructions for synthesis
The target structure that is produced by transfer has been described as a simple linguistic tree --- it does not

http://privatewww.essex.ac.uk/~doug/book/node35.html (5 de 6) [20/09/2000 18:03:11]

Comparative Grammar and Transfer

contain, for example, special instructions to guide synthesis . Some systems do contain this sort of information:
transfer attaches what are essentially small programs to nodes of the target tree , which are executed in
synthesis.
Choosing between possible translations
In general, several different transfer rules will be able to apply to a structure , giving alternative (not
necessarily correct) translations. The question arises as to how to choose between these. One crude possibility
is to organize the rules so they apply in sequence, taking the results of the first rule that produces a `correct'
target structure (correct in the sense of getting an acceptable target sentence, perhaps). Alternatively, one could
apply all these rules and find some way of scoring the results, so as to prefer the better ones. A complementary
question which arises in the case where no translation rule applies (because none matches the source structure)
is whether one should leave the structure untranslated (it may be, for example, a proper name), or to try to force
a rule to apply?
Declarative or procedural processing
If the answer to the problem above is to organize the rules so they apply in sequence then the result is the
contamination of declarative information in the comparative grammar with procedural information --
information about the order in which things should be done. This violates a widely accepted principle that it
should be possible to describe the relevant linguistic facts in an MT system independently of the ways the
engine actually uses them. The advantages of a declarative system are (a) ease of understanding, modification
and debugging, and (b) independence of particular implementations or algorithms: if a collection of rules is
declarative, it will be possible to consider alternative algorithms for applying them, with some confidence that
the same results will be produced, which allows one to find the most efficient way of processing. Despite these
advantages of declarativity there is a strong temptation to introduce non-declarative characteristics (e.g. to
ensure that the most likely transfer rules are tried early, and block the application of other rules, so cutting
down the space of possibilities that have to be processed). Thus, though declarativity is a generally accepted
goal, it is a property that systems have in different degrees, and it is not even generally agreed what the correct
compromise between efficiency and declarativity is.

Next: Interlinguas Up: Linguistic Knowledge Architectures Previous: Linguistic Knowledge Architectures

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node35.html (6 de 6) [20/09/2000 18:03:11]

Interlinguas

Next: LK Engines Summarised Up: Linguistic Knowledge Architectures Previous:

Comparative Grammar and

Interlinguas

The general idea suggested by the discussion of the like-plaire example at the end of the
previous section is that comparative grammar (hence transfer) becomes simpler as linguistic
analysis goes deeper --- as the representations become more abstract. In fact, a major objective
of MT research is to define a level of analysis which is so deep that the comparative grammar
component disappears completely. Given such a level of representation, the output of analysis
could be the direct input to the target synthesis component. Representations at such a level
would have to capture whatever is common between sentences (and expressions of other
categories) and their translations --- that is they would have to be representations of `meaning'
(in some sense). Moreover, such a level of representation would have to be entirely language
independent --- for example, if it preserved features of the source language, one would still
require a transfer component of some kind to produce the corresponding features of the target
language. For this reason, such a level of representation is normally called an Interlingua , and
systems that use such a level are called Interlingual.

The relationship between transfer and interlingual systems can be pictured as in Figure . As
one can see, the size of the contrastive grammar (hence the transfer component) between two
languages decreases as the level of representation becomes more abstract. As this diagram
perhaps suggests, the difference between transfer representations and interlinguas is a matter
of degree rather than absolute distinction (for example, Chapter shows how one might
combine an interlingual representation of tense and aspect with a transfer approach to other
phenomena).
There are a number of clear attractions to an interlingual architecture. First, from a purely
intellectual or scientific point of view, the idea of an interlingua is interesting, and exciting.
Second, from a more practical point of view, an interlingual system promises to be much easier
to extend by adding new language pairs, than a transfer system (or a transformer system) . This
is because, providing the interlingua is properly designed, it should be possible to add a new
language to a system simply by adding analysis and synthesis components for it. Compare this
with a transfer system, where one needs not only analysis and synthesis , but also transfer
components into all the other languages involved in the system. Since there is one transfer for
each language pair, N languages require transfer components (one does not need a transfer
component from a language into itself). For example, extending a system for 3 languages into
one for 5 means writing 14 new transfer components (as one goes from 6 to 20 transfer
components) , and going from a 5 language system to a 9 language system means going from
20 components to 72.

http://privatewww.essex.ac.uk/~doug/book/node36.html (1 de 3) [20/09/2000 18:03:13]

Interlinguas

Figure: Transfer and Interlingua

Ideas about interlinguas are intimately tied up with ideas about the representation of meaning.
We will look at this in more detail in Chapter . However, one can get a flavour of the
problems that are involved in defining an interlingua by considering the following.
Producing an interlingual representation involves producing a representation that is entirely
language independent (for the languages one wants to translate, at least). This involves
producing a language independent representation of words, and the structures they appear in.
Under the latter heading, one would have to make sure one could represent the difference in
meaning between examples like those in ( ) --- assuming one does not want them all to
translate alike, that is --- and find a way of representing the meaning that is expressed by
various tenses, and by the distinction between definite, and indefinite NPs (e.g. a printer vs. the
printer).
● ❍ It was the printer that was serviced yesterday.

❍ It was yesterday that the printer was serviced.

❍ The printer was serviced yesterday.

While this raises many unsolved linguistic problems, it is the language independent
representation of word meaning that seems to pose the most difficult problems. The central
problem is how to choose the vocabulary of the interlingua --- what are the primitive concepts
of the meaning representation to be. Notice that this is not a question of what names we should
give the concepts --- how we should write them down or represent them. Of course, we should
make sure that we do not use one name for two concepts, which might be confusing, but
beyond this, we can give them, for example, names from an existing language (e.g. English, or
Esperanto ), or numbers, or codes in some invented language --- the only difference here will be
how easy they are to write or remember. The problem is one of identity. For example, are we to
include a concept that we might write as CORNER --- this being the interlingual representation
of the English noun corner? This seems natural enough from the point of view of English, but
from the point of view of, for example, Spanish it is not so natural, because in Spanish there
are different words for inside corners ( rincón) and outside corners ( esquina). Is there any
reason why we should not choose a more specific primitive word for our representation, for
example, OUTSIDE-CORNER and INSIDE-CORNER. Similar problems will arise wherever
one language has several words that correspond to one word in another. The point is that
different languages `carve the world up' differently, so settling the choice of vocabulary for the
interlingua will involve either (i) some apparently arbitrary decisions about which language's
conceptualization to take as basic, or (ii) `multiplying out' all the distinctions found in any
language. In the latter case one will have two interlingual items for English corner (because of
Spanish ), two for English river (because of the distinction between rivière and fleuve in
French ), and two for English eat, because of the distinction between essen (for humans) and
fressen (for animals) in German . When one consider more distant languages like Japanese,
even more distinctions will arise --- Japanese does not distinguish between wearing and putting
on, as does English, but does make a distinction according to where the item is worn or put on
(e.g. on the head vs on the hands). Of course, one solution to this multiplicity of concepts is to
try to reduce the set of primitive concepts, defining complex concepts in terms of the primitive
ones. For example, one might think that EAT is not a primitive, but that INGEST is, and that
the interlingual representation of the meaning of eat should involve INGEST, and some other
primitives. However, though this solves the problem of the number of concepts, it does not
overcome the problem of arbitrariness, and it raises the problem of finding an adequate set of

http://privatewww.essex.ac.uk/~doug/book/node36.html (2 de 3) [20/09/2000 18:03:13]

Interlinguas

primitives to capture the relevant distinctions (the reader might, as an exercise, like to consider
what a set of primitives would look like to distinguish a handful of verbs like eat, drink, gobble
up, feed on, or find a set of primitives that will distinguish between different kinds of furniture
(chairs, stools, tables, etc.)).

Figure: The Components of an Interlingual System

A further problem is that using an interlingua in MT can lead to extra, unnecessary work, in
some cases. For example, suppose one has an interlingua intended for translation between
English, French , and Japanese . Japanese distinguishes terms for older and younger brother and
sister, and for various relatives depending on whether they belong to the speaker, or to the
hearer (i.e. the term for my mother is different from that for your mother, or mothers in
general). The problem is that this distinction has to be encoded in the interlingua , so one must
decide if English brother is an older brother or a younger brother, even if one is not translating
into Japanese. For example, translating Sam's brother has already left into French will involve
dealing with an ambiguity, since there will be two interlingual representations differing as to
whether the brother is older or younger than Sam. But of course, this is irrelevant for both
English and French , and one can manage with a very simple transfer rule (along the lines of
brother frère).
These are problems for general vocabulary. One should note, however, that these problems do
not occur for all kinds of vocabulary. In particular, in domains where there is a codified system
of terminology , the conceptual organization is generally relatively clear. In such cases, the set
of concepts, and thus at least some of the vocabulary of the interlingua , is already settled.
Interlinguas are rather metaphysical things. Implicitly or explicitly, they say what the universe
is made of (events, processes, individuals, relations, etc.) and how it is put together. It is not at
all surprising that many aspects of interlinguas are in dispute and are likely to remain so for
some time to come. Given these difficulties, interlinguas in the sense described here are more
popular as a basis for theoretical research in MT rather than for full-scale commercial
development. For the next few years, most general purpose LK MT systems on the market are
unlikely to analyse any deeper than to the level of semantic relations --- and even that will be
considered impractically deep by many developers and vendors. Nonetheless, we can certainly
expect a tendency towards increasingly deep analysis over the next decade or so.

Next: LK Engines Summarised Up: Linguistic Knowledge Architectures Previous:

Comparative Grammar and

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node36.html (3 de 3) [20/09/2000 18:03:13]

LK Engines Summarised

Next: Summary Up: Linguistic Knowledge Architectures Previous: Interlinguas

LK Engines Summarised
Having looked at some of the components of an LK engine and having seen something of how
they might work, we can conclude this discussion of MT architectures by setting out what the
performance characteristics of an LK engine might be.
● Because the system has a (partial) grammar of the target language, output will tend to be
grammatical. At any rate, it will be far less strange and far less source-language
grammar- dependent than output from transformer engines.
● Because the comparative grammar completely specifies a relationship between
representations of two languages, translational quality will tend to be more reliable than
for transformer engines.
● Because the system tends to separate language into separate modules (one grammar for
each language and one comparative grammar for each pair of languages), it is relatively
easy in principle to add new languages to the system. For example, adding Dutch to a
German-English system would require only the addition of a Dutch grammar module
and Dutch-English and German-English comparative grammar modules. Individual
language modules can be designed and constructed without specifying which other
language modules they will have to work with in the final system. Of course, this matters
more to the developer than the user since it is the former that writes and supplies basic
language modules.
● The system will be upset by unusual, marginally acceptable or frankly unacceptable input
sentences because it has a grammar for the source language and hence a strong notion of
grammaticality.
● Because the grammar s that computational linguists are able to write are invariably less
complete than the `real' complete grammar of any language, there will be some
complicated grammatical input sentences that the system fails to recognise.
From the engine manufacturer's point of view, the transformer architecture has the advantage
that it accepts anything that is given to it (though the translations it produces are another
matter). The LK architecture is at a disadvantage here: because it thinks it knows something
about the languages involved, it tends to think that anything it doesn't know isn't language and
hence unacceptable. As a consequence, a pure LK engine during its development phase tends
to grind to a halt on anything unusual, or even on something quite common which the developer
has forgotten to include.
For commercial purposes, this means that pure LK engines must be supplemented with various
coping strategies. For example, if they cannot parse a particular sentence completely, then they
at least ought to be able to use some of the information on those parts of the sentence for which
they did find a parse --- and perhaps they can guess how those well-parsed bits might be fitted
together.
LK systems are clearly superior in principle to transformers. However, MT systems require a

http://privatewww.essex.ac.uk/~doug/book/node37.html (1 de 2) [20/09/2000 18:03:14]

LK Engines Summarised

considerable development effort and some commercial transformer systems which have
undergone extensive revision, refinement and updating over the years can achieve a good
overall performance. Furthermore, some MT systems have sufficient flexibility in the design of
the engine to allow developers to increase the depth and sophistication of their linguistic
knowledge and even the overall arrangement of grammar s. We can therefore expect highly
developed transformer MT systems to survive in some sectors of the marketplace for some
years to come.

Next: Summary Up: Linguistic Knowledge Architectures Previous: Interlinguas

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node37.html (2 de 2) [20/09/2000 18:03:14]

Summary

Next: Further Reading Up: Machine Translation Engines Previous: LK Engines Summarised

Summary
In this chapter we have looked inside two different kinds of MT system, transformer systems ,
and linguistic knowledge systems, discussing, under the latter heading the distinction between
transfer and interlingual systems. The following chapters will amplify this picture in various
ways, by looking in more detail at the sorts of knowledge that are involved, for example, in
dictionaries , and the representation of `meaning', and looking at some particular translation
problems. In Chapter we will give some more discussion of the limitations of LK
approaches , and describe a recently developed alternative.

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node38.html [20/09/2000 18:03:15]

Further Reading

Next: Dictionaries Up: Machine Translation Engines Previous: Summary

Further Reading
Probably the most famous example of a system with what we have called a transformer
architecture is SYSTRAN. This is described in [Hutchins and Somers1992]. A recent
discussion can be found in [Wilks1992].

A more detailed overview of transfer systems can be found in [Arnold1993].

Examples of transfer systems include the following, ARIANE [Vauquois and Boitet1985],
SUSY [Maas1987], MU (the Japanese National Project) [Nagao et al. July 1986], METAL
[Slocum et al. 1987], [Bennett and Slocum1988], TAUM-AVIATION [Isabelle1987],
ETAP-2 [Apresian et al. 1992], LMT [McCord1989], EUROTRA [Arnold1986,Arnold and
des Tombe1987,Copeland et al. 1991a,Copeland et al. 1991b], , CAT-2 [Sharp1988],
MIMO [Arnold and Sadler1990], MIMO-2 [van Noord et al. 1990], ELU [Estival et al.
1990]. Several of these systems are discussed in detail in [Hutchins and Somers1992].

Among interlingual systems , the following are noteworthy: Rosetta

[Landsbergen1987b,Landsbergen1987a], KBMT [Goodman1989], [Goodman and
Nirenburg1991]. A recent overview is given in [Nirenburg1993]. [Hutchins and Somers1992,
Chapter 6,] is also recommended. One interlingual approach that we have not mentioned here
is that which uses a human language as the interlingual. The best known example of this is
DLT, which uses Esperanto, see [Schubert1992] and [Hutchins and Somers1992, Chapter 17,].

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node39.html [20/09/2000 18:03:15]

Dictionaries

Next: Introduction Up: Machine Translation Previous: Further Reading

Dictionaries

● Introduction
● Paper Dictionaries
● Types of Word Information
● Dictionaries and Morphology
❍ Inflection
❍ Derivation
❍ Compounds
● Terminology
● Summary
● Further Reading

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node40.html [20/09/2000 18:03:16]

Introduction

Next: Paper Dictionaries Up: Dictionaries Previous: Dictionaries

Introduction
This Chapter is about the role played by dictionaries in MT. Our decision to devote a whole
chapter to this discussion reflects the importance of dictionaries in MT:
● Dictionaries are the largest components of an MT system in terms of the amount of
information they hold. If they are more then simple word lists (and they should be, if a
system is to perform well), then they may well be the most expensive components to
construct.
● More than any other component, the size and quality of the dictionary limits the scope
and coverage of a system, and the quality of translation that can be expected.
● The dictionaries are where the end user can expect to be able to contribute most to a
system --- in fact, an end user can expect to have to make some additions to system
dictionaries to make a system really useful. While MT suppliers rarely make it possible
for users to modify other components, they normally expect them to make additions to
the dictionary. Thus, from the point of view of a user, a basic understanding of dictionary
construction and sensitivity to the issues involved in `describing words' is an important
asset.
● In discussing dictionaries here, we include also some discussion of terminology --- it is
with respect to the treatment of terminology that MT provides some of its most useful
benefits.
We shall approach the question of dictionaries in MT obliquely, by considering in some detail
the information contained in, and issues raised by, the paper dictionaries with which we are all
familiar. There are a number of reasons for this, but the most important is that the dictionaries
in existing MT systems are diverse in terms of formats, coverage, level of detail and precise
formalism for lexical description. This diversity should not be a surprise. Different theories of
linguistic representation can give rise to different views of the dictionary, and different
implementation strategies can make even fundamentally similar views of the dictionary look
very different in detail. Moreover, the different kinds of MT engine obviously put quite
different requirements on the contents of the dictionary. For example, dictionaries in an
interlingual system need not contain any translation information per se, all that is necessary is
to associate words with the appropriate (collections of) interlingual concepts. By contrast,
transformer systems will typically give information about source language items, and their
translations, including perhaps information that is really about the target language, and which is
necessary to trigger certain transformations (e.g. to do with the placement of particles like up in
look it up, and look up the answer). Since transfer systems typically use more abstract levels of
representation, the associated dictionaries have to contain information about these levels.
Moreover, in a transfer system, especially one which is intended to deal with several languages,
it is common to separate monolingual dictionaries for source and target languages (which give
information about the various levels of representation involved in analysis and synthesis ),
from bilingual dictionaries which are involved in transfer (which normally relate source and
http://privatewww.essex.ac.uk/~doug/book/node41.html (1 de 2) [20/09/2000 18:03:17]
Introduction

target lexical items, and which normally contain information only about the levels of
representation that are involved in transfer).
We would like to abstract away from these divergences and points of detail in order to focus on
the main issues. Accordingly, we will begin with a brief discussion of typical entries that one
might find in a good monolingual `paper' dictionary, and a good bilingual `paper' dictionary.
We will then briefly discuss the sort of information about words that one typically finds in
MT dictionaries, outlining some of the different ways such information can be represented. As
we have said, a simple view is that a dictionary is a list of words. However, it is impractical,
and perhaps impossible to provide an exhaustive list of words for most languages. This is
because of the possibility of forming new words out of existing ones, by various morphological
processes. In Section we will look briefly at these, and provide some discussion of how they
can be dealt with, and the problems they raise in an MT context. In Section we will briefly
describe the difference between terminology and general vocabulary.

Next: Paper Dictionaries Up: Dictionaries Previous: Dictionaries

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node41.html (2 de 2) [20/09/2000 18:03:17]

Paper Dictionaries

Next: Types of Word Up: Dictionaries Previous: Introduction

Paper Dictionaries
The best place to start our discussion is by looking at typical entries that one might find in a
monolingual English dictionary (cf. page ), and a bilingual dictionary (cf. page ).

We will start by looking at the layout of the first half of the monolingual entry. The entry for
button starts off with the word itself in bold print. This is called the head word. The dot in the
word indicates where the word may be broken off (e.g. for hyphenation). After that there is a
phonetic transcription of the word's pronunciation . Then the entry falls into two main parts,
describing first the noun and then the verb button. Definitions identify two different meanings,
or readings of the noun button, with examples of usage given in italics. The refers the reader
to a related entry. Idiomatic expressions are given under 3. As for the verb, the code [Tn,Tn.p]
indicates that the verb is transitive, i.e. appears in a sentence with a subject and an object (Tn),
or is transitive with an adverbial particle (Tn.p). In this case the adverbial particle is the
preposition up. Under b another usage is described where button is an intransitive verb and thus
takes only a subject (I), or a subject plus the preposition up (Ip). Idioms appear under 2. The
box halfway through the entry signals the start of a list of complex forms, a phrasal verb (
button up), and several compound s, which we will discuss later in this chapter. The verb, and
noun, phrasal verbs and compound s are given in a standard form (the citation form), with
information about stress (given by raised or lowered apostrophes). By convention, this is
normally the singular form of nouns, and the infinitive form of verbs (i.e. the form that one
finds after to, as in to button, to be, etc.)

The bilingual entry for the noun printer begins with the head word, its pronunciation and word
class, in this case noun. Logically, the entry then divides into three component parts (a), (b),
and (c), essentially distinguishing three different uses or meaning of the noun in English which
have distinct translations into French . Where a particular meaning can be identified by
reference to a subject field, this information is given (bracketed, in italics) --- here computation
and photography are identified as subject fields. If the context of use is other than these two
fields, then the translation given under (a) is assumed to be appropriate. For each reading, the
gender of the translation is given: m or f (for masculine or feminine, mf indicates either is
possible; where the masculine and feminine forms differ, both are indicated --- printer's reader
is thus either correcteur or correctrice). If two different translations are possible they are both
given, separated by a comma (thus, either typographe, or imprimeur are possible `general'
translations). The entry also contains some examples of idioms , or other usages, again with the
appropriate translations.
Normal, `paper' dictionaries, are collections of entries such as these. That is, they are basically
lists of words, with information about the various properties. While grammar rules define all the
http://privatewww.essex.ac.uk/~doug/book/node42.html (1 de 2) [20/09/2000 18:03:18]
Paper Dictionaries

possible linguistic structures in a language, the descriptions of individual words that are found
in the dictionary or dictionaries state which words can appear in which of the different
structures. A common (though not completely correct) view is that dictionaries contain all the
`idiosyncratic', `irregular', or unpredictable information about words, while grammars provide
general rules about classes of word, and phrases (this is only true if one excludes morphological
rules and idiom s from the dictionary --- the former can be viewed as dealing with classes of
word, and the latter are phrases).
One can get an idea of the sheer volume of information of this kind that may be needed by
considering that for commercial purposes a lexicon with 20 000 entries is often considered as
the minimum. This however is still only a modest percentage of existing words --- the Oxford
English Dictionary contains about 250 000 entries without being exhaustive even of general
usage. In fact, no dictionary can ever be really complete. Not only do dictionaries generally
restrict themselves to either general, or specialist technical vocabulary (but not both), in
addition, new words are constantly being coined, borrowed, used in new senses, and formed by
normal morphological processes.

Next: Types of Word Up: Dictionaries Previous: Introduction

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node42.html (2 de 2) [20/09/2000 18:03:18]

Types of Word Information

Next: Dictionaries and Morphology Up: Dictionaries Previous: Paper Dictionaries

Types of Word Information

We have already observed that dictionaries are a, perhaps the, central component of MT
systems. In earlier Chapters, we have presented a highly simplified view of dictionaries --- for
example, in Chapter the dictionary was sometimes little more than a list of rules such as v
walk, which only allows information about part of speech to be represented, and in
Chapter we gave translation rules which simply paired up the citation forms of source and
target words (e.g. temperature temperatur). However, though some of the
information that is found in a typical paper dictionary is of limited value in MT (e.g.
information about pronunciation is only useful in speech to speech systems), in general the
quality and detail of the information one needs for MT is at least equal to that which one finds
in paper dictionaries. In this section we discuss the various pieces of information about words
that a good MT system must contain, basing ourselves on the dictionary entries above. An issue
we will not address in this Chapter is the treatment of idiom s, which one typically finds in
paper dictionary entries. We discuss the treatment of idiom s in Chapter .
It is useful to make a distinction between the characteristics of a word itself (its inherent
properties) and the restrictions it places on other words in its grammatical environment.
Although this distinction is not explicitly drawn in paper dictionaries, information of both types
is available in them. Information about grammatical properties includes the indication of
gender in the French part of the bilingual dictionary entry, and the indication of number on
nouns (typically, the citation form of nouns is the singular form, and information about number
is only explicitly given for nouns which have only plural forms, such as scissors, and trousers).
Information about the grammatical environment a word can appear in is normally thought of as
dividing into two kinds: subcategorization information, which indicates the syntactic
environments that a word can occur in, and selectional restrictions which describe semantic
properties of the environment. Typical information about subcategorization is the information
that button is a transitive verb. This is expressed in the verb code [Tn] in the dictionary entry on
page . More precisely, this indicates that it is a verb that appears as the HEAD of sentences
with a (noun phrase) SUBJECT and a (noun phrase) OBJECT. The following gives some
examples, together with the appropriate verb codes from OALD:
● ❍ The president died. [I]

❍ The Romans destroyed the city. [Tn]

❍ Sam gave roses to Kim. [Dn.pr]

❍ Sam gave Kim roses. [Dn.n]

❍ Sam persuaded Kim to stay at home. [Cn.t]

❍ Kim believed that the library was closed. [Tf]

❍ The quality is low. [La]

http://privatewww.essex.ac.uk/~doug/book/node43.html (1 de 7) [20/09/2000 18:03:22]

Types of Word Information

❍ Sam appeared the best man for the job. [Ln]

Note that [I] refers to intransitive verbs that only need a subject to form a grammatical sentence,
[Tn] to transitive verbs (like button) that need a subject and an object, [Dn.pr] to ditransitive
verbs which take a subject and two objects, where the second one is introduced by the
preposition to, [Dn.n] to ditransitive verbs that take a subject plus two object nouns, [Cn.t] to
complex transitive verbs which require a subject, object and an infinitival (non-tensed) clause
introduced by to, [Tf] to transitive verbs taking a subject, object and a finite (tensed) sentence
introduced by that, [La] to linking verbs which link an adjectival phrase (which describes in
some way the subject), to the subject, and [Ln] refers to linking verbs which link a noun phrase
to the subject.
Verbs are not the only word categories that subcategorize for certain elements in their
environment. Nouns exhibit the same phenomenon, like those nouns that have been derived
from verbs (deverbal nouns).
● ❍ The death of the president shocked everybody.

❍ The destruction of the city by the Romans was thorough.

Similarly, there are some adjectives that subcategorize for certain complements. Note that in
the examples below we find three different types of complements, and that b and c differ
from each other because in b the subject of the main clause is also the understood subject of
the subclause, whereas in c the subject of the main clause is the understood object of the
subclause.
● ❍ Mary was proud of her performance.

❍ He was eager to unwrap his present.

❍ That matter is easy to deal with.

An adequate dictionary of English would probably have to recognize at least twenty different
subcategorization classes of verb, and a similar number for adjectives and nouns.
The reason one cannot be precise about the number of different subcategorization classes is that
it depends (a) on how fine the distinctions are that one wants to draw, and (b) on how far one
relies on rules or general principles to capture regularities. For example, probably all verbs
allow coordinated subjects such as Sam and Leslie, but there are some, like meet, where this is
equivalent to an ordinary transitive SUBJECT-VERB-OBJECT construction (cf. ( a), and ( b)

mean the same, but ( c) and (

d) do not). One could decide to recognise this distinction by creating a separate
subcategorization class, thus extending the number of classes. But one could also argue that
this fact about meet and similar verbs is probably related to their semantics (they describe
symmetric relations, in the sense that if A meets B, then B meets A), and is thus regular and
predictable. The appropriate approach could then be to treat it by means of a general linguistic
rule (perhaps one that transforms structures like ( a) into ones of the form ( b)) Of course,
unless one can rely on semantic information to pick out verbs like meet, one will have to

http://privatewww.essex.ac.uk/~doug/book/node43.html (2 de 7) [20/09/2000 18:03:22]

Types of Word Information

introduce some mark on such verbs to ensure that they, and only they, undergo this rule.
However, this is not necessarily the same as introducing a subcategorization class.
● ❍ Sam met Mary

❍ Sam and Mary met

❍ Sam saw Mary

❍ Sam and Mary saw

Subcategorization information indicates that, for example, the verb button occurs with a noun
phrase OBJECT. In fact, we know much more about the verb than this --- the OBJECT, or in
terms of semantic roles , the PATIENT, of the verb has to be a `buttonable' thing, such as a
piece of clothing, and that the SUBJECT (more precisely AGENT) of the verb is normally
animate. Such information is commonly referred to as the selectional restrictions that
words place on items that appear in constructions where they are the HEAD. This information
is implicit in the paper dictionary entry above --- the information that the object of button is
inanimate, and normally an item of clothing has to be worked out from the use of sth (= `some
thing') in the definition, and the example, which gives coat, jacket, shirt as possibilities. The
entry nowhere says the SUBJECT of the verb has to be an animate entity (probably human),
since no other entity can perform the action of `buttoning'. It is assumed (rightly) that the
human reader can work this sort of thing out for herself. This information has to be made
explicit if it is to be used in analysis , transfer or synthesis , of course.
Basic inherent information and information about subcategorization and selectional restrictions
can be represented straightforwardly for MT purposes. Essentially, entries in an MT dictionary
will be equivalent to collections of attributes and values (i.e. features). For example, one might
have something like the following for the noun button, indicating that its base, or citation form
is button, that it is a common noun, which is concrete (rather than abstract, like
happiness, or sincerity)
lex = button
cat = n
ntype = common
number =
human = no
concrete = yes
An obvious way to implement such things is as records in a database , with attributes naming
fields (e.g. cat), and values as the contents of the fields (e.g. n). But it is not always necessary
to name the field --- one could, for example, adopt a convention that the first field in a record
always contains the citation form (in this case the value of the feature lex) , that the second
field indicates the category, and that the third field some sort of subdivision of the category.
Looking at the dictionary entry for the noun button it becomes clear that different parts of
speech will have a different collection of attributes. For example, verbs will have a vtype,
rather than an ntype feature , and while verbs might have fields for indications of number ,
person and tense , one would not expect to find such fields for prepositions. In the entry we
have given we also find one attribute --- number --- without a value. The idea here is to
indicate that a value for this attribute is possible, but is not inherent to the word button, which
may have different number values on different occasions (unlike e.g. trousers, which is always
plural). Of course, this sort of blank field is essential if fields are indicated by position, rather

http://privatewww.essex.ac.uk/~doug/book/node43.html (3 de 7) [20/09/2000 18:03:22]

Types of Word Information

than name. In systems which name attribute fields it might simply be equivalent to omitting the
attribute, but maintaining the field is still useful because it helps someone who has to modify
the dictionary to understand the information in the dictionary. An alternative to giving a blank
value, is to follow the practice of some paper dictionaries and fill in the default, or (in some
sense) normal value. For an attribute like number, this would presumably be singular. This
alternative, however, is unfashionable these days, since it goes against the generally accepted
idea that in the best case linguistic processing only adds, and never changes information. The
attraction of such an approach is that it makes the order in which things are done less critical
(cf. our remarks about the desirability of separating declarative and procedural information in
Chapter ).
In order to include information about subcategorization and selectional restrictions, one has
two options. The first is to encode it via sets of attributes with atomic values such as those
above. In practice, this would mean that one might have features such as
subcat=subj_obj, and sem_patient=clothing. As regards subcategorization
information, this is essentially the approach used in the monolingual paper dictionary above.
The resulting dictionary entry could then look something like the following:
lex = button
cat = v
vtype = main
finite =
person =
number =
subcat = subj_obj
sem_agent = human
sem_patient = clothing
In some systems this may be the only option. However, some systems may allow values to be
sets, or lists, in which case one has more flexibility. For example, one might represent
subcategorization information by means of a list of categories, for example subcat =
[np,np,np] might indicate a verb that allows three NPs (such as give), and [np,np,pp]
might indicate a verb that takes two NPs and a PP (again like give).
● ❍ Sam gave roses to Kim. ( subcat = [np,np,pp])

❍ Sam gave Kim roses. ( subcat = [np,np,np])

A further refinement would be to indicate the actual grammatical relations involved, perhaps as
in subcat = [SUBJ:np, OBJ:np, IOBJ:pp]. A notation which allows the
lexicographer to indicate other properties of the items would be still more expressive. For
example, it would be useful to indicate that with give, the preposition in the PP has to be to.
This would mean that instead of ` pp' and ` np' one would have collections of features , and
perhaps even pieces of syntactic structure. (A current trend in computational linguistics
involves the development of formalisms that allow such very detailed lexical entries, and we
will say a little more about them in Chapter ).
Turning now to the treatment of translation information in MT dictionaries, one possibility is to
attempt to represent all the relevant information by means of attributes and values. Thus, as an
addition to the dictionary entry for button given above, a transformer system could specify a
`translation' feature which has as its value the appropriate target language word; e.g. trans =
bouton for translation into French . One might also include features which trigger certain

http://privatewww.essex.ac.uk/~doug/book/node43.html (4 de 7) [20/09/2000 18:03:22]

Types of Word Information

transformations (for example for changing word order for certain words). However, this is not a
particularly attractive view. For one thing, it is clearly oriented in one direction , and it will be
difficult to produce entries relating to the other direction of translation from such entries. More
generally, one wants a bilingual dictionary to allow the replacement of certain source language
oriented information with corresponding target language information --- i.e. replace the
information one derives from the source dictionary by information derived from the target
dictionary. This suggests the usage of translation rules which relate head words to head words.
That is, rules of the type we introduced in Chapter , like temperature temperatur.
As we noted before, not all translation rules can be a simple mapping of source language words
onto their target language equivalents. One will have to put conditions on the rules.
For example, one might like to be able to describe in the bilingual entry that deals with like and
plaire, the change in grammatical relations that occurs if one is working with relatively shallow
levels of representation,. In effect, the transfer rule that we gave for this example in Chapter
might be seen as a bilingual lexical entry. Other translation rules that may require more than
just a simple pairing of source and target words are those that treat phenomena like idiom s and
compound s, and some cases of lexical holes (cf. Chapter ). To deal with such phenomena
bilingual dictionary entries may have a single lexical item on the side of one language, whereas
the other side describes a (possibly quite complex) linguistic structure.
The entry for button taken from a paper dictionary at the beginning of this Chapter illustrates an
issue of major importance to the automatic processing of some languages, including English.
This is the very widespread occurrence of homography in the language. Loosely speaking,
homographs are words that are written in the same way. However, it is important to distinguish
several different cases (sometimes the term homography is restricted to only one of them).
1. The case where what is intuitively a single noun (for example) has several different
readings. This can be seen with the entry for button on page , where a reading relating
to clothing is distinguished from a `knob' reading.
2. The case where one has related items of different categories which are written alike. For
example, button can be either a noun or a verb.
3. The case where one has what appears to be unrelated items which happen to be written
alike. The classic example of this is the noun bank, which can designate either the side of
a river, or a financial institution.
These distinctions have practical significance when one is writing (creating, extending, or
modifying) a dictionary, since they relate to the question of when one should create a new entry
(by defining a new headword). The issues involved are rather different when one is creating a
`paper' dictionary (where issues of readability are paramount) or a dictionary for MT, but it is in
any case very much a pragmatic decision. One good guiding principle one might adopt is to
group entries hierarchically in terms of amounts of shared information. For example, there is
relatively little that the two senses of bank share apart from their citation form and the fact that
they are both common nouns, so one may as well associate them with different entries. In a
computational setting where one has to give unique names to different entries, this will involve
creating headwords such as bank_1 and bank_2, or ( bank_finance, and
bank_river). As regards the noun and verb button, though one might want to have some
way of indicating that they are related, they do not share much information, and can therefore
be treated as separate entries. For multiple readings of a word, for example, the two readings of

http://privatewww.essex.ac.uk/~doug/book/node43.html (5 de 7) [20/09/2000 18:03:22]

Types of Word Information

the noun button, on the other hand, most information is shared --- they differ mainly in their
semantics. In this case, it might be useful to impose an organization in the lexicon in which
information can be inherited from an entry into sub-entries (or more generally, from one entry
to another), or to see them as subentries of an abstract `protoentry' of some sort. This will
certainly save time and effort in dictionary construction --- though the savings one makes may
look small in one case, it becomes significant when multiplied by the number items that have
different readings (this is certainly in the thousands, perhaps the hundreds of thousands, since
most words listed in normal dictionaries have at least two readings). The issues this raises are
complex and we cannot do them justice here, however, the following will give a flavour of what
is involved.
More generally, what one is talking about here is inheritance of properties between entries (or
from entries into subentries). This is illustrated in Figure . One could imagine extending this,
introducing abstract entries expressing information true of classes of (real) entry. For example,
one might want to specify certain facts about all nouns (all noun readings) just once, rather than
stating them separately in each entry. The entry for a typical noun might then be very simple,
saying no more than `this is a typical noun', and giving the citation form (and semantics, and
translation, if appropriate). One allows for subregularities (that is lexical elements which are
regular in some but not all properties), by allowing elements to inherit some information while
expressing the special or irregular information directly in the entry itself. In many cases, the
optimal organization can turn out to be quite complicated, with entries inheriting from a number
of different sources. Such an approach becomes even more attractive if default inheritance is
possible. That is, that information is inherited, unless it is explicitly contradicted in an
entry/reading --- it would then be possible to say, for example, `this is a typical noun, except for
the way it forms its plural'.

Figure: Inheritance
One final and important component of an MT dictionary, which is entirely lacking in paper
dictionaries (at least in their printed, public form) is documentation. Apart from general
documentation describing design decisions, and terminology , and providing lists and
definitions (including operational tests) for the attributes and values that are used in the
dictionary (it is, obviously, essential that such terms are used consistently --- and consistency is
a problem since creating and maintaining a dictionary is not a task that can be performed by a
single individual), it is important that each entry include some lexicographers' comments ---
information about who created the entry, when it was last revised, the kinds of example it is
based on, what problems there are with it, and the sorts of improvement that are required. Such
information is vital if a dictionary is to be maintained and extended. In general, though the
quality and quantity of such documentation has no effect on the actual performance of the
dictionary, it is critical if a dictionary is to be modified or extended.

Next: Dictionaries and Morphology Up: Dictionaries Previous: Paper Dictionaries

http://privatewww.essex.ac.uk/~doug/book/node43.html (6 de 7) [20/09/2000 18:03:22]

Types of Word Information

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node43.html (7 de 7) [20/09/2000 18:03:22]

Dictionaries and Morphology

Next: Inflection Up: Dictionaries Previous: Types of Word

Dictionaries and Morphology

Morphology is concerned with the internal structure of words, and how words can be formed. It
is usual to recognize three different word formation processes.
1
Inflectional processes , by means of which a word is derived from another word form,
acquiring certain grammatical features but maintaining the same part of speech or
category (e.g. walk, walks);
2
Derivational processes in which a word of a different category is derived from another
word or word stem by the application of some process (e.g. grammar grammatical,
grammatical grammaticality);
3
Compounding, in which independent words come together in some way to form a new
unit ( buttonhole).
In English, inflectional and derivational processes involve prefixes (as in undo) and suffixes
(as in stupidity), and what is called conversion, or zero-affixation where there is a change of
category, but no change of form (and example would be the process that relates the noun button
to the verb). In other languages, a range of devices such as changes in the vowel patterns of
words, doubling or reduplication of syllables, etc., are also found. Clearly, these prefixes and
suffixes (collectively known as affixes) cannot `stand alone' as words. Compound ing is quite
different in that the parts can each occur as individual words. Compounding is a very
productive phenomenon in the Germanic languages, and poses some particular problems in
MT, which we will discuss later.

● Inflection
● Derivation
● Compounds

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node44.html [20/09/2000 18:03:23]

Inflection

Next: Derivation Up: Dictionaries and Morphology Previous: Dictionaries and Morphology

Inflection

As a rule, paper dictionaries abstract away from inflection. Head words are generally
uninflected, that is, nouns appear in singular form and verbs have the base (or infinitival) form.
There are a number of reasons for this. The first is that inflection is a relatively regular process,
and once the exceptional cases have been separated out, inflectional processes apply to all
members of a given category. For example, to form the third person singular of the present
tense of verbs one simply suffixes s (or its variant es) to the citation form of the verb. There are
very few exceptions to this rule. Since it is a regular process, the dictionary user can be relied
upon to form regularly inflected words from the citation forms given in the dictionary at will.
Of course, irregularities, such as irregular plurals ( sheep, oxen, phenomena, etc.) and plural
only nouns ( trousers) must be stated explicitly. A second important reason is eminently
practical --- it saves space, time and effort in constructing entries. Since English inflectional
morphology is rather impoverished, these savings are not enormous. But Spanish , for example,
has six different verb forms for the present tense , and if we add those for the past tense (either
imperfecto or pretérito in Spanish) it amounts to 16 different verb forms. Other languages
make even more use of inflections, like, for example, Finnish where there are said to be in the
region of 2000 forms for most nouns, and 12 000 forms for each verb. It will be obvious that
the need to describe inflectional variation by means of rules is pressing in such cases.
Within the context of MT, it is clearly desirable to have a similar approach, where monolingual
and transfer dictionaries only contain the head words and no inflected words. In order to
achieve this a system must be capable of capturing the regular patterns of inflection . This can
be done by adding a morphological component to the system, which describes all the regular
inflections in general rules, with additional explicit rules for irregular inflection, thus allowing
dictionary writers to abstract away from inflected forms as much as possible. The
morphological component will be able to map inflected words onto the appropriate head words
and will retain the information provided by the inflectional affix by adding the relevant
features.
Let us consider again the verb affects in the simple sentence Temperature affects density. First,
we want our morphological component to recognize affects as an inflected form of affect.
Secondly, we want to retain the information carried by the affix so we can use it later when
generating the output sentence. In the case of affects this means we want to state that the verb is
finite, or tense d (in fact, present tense ). This is important since it allows the verb to occur as
the only verb of a main clause. The tense also prevents the verb from occurring behind
auxiliary verbs like will. Other information that we gather from the inflection is the fact that the
verb is third person (as opposed to first person, occurring with I or we, and as opposed with
second person, occurring with you), and that it is singular (rather than third person plural, which
occurs with they, or with a plural noun).

http://privatewww.essex.ac.uk/~doug/book/node45.html (1 de 3) [20/09/2000 18:03:25]

Inflection

There are various ways of describing this, but perhaps the simplest is to use rules of the
following form:
(lex=V,cat=v,+finite,person=3rd,number=sing,tense=pres)
V+ s
Here we have introduced a rule which says that finite verbs which are third person singular and
have present tense ( cat=v, +finite, person=3rd, number=sing,
tense=pres) can be formed by adding s to the base form (the base form is represented as
the value of the attribute lex). The rule can also be read in the opposite direction: if a word can
be divided into a string of characters and s, then it may be a finite verb with third person
singular in present tense . Other rules would have to be given to indicate that the + s ending
can be added to all verbs, except for those that end in +s, themselves, in which case es is
added (cf. kiss, kisses).
Whether something is indeed the base form of the verb can be verified in the monolingual
dictionary . So, if the morphological analyser encounters a word like affects, it will check
whether the monolingual dictionary contains an entry with the features cat = v, lex =
affect . Since it does, affects can be represented by means of the lexical entry, with some of
the information supplied by the rule.The result of morphological analysis then is a
representation which consists of both the information provided by the dictionary and the
information contributed by the affix.
lex = affect
cat = v
vtype = main
subcat = subj_obj
sem_agent = ?
sem_patient = ?
vform = finite
person = 3rdsing
tense = pres
In order to recognize irregular forms the morphological component has to contain explicit rules.
One approach here is to try to normalise the spelling , so that the ordinary morphological rules
can deal with the result. For example, one might have rules like the following to deal with the
irregular third person singular forms of be and have.
be+s is have+s has
Under this approach, morphological analysis for is and has is a two stage process.
The alternative is to state the relationship between the forms is and has directly, via rules like
the following:
(lex=be,cat=v,+finite,person=3rd,number=sing,tense=pres)
is
(lex=have,cat=v,+finite,person=3rd,number=sing,tense=pres)
has

http://privatewww.essex.ac.uk/~doug/book/node45.html (2 de 3) [20/09/2000 18:03:25]

Inflection

A graphic interpretation of the two alternative approaches is given in Figure .

Figure: Treatment of Irregular Verbs

Notice that we must ensure that these rules apply in the right cases. For example, dies should
not be analysed as di+ es. This is not problematic, providing we ensure that the analyses we
produce contain actual lexical items.
In synthesis , there is a related problem of making sure that the regular rules do not produce *
bes, and * haves. One approach to this is to try to divide rules into exceptional and default
groups, and to make sure that no default rule applies if a an exceptional rule can apply. Thus,
for example, the fact that there is a special rule for the third person singular form of is would
prevent the application of the normal or default rule that simply adds s to the base form of the
verb.
Alternatively, one could add feature s to control which rules apply to lexical entries, and have
the morphological rules check for the presence of the particular feature . This approach is
particularly attractive in cases where a language has a number of conjugation or declension
classes --- lexical items can contain features indicating their conjugation/declension class,
which the morphological rules can check.
So far, we have talked about morphological rules as things that actually apply as a sentence is
being analysed. Another way in which one could use them is to compile out a full form
dictionary from a dictionary of uninflected words, essentially by running the morphological
rules over the dictionary of uninflected forms. Note, however, that this strategy would build a
monolingual dictionary of an enormous size for languages like Spanish , or Finnish .

Next: Derivation Up: Dictionaries and Morphology Previous: Dictionaries and Morphology

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node45.html (3 de 3) [20/09/2000 18:03:25]

Derivation

Next: Compounds Up: Dictionaries and Morphology Previous: Inflection

Derivation
Derivation processes form new words (generally of a different category) from existing words,
in English this is mainly done by adding affixes. For example, industrialization, and destruction
can be thought of as being derived in the way illustrated below. As one can see from
destruction, it is not necessarily the citation form of a word that appears in derivations, for this
reason it is common to talk of derivational processes involving stems and affixes (rather than
words and affixes).
● ❍ industry +ial

+ize +ation

❍ destroy +ion

In a paper dictionary, some derived words are listed, under the relevant head word. This is
partly because affixes differ in their productivity and in the regularity of the effect they have on
the words or stems that they combine with. For example, there seems to be no real basis on
which to predict which of the noun-forming affixes produce nouns from particular verbs. This
is illustrated below by the verbs arrive, destroy, and deport:

● ❍

However, some derivational processes are quite regular and can be described by means of a bf
word grammar. This involves: (i) entering the affix in the dictionary; (ii) allowing it to
subcategorize for what it combines with (e.g. -able combines with transitive verbs: witness
read- readable) --- this is just like normal syntactic subcategorization ; (iii) making sure that the
rules to combine words and affixes give the derived word the correct features for the result, and
take care of any spelling changes in word or affix; (iv) finding some way of specifying the
meaning in terms of the meanings of the word and affix.
As with inflection , the rules must be set up so as to produce only genuine lexical items. For
example, we can ensure that the rules that analyse cordiality as cordial+ -ity do not produce
qual+ -ity from quality, because there is no lexical item * qual.
One approach to handling derivational morphology in MT is to simply list all derived words,
and for some derived words (e.g. landing, in the sense of area at the top of stairs), this is clearly
the right approach, because their meaning is unpredictable. But not all derivational morphology
is unpredictable. Some affixes almost always have just one sense, like the prefix un which
(when combined with an adjective) normally means `not X' ( unhappy means not happy) ,
and for others there are certain tendencies or regularities: with the examples in ( ) the
addition of the suffix -ing to the verb stem seems to have the same, regular consequence for the
meaning of the word, so the derived word denotes the action or process associated with the verb

http://privatewww.essex.ac.uk/~doug/book/node46.html (1 de 3) [20/09/2000 18:03:29]

Derivation

(the act of Xing). Speakers exploit this fact by creating new words which they expect hearers to
understand.
● ❍ The killing of elephants is forbidden.

❍ Driving off went without any problems.

❍ The painting of still lives never appealed to me.

In contrast with the examples in ( ), one should consider the nouns in ( ), where the meaning,
although common, is not predictable from the suffix -ing:
● ❍ Painting: a picture produced with paint

❍ Covering: something which covers something

❍ Cutting: something which has been cut out

❍ Crossing: a place were e.g. roads cross

We see here that a verb+ing noun can refer to a product ( a), a thing which performs an action
( b), a thing which undergoes an action ( c), or a place ( d). At the same time, however, it
is true that in most cases the regular interpretation `the act of Xing' is also available. What this
means is that there is almost always a problem of ambiguity with derived words.
Moreover, there are cases where one can translate derived words by translating the stem, and
translating the affix. For example, the French translation of English adverbs formed from an
adjective plus - ly is often made up of the translation of the adjective plus - ment (e.g. quick+ly
rapide+ment, easy+ly facile+ment), etc. But this is only possible for some affixes, and
only when the interpretation of the derived word is predictable. The difficulties of translating
derived words by translating stems and affixes can can be seen from the translation of the
previous examples into Dutch .
● ❍ killing doden
❍ driving off wegrijden
❍ painting (the act) schilderen
❍ painting (the product) schilderen, but schilderij

❍ covering bedekken, but bedekking

❍ cutting knippen, but knipsel

❍ crossing kruisen, but kruispunt

Thus, though the idea of providing rules for translating derived words may seem attractive, it
raises many problems and so it is currently more of a research goal for MT than a practical
possibility.

http://privatewww.essex.ac.uk/~doug/book/node46.html (2 de 3) [20/09/2000 18:03:29]

Derivation

Next: Compounds Up: Dictionaries and Morphology Previous: Inflection

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node46.html (3 de 3) [20/09/2000 18:03:29]

Compounds

Next: Terminology Up: Dictionaries and Morphology Previous: Derivation

Compounds

A compound is a combination of two or more words which functions as a single word. In English, the most
common type of compound is probably a compound made up of two nouns (noun-noun compounds), such
as those in the dictionary entry for button:
● ❍ buttonhole:
button hole

❍ buttonhook:
button hook

❍ button mushroom:
button mushroom

In Spanish , for example, other types of compound s are equally important, including adjective-adjective
compounds:
● ❍ guardacostas (`coastguard'):
guarda

http://privatewww.essex.ac.uk/~doug/book/node47.html (1 de 3) [20/09/2000 18:03:34]

Compounds

costas
❍ rojiblanco (`red and white'):
roji blanco

Orthographically , different languages follow different conventions for compounds . For example, in
German compounds are generally written as one word, but in English some are written as one word (as
buttonhole and buttonhook above), some as hyphenated words (e.g. small-scale) and some as juxtaposed
words (e.g. button mushroom).
As with derivations, it is possible to describe the range of possible compounds by means of a word
grammar, and as with derivations the possibility that one might be able to translate compounds by
translating the component parts is very attractive --- especially since it is in principle not possible to list all
English compounds, because compounding can give rise to words that are arbitrarily long. To see this,
consider that one can form, in addition to film society:
● ❍ student film

❍ student film society

❍ student film society committee

❍ student film society committee scandal

❍ student film society committee scandal inquiry

Unfortunately, though there are cases where decomposing a compound and translating its parts gives
correct results (e.g. the German compound Wassersportverein translates as water sport club), the problems
of interpretation and translation are even worse for compounds than for derivation s. Apart from the fact

http://privatewww.essex.ac.uk/~doug/book/node47.html (2 de 3) [20/09/2000 18:03:34]

Compounds

that some compounds have completely idiosyncratic interpretations (e.g. a redhead is a person with ginger
coloured hair), there are problems of ambiguity . For example, student film society could have either of the
structures indicated, with different interpretations (the first might denote a society for student films, the
second a film society for students):
● ❍ student film society

❍ student film society

A different type of ambiguity can be illustrated by giving an example: satellite observation may on one
occasion of use mean observation by satellite, while on another occasion of use it might mean observation
of satellites. Most of the time humans are able to rely on either our world knowledge or on the context to
unravel a compound's meaning. Moreover, it is frequently important for translation purposes to work out
the exact relation expressed by a compound. In Romance languages, for example, this relation may be
explicitly realised by a preposition. For example, research infrastructure in Spanish translates as
infraestructura para la investigación (literally, `infrastructure for research'). Nor can we happily assume
that an ambiguity in one language will be preserved in another. Thus satellite observation has two possible
translations in Spanish , depending on its meaning: observación por satelite (`observation by satellite') and
observación de satelites (`observation of satellites').
A further problem with compounds is that a wide variety of relations are possible between the elements of
a compound. Thus buttonhole is a hole for buttons, but button mushroom is a mushroom that resembles a
button. It is not clear how to capture these relations.
Thus, as with derivation s, a really general approach to the treatment of compounds remains a research goal
for MT.

Next: Terminology Up: Dictionaries and Morphology Previous: Derivation

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node47.html (3 de 3) [20/09/2000 18:03:34]

Terminology

Next: Summary Up: Dictionaries Previous: Compounds

Terminology

The discussion so far has been about issues relating to general vocabulary. However, a slightly
different, and somewhat less troublesome, set of issues arise when one turns to the specialist
vocabulary that one finds in certain types of text in certain subject fields (the vocabulary of
weather reports is an extreme example, other examples might be the vocabulary of reports on
trials for medical reports, reports of tests of pharmaceutical drugs, or reports of particular kinds
of sporting event). Such fields often have a relatively well-defined terminology, which is
sometimes even codified, and given official recognition by professional bodies. What this
codification involves is settling on a collection of concepts, and assigning each a name (or
perhaps several names, one in each of several languages). When a word (or collection of words
in several languages) designate a single concept in this way, it is called a term. Examples of
terms include the names for material objects, but also the abstract entities (processes, properties,
functions, etc). Concepts, and hence the associated terms, can be organized into conceptual
structures, based on the relationship between them. For example tables, chairs, cupboards, etc.
can be grouped together as furniture, with a possible subdivision into household furniture and
office furniture.
Terms may be simple words or multiword expressions. Syntactically, there is nothing to
distinguish terms from ordinary language, although there is a strong tendency for terms to be
nouns, often compound nouns.
Terms are potentially more tractable for MT systems than general language vocabulary, since
for the most part they tend to be less ambiguous . While a general language word may represent
more than one concept in a system of concepts, there is frequently a one-to-one mapping
between terms and the concepts they represent. Take for example the word graduation, which
in machine tool terminology has the very precise meaning: ``distribution of divisions on the
scale of an apparatus (linear, logarithmic, quadratic, etc)'' The general language word
graduation, on the other hand, has many more meanings, including ``the ceremony at which
degrees are conferred''. What this means, of course, is that one can in principle adopt an
interlingual approach to terminology. For example, even in a transfer system, one need not
deal with terms on a language pair basis --- all one may need is to have analysis and synthesis
rules which relate the words for individual terms to an interlingual name for the associated
concept (this could be an arbitrary numerical code, a collection of features , or even the actual
term used in one of the languages, of course).
It is not always the case that a term represents one and only one concept --- there are examples
of terms which are ambiguous . For example, in machine tool terminology the term screw is
defined as follows: ``a machine thread whose essential element is a screw thread. A screw is
either an external screw or an internal screw.'' (Likewise, synonymy amongst terms occurs,
though much less frequent than in general language. In machine tool terminology, for example,

http://privatewww.essex.ac.uk/~doug/book/node48.html (1 de 3) [20/09/2000 18:03:36]

Terminology

cramp and clamp appear to designate the same concept.) However, the problems of ambiguity
are small when compared to the problems one has with general vocabulary.
There are still some translational problems with terminology, however. In particular, there are
problems whenever there is a mismatch between the conceptual systems of the two languages
to be translated. An example of a concept mismatch from wine-making terminology is the
difference between the English acid and the French acide which are defined as follows:
● ❍ acid: term applied to wine containing an excessive amount of , usually a wine
made from grapes not completely ripe.
❍ acide: caractère d'un vin dont la teneur elevée en provient généralement de
raisins incomplètement mûrs.
While the French definition speaks of acides organiques (`organic acids'), the English speaks
only of acids. If the mismatch is considered significant enough, the term may need to be
paraphrased in the other language. In such cases translating terminology raises the same
problems as dealing with general vocabulary.
Fortunately, problem cases in terminology translation are much less frequent than in general
vocabulary.
From the point of view of the human translator, and more particularly, groups of human
translators collaborating on the translation of documents, terminology poses other sorts of
problem. First, there is the problem of size --- the sheer number of terms there are to deal with.
Second, there is the problem of consistency.
With respect to the second problem, MT offers a considerable advantage. This is because once
a term has been translated, it is possible to store the term and its translation, and ensure that the
term is translated consistently throughout texts.
Of course, this is partly a solution to the problem of size also, because it ensures that the
research and effort that goes into finding a translation for a term is not duplicated by other
translators working with the same system. However, it is only a partial solution, because there
is a seemingly inexorable increase in terminology in many subject areas. Many hours of
research are put into the recognition and documentation of new terms and their translational
equivalents in other languages. To alleviate this problem, many translators and translation
bureaux make use of termbanks , either pre-existing, or constructed in-house.
Termbanks are basically database s which contain many thousands of entries, one for every
term. These entries consist, just like dictionary entries, of several fields, but the type of
information given in these fields is rather different from that which one finds in an ordinary
dictionary. Partly, this is because the proper documentation of a term requires specific
information about the provenance of the entry, and about when it was created, and when
modified (of course, one would expect to find information of this kind available to the builders
of a properly documented dictionary too). Other information will typically concern related
terms (synonyms, antonyms , abbreviations, superordinate terms and hyponyms ), subject area
(e.g. pharmaceutical products vs. sports goods), and sources of further information (e.g.
specialist dictionaries or reference books). On the other hand, information about grammatical
properties and pronunciation is typically rather scant. This is partly because terms are very
often new words, or loan words, and typically follow the regular morphological rules of a
language. Similarly, the lack of phonological information is partly because the entries are
oriented towards written material, but also because it is expected that the terms will be

http://privatewww.essex.ac.uk/~doug/book/node48.html (2 de 3) [20/09/2000 18:03:36]

Terminology

phonologically regular (i.e. they will follow the normal rules for the language, or the normal
rules that apply to loans words).
Apart from in-house termbanks which are local to a single organization, there are a number of
large termbanks which offer open access (sometimes at a small charge). Examples are
Eurodicautom (European Commission) , Termium (Canadian Government), Normaterm (the
French standards organization) and Frantext (National Institute of the French Language),
which offer a range of terminology areas including science, technology, administration,
agriculture, medicine, law and economics.
It should be evident from even this brief discussion that ensuring clear and consistent use and
translation of terminology is a significant factor in the translation process, which in most
technical domains necessitates the creation and maintenance of termbanks --- itself a costly and
time-consuming endeavour. It is not surprising, therefore, that with the increasing availability
of large amounts of on-line texts, researchers have begun to experiment with the automatic
extraction of terms from running text, using a variety of statistical methods to determine the
likelihood that a word, or string of words, constitutes a term. Of course, lists of (putative) terms
cannot be made to emerge magically from a corpus of texts - the process takes into account the
frequency of items in the texts and is often guided by some information provided by the user,
such as a thesaurus of concepts or concept hierarchy or a list of already identified terms, or a
list of typical syntactic patterns for terms. There is no reason to expect such techniques to be
limited to the extraction of monolingual terminology, and in fact the idea of automating to
some degree the compilation of bilingual and multilingual termbank s is also gaining ground.

Next: Summary Up: Dictionaries Previous: Compounds

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node48.html (3 de 3) [20/09/2000 18:03:36]

Summary

Next: Further Reading Up: Dictionaries Previous: Terminology

Summary
This Chapter has dealt with a number of issues concerning dictionaries in MT, including issues
relating to various kinds of word structure (morphology) , and terminology. Apart from
stressing their importance, we have outlined the main sorts of information that one typically
finds in dictionaries, and raised some questions about how this information should be
represented.

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node49.html [20/09/2000 18:03:36]

Further Reading

Next: Translation Problems Up: Dictionaries Previous: Summary

Further Reading
A readable account of what is involved in producing a dictionary can be found in [Sinclair1987]
--- in this case the dictionary is monolingual , and intended for human readers, but many of the
issues are similar. A general discussion of what are taken to be the main theoretical issues in the
design and construction of dictionaries for NLP purposes is given in [Ritchie1987].

On morphology , [Spencer1991] provides an excellent up-to-date description of current

linguistic theory. For a more extensive discussion of compounding see [Bauer1983]. A
detailed description of the state of the art as regards computational treatments of morphological
phenomena is given in [Ritchie et al. 1992]. Almost the only discussion of morphology which
is specifically related to MT is [Bennett1993].

For a general introduction to the study of terminology, see [Sager1990] , on termbanks , see
[Bennett et al. 1986,McNaught1988b,McNaughtforthcoming,McNaught1988a]. For discussion
of computerized termbanks and translation, see [Thomas1992]. Experience of using a
terminological database in the translation process is reported in [Paillet1990].

These days, many paper dictionaries exist in machine readable form (i.e. they have been created
as `electronic documents' in the sense of Chapter , below). OALD , the Oxford Advanced
Learners' Dictionary [Hornby et al. 1974], from which the monolingual entry on page is
taken, and LDOCE, Longman's Dictionary of Contemporary English [Proctor1978], are typical
in this respect. They are sufficiently consistent and explicit to have been used in a number of
experiments which try to take `paper' dictionaries (or rather the machine readable versions of
them), and convert them into a form which can be used directly in NLP systems. Some of this
work is reported in [Boguraev and Briscoe1989].

The representation and use of lexical information in NLP is the focus of a great deal of research
currently. Some idea of the range of this can be obtained from [Evens1988] and [Pustejovsky
and Bergler1992]. The idea of structuring a dictionary
hierarchically so that individual entries can inherit information (and so be simplified), which we
mentioned briefly, is particularly important in this research. A clearer idea of what is involved
can be gained from [Pollard and Sag1987, Chapter 8,].

Next: Translation Problems Up: Dictionaries Previous: Summary

http://privatewww.essex.ac.uk/~doug/book/node50.html (1 de 2) [20/09/2000 18:03:37]

Further Reading

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node50.html (2 de 2) [20/09/2000 18:03:37]

Translation Problems

Next: Introduction Up: Machine Translation Previous: Further Reading

Translation Problems

● Introduction
● Ambiguity
● Lexical and Structural Mismatches
● Multiword units: Idioms and Collocations
● Summary
● Further Reading

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node51.html [20/09/2000 18:03:38]

Introduction

Next: Ambiguity Up: Translation Problems Previous: Translation Problems

Introduction
In this chapter we will consider some particular problems which the task of translation poses for
the builder of MT systems --- some of the reasons why MT is hard. It is useful to think of these
problems under two headings: (i) Problems of ambiguity , (ii) problems that arise from
structural and lexical differences between languages and (iii) multiword units like idiom s and
collocations . We will discuss typical problems of ambiguity in Section , lexical and
structural mismatches in Section , and multiword units in Section .
Of course, these sorts of problem are not the only reasons why MT is hard. Other problems
include the sheer size of the undertaking, as indicated by the number of rules and dictionary
entries that a realistic system will need, and the fact that there are many constructions whose
grammar is poorly understood, in the sense that it is not clear how they should be represented,
or what rules should be used to describe them. This is the case even for English, which has been
extensively studied, and for which there are detailed descriptions -- both traditional `descriptive'
and theoretically sophisticated -- some of which are written with computational usability in
mind. It is an even worse problem for other languages. Moreover, even where there is a
reasonable description of a phenomenon or construction, producing a description which is
sufficiently precise to be used by an automatic system raises non-trivial problems.

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node52.html [20/09/2000 18:03:38]

Ambiguity

Next: Lexical and Structural Up: Translation Problems Previous: Introduction

Ambiguity

In the best of all possible worlds (as far as most Natural Language Processing is concerned,
anyway) every word would have one and only one meaning. But, as we all know, this is not the
case. When a word has more than one meaning, it is said to be lexically ambiguous. When a
phrase or sentence can have more than one structure it is said to be structurally ambiguous.
Ambiguity is a pervasive phenomenon in human languages. It is very hard to find words that
are not at least two ways ambiguous, and sentences which are (out of context) several ways
ambiguous are the rule, not the exception. This is not only problematic because some of the
alternatives are unintended (i.e. represent wrong interpretations), but because ambiguities
`multiply'. In the worst case, a sentence containing two words, each of which is two ways
ambiguous may be four ways ambiguous ( ), one with three such words may be , ways
ambiguous etc. One can, in this way, get very large numbers indeed. For example, a sentence
consisting of ten words, each two ways ambiguous, and with just two possible structural
analyses could have different analyses. The number of analyses can be problematic, since
one may have to consider all of them, rejecting all but one.
Fortunately, however, things are not always so bad. In the rest of this section we will look at the
problem in more detail, and consider some partial solutions.
Imagine that we are trying to translate these two sentences into French :
● ❍ You must not abrasive cleaners on the printer casing.
❍ The of abrasive cleaners on the printer casing is not recommended.
In the first sentence use is a verb, and in the second a noun, that is, we have a case of lexical
ambiguity. An English-French dictionary will say that the verb can be translated by (inter alia)
se servir de and employer, whereas the noun is translated as emploi or utilisation. One way a
reader or an automatic parser can find out whether the noun or verb form of use is being
employed in a sentence is by working out whether it is grammatically possible to have a noun
or a verb in the place where it occurs. For example, in English, there is no grammatical
sequence of words which consists of the + V + PP --- so of the two possible parts of speech to
which use can belong, only the noun is possible in the second sentence ( b).

As we have noted in Chapter , we can give translation engines such information about
grammar, in the form of grammar rules. This is useful in that it allows them to filter out some
wrong analyses. However, giving our system knowledge about syntax will not allow us to
determine the meaning of all ambiguous words. This is because words can have several
meanings even within the same part of speech. Take for example the word button. Like the

http://privatewww.essex.ac.uk/~doug/book/node53.html (1 de 4) [20/09/2000 18:03:41]

Ambiguity

word use, it can be either a verb or a noun. As a noun, it can mean both the familiar small round
object used to fasten clothes, as well as a knob on a piece of apparatus. To get the machine to
pick out the right interpretation we have to give it information about meaning.
In fact, arming a computer with knowledge about syntax, without at the same time telling it
something about meaning can be a dangerous thing. This is because applying a grammar to a
sentence can produce a number of different analyses, depending on how the rules have applied,
and we may end up with a large number of alternative analyses for a single sentence. Now
syntactic ambiguity may coincide with genuine meaning ambiguity, but very often it does not,
and it is the cases where it does not that we want to eliminate by applying knowledge about
meaning.
We can illustrate this with some examples. First, let us show how grammar rules, differently
applied, can produce more than one syntactic analysis for a sentence. One way this can occur is
where a word is assigned to more than one category in the grammar. For example, assume that
the word cleaning is both an adjective and a verb in our grammar. This will allow us to assign
two different analyses to the following sentence.
● fluids can be dangerous.

One of these analyses will have cleaning as a verb, and one will have it as an adjective. In the
former (less plausible) case the sense is `to clean a fluid may be dangerous', i.e. it is about an
activity being dangerous. In the latter case the sense is that fluids used for cleaning can be
dangerous. Choosing between these alternative syntactic analyses requires knowledge about
meaning.
It may be worth noting, in passing, that this ambiguity disappears when can is replaced by a
verb which shows number agreement by having different forms for third person singular and
plural. For example, the following are not ambiguous in this way: ( a) has only the sense that
the action is dangerous, ( b) has only the sense that the fluids are dangerous.
● ❍ Cleaning fluids is dangerous.

❍ Cleaning fluids are dangerous.

We have seen that syntactic analysis is useful in ruling out some wrong analyses, and this is
another such case, since, by checking for agreement of subject and object, it is possible to find
the correct interpretations. A system which ignored such syntactic facts would have to consider
all these examples ambiguous, and would have to find some other way of working out which
sense was intended, running the risk of making the wrong choice. For a system with proper
syntactic analysis, this problem would arise only in the case of verbs like can which do not
show number agreement.
Another source of syntactic ambiguity is where whole phrases, typically prepositional phrases,
can attach to more than one position in a sentence. For example, in the following example, the
prepositional phrase with a Postscript interface can attach either to the NP the word processor
package, meaning ``the word-processor which is fitted or supplied with a Postscript interface'',
or to the verb connect, in which case the sense is that the Postscript interface is to be used to
make the connection.
● the printer to a word processor package with a Postscript interface.

Notice, however, that this example is not genuinely ambiguous at all, knowledge of what a
Postscript interface is (in particular, the fact that it is a piece of software, not a piece of

http://privatewww.essex.ac.uk/~doug/book/node53.html (2 de 4) [20/09/2000 18:03:41]

Ambiguity

hardware that could be used for making a physical connection between a printer to an office
computer) serves to disambiguate. Similar problems arise with ( ), which could mean that the
printer and the word processor both need Postscript interfaces, or that only the word processor
needs them.
● will require a printer and a word processor with Postscript interfaces.

This kind of real world knowledge is also an essential component in disambiguating the
pronoun it in examples such as the following
● the paper in the printer. Then switch it on.

In order to work out that it is the printer that is to be switched on, rather than the paper, one
needs to use the knowledge of the world that printers (and not paper) are the sort of thing one is
likely to switch on.
There are other cases where real world knowledge , though necessary, does not seem to be
sufficient. The following, where two people are re-assembling a printer, seems to be such an
example:
● : Now insert the cartridge at the back.
B: Okay.
A: By the way, did you order more toner today?
B: Yes, I got some when I picked up the new paper.
A: OK, how far have you got?
A: Did you get fixed?
It is not clear that any kind of real world knowledge will be enough to work out that it in the
last sentence refers to the cartridge, rather than the new paper, or toner. All are probably equally
reasonable candidates for fixing. What strongly suggests that it should be interpreted as the
cartridge is the structure of the conversation --- the discussion of the toner and new paper
occurs in a digression, which has ended by the time it occurs. Here what one needs is
knowledge of the way language is used. This is knowledge which is usually thought of as
pragmatic in nature. Analysing the meaning of texts like the above example is important in
dialogue translation, which is a long term goal for MT research, but similar problems occur in
other sorts of text.
Another sort of pragmatic knowledge is involved in cases where the translation of a sentence
depends on the communicative intention of the speaker --- on the sort of action (the speech act)

that the speaker intends to perform with the sentence. For example, ( ) could be a request
for action, or a request for information, and this might make a difference to the translation.
● you reprogram the printer interface on this printer?

In some cases, working out which is intended will depend on the non-linguistic situation, but it
could also depend on the kind of discourse that is going on --- for example, is it a discourse
where requests for action are expected, and is the speaker in a position to make such a request
of the hearer? In dialogues, such pragmatic information about the discourse can be important
for translating the simplest expressions. For example, the right translation of Thank you into
French depends on what sort of speech act it follows. Normally, one would expect the
translation to be merci. However, if it is uttered in response to an offer, the right translation
would be s'il vous plaît (`please').

http://privatewww.essex.ac.uk/~doug/book/node53.html (3 de 4) [20/09/2000 18:03:41]

Ambiguity

Next: Lexical and Structural Up: Translation Problems Previous: Introduction

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node53.html (4 de 4) [20/09/2000 18:03:41]

Lexical and Structural Mismatches

Next: Multiword units: Idioms Up: Translation Problems Previous: Ambiguity

Lexical and Structural Mismatches

At the start of the previous section we said that, in the best of all possible worlds for NLP, every word
would have exactly one sense. While this is true for most NLP, it is an exaggeration as regards MT. It would
be a better world, but not the best of all possible worlds, because we would still be faced with difficult
translation problems. Some of these problems are to do with lexical differences between languages ---
differences in the ways in which languages seem to classify the world, what concepts they choose to express
by single words, and which they choose not to lexicalize. We will look at some of these directly. Other
problems arise because different languages use different structures for the same purpose, and the same
structure for different purposes. In either case, the result is that we have to complicate the translation
process. In this section we will look at some representative examples.

Examples like the ones in ( ) below are familiar to translators, but the examples of colours ( c), and the
Japanese examples in ( d) are particularly striking. The latter because they show how languages need
differ not only with respect to the fineness or `granularity' of the distinctions they make, but also with
respect to the basis for the distinction: English chooses different verbs for the action/event of putting on, and
the action/state of wearing. Japanese does not make this distinction, but differentiates according to the
object that is worn. In the case of English to Japanese, a fairly simple test on the semantics of the NPs that
accompany a verb may be sufficient to decide on the right translation. Some of the colour examples are
similar, but more generally, investigation of colour vocabulary indicates that languages actually carve up the
spectrum in rather different ways, and that deciding on the best translation may require knowledge that goes
well beyond what is in the text, and may even be undecidable. In this sense, the translation of colour
terminology begins to resemble the translation of terms for cultural artifacts (e.g. words like English
cottage, Russian dacha, French château, etc. for which no adequate translation exists, and for which the
human translator must decide between straight borrowing, neologism, and providing an explanation). In this
area, translation is a genuinely creative act , which is well beyond the capacity of current computers.

Calling cases such as those above lexical mismatches is not controversial. However, when one turns to
cases of structural mismatch, classification is not so easy. This is because one may often think that the
reason one language uses one construction, where another uses another is because of the stock of lexical
items the two languages have. Thus, the distinction is to some extent a matter of taste and convenience.
A particularly obvious example of this involves problems arising from what are sometimes called lexical
holes --- that is, cases where one language has to use a phrase to express what another language expresses
in a single word. Examples of this include the `hole' that exists in English with respect to French ignorer
(`to not know', `to be ignorant of'), and se suicider (`to suicide', i.e. `to commit suicide', `to kill oneself').
The problems raised by such lexical holes have a certain similarity to those raised by idiom s: in both cases,
one has phrases translating as single words. We will therefore postpone discussion of these until Section
.
One kind of structural mismatch occurs where two languages use the same construction for different
purposes, or use different constructions for what appears to be the same purpose.
Cases where the same structure is used for different purposes include the use of passive constructions in
English, and Japanese . In the example below, the Japanese particle wa, which we have glossed as `TOP'

http://privatewww.essex.ac.uk/~doug/book/node54.html (1 de 5) [20/09/2000 18:03:46]

Lexical and Structural Mismatches

here marks the `topic' of the sentence --- intuitively, what the sentence is about.
● ❍ Satoo-san wa shyushoo ni erabaremashita.
Satoo-hon TOP Prime Minister in was-elected
❍ Mr. Satoh was elected Prime Minister.

Example ( ) indicates that Japanese has a passive-like construction, i.e. a construction where the
PATIENT, which is normally realized as an OBJECT, is realized as SUBJECT. It is different from the
English passive in the sense that in Japanese this construction tends to have an extra adversive nuance which
might make ( a) rather odd, since it suggests an interpretation where Mr Satoh did not want to be elected,
or where election is somehow bad for him. This is not suggested by the English translation, of course. The
translation problem from Japanese to English is one of those that looks unsolvable for MT, though one
might try to convey the intended sense by adding an adverb such as unfortunately. The translation problem
from English to Japanese is on the other hand within the scope of MT, since one must just choose another
form. This is possible, since Japanese allows SUBJECTs to be omitted freely, so one can say the equivalent
of elected Mr Satoh, and thus avoid having to mention an AGENT . However, in general, the result of
this is that one cannot have simple rules like those described in Chapter for passives. In fact, unless one
uses a very abstract structure indeed, the rules will be rather complicated.
We can see different constructions used for the same effect in cases like the following:
● ❍ He is called Sam.

❍ Er heiß t Sam.
`He is-named Sam'
❍ Il s'appelle Sam.
`He calls himself Sam'
●

❍ Sam has just seen Kim.

❍ Sam vient de voir Kim.
`Sam comes of see Kim'
● ❍ Sam likes to swim.
❍ Sam zwemt graag.
`Sam swims likingly'

Figure: venir-de and have-just

The first example shows how English, German and French choose different methods for expressing
`naming'. The other two examples show one language using an adverbial ADJUNCT ( just, or graag(Dutch)
`likingly' or `with pleasure'), where another uses a verbal construction. This is actually one of the most
discussed problems in current MT, and it is worth examining why it is problematic. This can be seen by
looking at the representations for ( ) in Figure .
These representations are relatively abstract (e.g. the information about tense and aspect conveyed by the
auxiliary verb have has been expressed in a feature) , but they are still rather different. In particular, notice
that while the main verb of ( a) is see, the main verb of ( b) is venir-de. Now notice what is involved in
writing rules which relate these structures (we will look at the direction English French).

1. The adverb just must be translated as the verb venir-de (perhaps this is not the best way to think about
it --- the point is that the French structure must contain venir-de, and just must not be translated in any
other way).

http://privatewww.essex.ac.uk/~doug/book/node54.html (2 de 5) [20/09/2000 18:03:46]

Lexical and Structural Mismatches

2. Sam, the SUBJECT of see, must become the SUBJECT of venir-de.

3. Some information about tense, etc. must be taken from the S node of which see is the HEAD, and put
on the S node of which venir-de is the HEAD. This is a complication, because normally one would
expect such information to go on the node of which the translation of see, voir, is the HEAD.
4. Other parts of the English sentence should go into the corresponding parts of the sentence HEADed
by voir. This is simple enough here, because in both cases Kim is an OBJECT, but it is not always the
case that OBJECTs translate as OBJECTs, of course.
5. The link between the SUBJECT of venir-de and the SUBJECT of voir must be established --- but this
can perhaps be left to French synthesis.

All this is summarized in Figure and Figure .

Figure: Translating have-just into venir-de

Of course, given a complicated enough rule, all this can be stated. However, there will still be problems
because writing a rule in isolation is not enough. One must also consider how the rule interacts with other
rules. For example, there will be a rule somewhere that tells the system how see is to be translated, and what
one should do with its SUBJECT and OBJECT. One must make sure that this rule still works (e.g. its
application is not blocked by the fact that the SUBJECT is dealt with by the special rule above; or that it
does not insert an extra SUBJECT into the translation, which would give * Sam vient de Sam voir Kim).
One must also make sure that the rule works when there are other problematic phenomena around. For
example, one might like to make sure the system produces ( b) as the translation of ( a).

Figure: The Representation of venir-de

● ❍ Sam has probably just seen Kim.

❍ Il est probable que Sam vient de voir Kim.
`It is probable that Sam comes of see Kim'
We said above that everything except the SUBJECT, and some of the tense information goes into the
`lower' sentence in French. But this is clearly not true, since here the translation of probably actually
becomes part of the main sentence, with the translation of ( a) as its COMPLEMENT.
Of course, one could try to argue that the difference between English just and French venir de is only
superficial. The argument could, for example, say that just should be treated as a verb at the semantic level.
However, this is not very plausible. There are other cases where this does not seem possible. Examples like
the following show that where English uses a `manner' verb and a directional adverb/prepositional phrase,
French (and other Romance languages ) use a directional verb and a manner adverb. That is where English
classifies the event described as `running', French classifies it as an `entering':
● ❍ She ran into the room.

❍ Elle entra dans la salle en courant.

`She entered into the room in/while running'
The syntactic structures of these examples are very different, and it is hard to see how one can naturally
reduce them to similar structures without using very abstract representations indeed.
A slightly different sort of structural mismatch occurs where two languages have `the same' construction
(more precisely, similar constructions, with equivalent interpretations), but where different restrictions on
the constructions mean that it is not always possible to translate in the most obvious way. The following is a
relatively simple example of this.

http://privatewww.essex.ac.uk/~doug/book/node54.html (3 de 5) [20/09/2000 18:03:46]

Lexical and Structural Mismatches

● ❍ These are the letters which I have already replied to.

❍ *Ce sont les lettres lesquelles j'ai déjà répondu à.
❍ These are the letters to which I have already replied.
❍ Ce sont les lettres auxquelles j'ai déjà répondu.
What this shows is that English and French differ in that English permits prepositions to be `stranded' (i.e.
to appear without their objects, like in a). French normally requires the preposition and its object to
appear together, as in ( d) --- of course, English allows this too. This will make translating ( a) into
French difficult for many sorts of system (in particular, for systems that try to manage without fairly abstract
syntactic representations). However, the general solution is fairly clear --- what one wants is to build a
structure where ( a) is represented in the same way as ( c), since this will eliminate the translation

problem. The most obvious representation would probably be something along the lines of ( a), or

perhaps ( b).
● ❍ These are the letters I have already replied to which

❍ These are the letters I have already replied

to the letters

While by no means a complete solution to the treatment of relative clause constructions, such an approach
probably overcomes this particular translation problem. There are other cases which pose worse problems,
however.
In general, relative clause constructions in English consist of a head noun ( letters in the previous example),
a relative pronoun (such as which), and a sentence with a `gap' in it. The relative pronoun (and hence the

http://privatewww.essex.ac.uk/~doug/book/node54.html (4 de 5) [20/09/2000 18:03:46]

Lexical and Structural Mismatches

head noun) is understood as if it filled the gap --- this is the idea behind the representations in ( ). In
English, there are restrictions on where the `gap' can occur. In particular, it cannot occur inside an indirect

question, or a `reason' ADJUNCT. Thus, ( b), and ( d) are both ungrammatical. However, these
restrictions are not exactly paralleled in other languages. For example, Italian allows the former, as in (
a), and Japanese the latter, as in ( c). These sorts of problem are beyond the scope of current MT systems
--- in fact, they are difficult even for human translators.
● ❍ Sinda node minna ga kanasinda hito wa yumei desita.
`died hence everyone SUBJ distressed-was man TOP famous was'
❍ *The man who everyone was distressed because (he) died was famous.

❍ L'uomo che mi domando chi abbia visto fu arrestato.

❍ *The man that I wonder who (he) has seen was arrested.

Next: Multiword units: Idioms Up: Translation Problems Previous: Ambiguity

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node54.html (5 de 5) [20/09/2000 18:03:46]

Multiword units: Idioms and Collocations

Next: Summary Up: Translation Problems Previous: Lexical and Structural

Multiword units: Idioms and Collocations

Roughly speaking, idioms are expressions whose meaning cannot be completely understood from the meanings
of the component parts. For example, whereas it is possible to work out the meaning of (

a) on the basis of knowledge of English grammar and the meaning of words, this would not be sufficient to
work out that (

http://privatewww.essex.ac.uk/~doug/book/node55.html (1 de 7) [20/09/2000 18:03:50]

Multiword units: Idioms and Collocations

b) can mean something like `If Sam dies, her children will be rich'. This is because kick the bucket is an idiom.
● ❍ If Sam mends the bucket, her children will be rich.

❍ If Sam kicks the bucket, her children will be rich.

The problem with idioms, in an MT context, is that it is not usually possible to translate them using the normal
rules. There are exceptions, for example take the bull by the horns (meaning `face and tackle a difficulty
without shirking') can be translated literally into French as prendre le taureau par les cornes, which has the
same meaning. But, for the most part, the use of normal rules in order to translate idioms will result in
nonsense. Instead, one has to treat idioms as single units in translation.
In many cases, a natural translation for an idiom will be a single word --- for example, the French word mourir
(`die') is a possible translation for kick the bucket. This brings out the similarity, which we noted above, with
lexical holes of the kind shown in (

http://privatewww.essex.ac.uk/~doug/book/node55.html (2 de 7) [20/09/2000 18:03:50]

Multiword units: Idioms and Collocations

).
● ❍ J'ignore la solution.
❍ I do not know the solution.
❍ se suicider.
❍ commit suicide.
Lexical holes and idioms are frequently instances of word

http://privatewww.essex.ac.uk/~doug/book/node55.html (3 de 7) [20/09/2000 18:03:50]

Multiword units: Idioms and Collocations

phrase translation. The difference is that with lexical hole s, the problem typically arises when one translates
from the language with the word into the language that uses the phrase, whereas with idioms, one usually gets
the problem in translating from the language that has the idiom (i.e. the phrase) into the language which uses a
single word. For example, there is no problem in translating I do not know the solution literally into French ---
the result is perfectly understandable. Similarly, there is no problem in translating mourir `literally' into English
(as die) --- one is not forced to use the idiom kick the bucket.
In general, there are two approaches one can take to the treatment of idioms. The first is to try to represent them
as single units in the monolingual dictionaries. What this means is that one will have lexical entries such as
kick_the_bucket. One might try to construct special morphological rules to produce these representations
before performing any syntactic analysis --- this would amount to treating idioms as a special kind of word,
which just happens to have spaces in it. As will become clear, this is not a workable solution in general. A more
reasonable idea is not to regard lexical lookup as a single process that occurs just once, before any syntactic or
semantic processing, but to allow analysis rules to replace pieces of structure by information which is held in
the lexicon at different stages of processing, just as they are allowed to change structures in other ways. This
would mean that kick the bucket and the non-idiomatic kick the table would be represented alike (apart from the
difference between bucket and table) at one level of analysis, but that at a later, more abstract representation
kick the bucket would be replaced with a single node, with the information at this node coming from the lexical
entry kick_the_bucket. This information would probably be similar to the information one would find in
the entry for die.
In any event, this approach will lead to translation rules saying something like the following, in a transformer
or transfer system (in an interlingual system , idioms will correspond to collections of concepts, or single
concepts in the same way as normal words).
in_fact => en_fait in_view_of => étant_donné kick_the_bucket =>
mourir kick_the_bucket => casser_sa_pipe
The final example shows that one might, in this way, be able to translate the idiom kick the bucket into the
equivalent French idiom casser sa pipe --- literally `break his/her pipe'. The overall translation process is
illustrated in Figure .
The second approach to idioms is to treat them with special rules that change the idiomatic source structure into
an appropriate target structure. This would mean that kick the bucket and kick the table would have similar

http://privatewww.essex.ac.uk/~doug/book/node55.html (4 de 7) [20/09/2000 18:03:50]

Multiword units: Idioms and Collocations

representations all through analysis . Clearly, this approach is only applicable in transfer or transformer
systems, and even here, it is not very different from the first approach --- in the case where an idiom translates
as a single word, it is simply a question of where one carries out the replacement of a structure by a single
lexical item, and whether the item in question is an abstract source language word such as kick_the_bucket or a
normal target language word (such as mourir).

Figure: Dealing with Idioms 1

One problem with sentences which contain idioms is that they are typically ambiguous , in the sense that either
a literal or idiomatic interpretation is generally possible (i.e. the phrase kick the bucket can really be about
buckets and kicking). However, the possibility of having a variety of interpretations does not really distinguish
them from other sorts of expression. Another problem is that they need special rules (such as those above,
perhaps), in addition to the normal rules for ordinary words and constructions. However, in this they are no
different from ordinary words, for which one also needs special rules. The real problem with idioms is that they
are not generally fixed in their form, and that the variation of forms is not limited to variations in inflection (as
it is with ordinary words). Thus, there is a serious problem in recognising idioms.
This problem does not arise with all idioms. Some are completely frozen forms whose parts always appear in
the same form and in the same order. Examples are phrases like in fact, or in view of. However, such idioms are
by far the exception. A typical way in which idioms can vary is in the form of the verb, which changes
according to tense , as well as person and number. For example, with bury the hatchet (`to cease hostilities and
becomes reconciled', one gets He buries/buried/will bury the hatchet, and They bury/buried/shall bury the
hatchet. Notice that variation in the form one gets here is exactly what one would get if no idiomatic
interpretation was involved --- i.e. by and large idioms are syntactically and morphologically regular --- it is
only their interpretations that are surprising.
A second common form of variation is in the form of the possessive pronoun in expressions like to burn one's
bridges (meaning `to proceed in such a way as to eliminate all alternative courses of action'). This varies in a
regular way with the subject of the verb:
● ❍ He has burned bridges.
❍ She has burned bridges.
In other cases, only the syntactic category of an element in an idiom can be predicted. Thus, the idiom pull X's
leg (`tease') contains a genitive NP, such as Sam's, or the king of England's. Another common form of variation
arises because some idioms allow adjectival modifiers. Thus in addition to keep tabs on (meaning observe) one
has keep close tabs on (`observe closely'), or put a political cat among the pigeons (meaning `do or say
something that causes a lot of argument politically'). Some idioms appear in different syntactic configurations,
just like regular non-idiomatic expressions. Thus, bury the hatchet appears in the passive, as well as the active
voice.
● ❍ He buried the hatchet

❍ The hatchet seems to have been buried

Of course, not all idioms allow these variations (e.g. one cannot passivize kick the bucket meaning `die'), and,
as noted, some do not allow any variation in form. But where variation in form is allowed, there is clearly a

http://privatewww.essex.ac.uk/~doug/book/node55.html (5 de 7) [20/09/2000 18:03:50]

Multiword units: Idioms and Collocations

problem. In particular, notice that it will not be possible to recognise idioms simply by looking for sequences of
particular words in the input. Recognising some of these idioms will require a rather detailed syntactic
analysis . For example, despite the variation in form for bury the hatchet, the idiomatic interpretation only
occurs when the hatchet is always DEEP OBJECT of bury. Moreover, the rules that translate idioms or which
replace them by single lexical items may have to be rather complex. Some idea of this can be gained from
considering what must happen to pull Sam's leg in order to produce something like equivalent to tease Sam, or
the French translation involving taquiner (`tease'), cf. Figure . This figure assumes the input and output of
transfer are representations of grammatical relations, but the principles are the same if semantic representations
are involved, or if the process involves reducing pull X's leg to a single word occurs in English analysis.

Figure: Dealing with Idioms 2

Rather different from idioms are expressions like those in (

http://privatewww.essex.ac.uk/~doug/book/node55.html (6 de 7) [20/09/2000 18:03:50]

Multiword units: Idioms and Collocations

), which are usually referred to as collocations . Here the meaning can be guessed from the meanings of the
parts. What is not predictable is the particular words that are used.
● ❍ This butter is rancid (*sour, *rotten, *stale).

❍ This cream is sour (*rancid, *rotten, *stale).

❍ They took (*made) a walk.

❍ They made (*took) an attempt.

❍ They had (*made, *took) a talk.

For example, the fact that we say rancid butter, but not * sour butter, and sour cream, but not * rancid cream
does not seem to be completely predictable from the meaning of butter or cream, and the various adjectives.
Similarly the choice of take as the verb for walk is not simply a matter of the meaning of walk (for example,
one can either make or take a journey).
In what we have called linguistic knowledge (LK) systems, at least, collocations can potentially be treated
differently from idioms. This is because for collocations one can often think of one part of the expression as
being dependent on, and predictable from the other. For example, one may think that make, in make an attempt
has little meaning of its own, and serves merely to `support' the noun (such verbs are often called light verbs,
or support verbs). This suggests one can simply ignore the verb in translation, and have the generation or
synthesis component supply the appropriate verb. For example, in Dutch , this would be doen, since the Dutch
for make an attempt is een poging doen (`do an attempt').
One way of doing this is to have analysis replace the lexical verb (e.g. make) with a `dummy verb' (e.g.
VSUP). This can be treated as a sort of interlingual lexical item, and replaced by the appropriate verb in
synthesis (the identity of the appropriate verb has to be included in the lexical entry of nouns, of course --- for
example, the entry for poging might include the feature support_verb=doen. The advantage is that
support verb constructions can be handled without recourse to the sort of rules required for idioms (one also
avoids having rules that appear to translate make into poging `do').
Of course, what one is doing here is simply recording, in each lexical entry, the identity of the words that are
associated with it, for various purposes --- e.g. the fact that the verb that goes with attempt is make (for some
purposes, anyway). An interesting generalisation of this is found in the idea of lexical functions . Lexical
functions express a relation between two words. Take the case of heavy smoker, for example. The relationship
between heavy and smoker is that of intensification, which could be expressed by the lexical function Magn as
follows, indicating that the appropriate adjective for English smoker is heavy, whereas that for the
corresponding French word fumeur is grand (`large') and that for the German word Raucher is stark (`strong').
(English) Magn(smoker) = heavy (French) Magn(fumeur) = grand
(German) Magn(Raucher) = stark
If one wants to translate heavy smoker into French , one needs to map smoker into fumeur, together with the
information that fumeur has the lexical function Magn applied to it, as in English. It would be left to the
French synthesis module to work out that the value Magn(fumeur) = grand, and insert this adjective
appropriately. Translation into German is done in the same way.

Next: Summary Up: Translation Problems Previous: Lexical and Structural

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node55.html (7 de 7) [20/09/2000 18:03:50]

Summary

Next: Further Reading Up: Translation Problems Previous: Multiword units: Idioms

Summary
This chapter looks at some problems which face the builder of MT systems. We characterized
them as problems of ambiguity (lexical and syntactic) and problems of lexical and structural
mismatches. We saw how different types of linguistic and non-linguistic knowledge are
necessary to resolve problems of ambiguity, and in the next chapter we examine in more detail
how to represent this knowledge. In this chapter we discussed instances of lexical and structural
mismatches and the problem of non-compositionality (as exemplified by idioms and
collocation s) and looked at some strategies for dealing with them in MT systems.

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node56.html [20/09/2000 18:03:51]

Further Reading

Next: Representation and Processing Up: Translation Problems Previous: Summary

Further Reading
The problem of ambiguity is pervasive in NLP, and is discussed extensively in the
introductions to the subject such as those mentioned in the Further Reading section of Chapter
.
Examples of lexical and structural mismatches are discussed in [Hutchins and Somers1992,
Chapter 6,]. Problems of the venir-de/ have just sort are discussed extensively in the MT
literature. A detailed discussion of the problem can be found in [Arnold et al. 1988], and in
[Sadler1993]. On light verbs or support verbs, see [Danlos and Samvelian1992,Danlos1992].

Treatments of idioms in MT are given in [Arnold and Sadler1989], and [Schenk1986]. On

collocations , see for example [Allerton1984], [Benson et al. 1986a], [Benson et al. 1986b]
and [Hanks and Church1989]. The notion of lexical functions is due to Mel'cuk, see for
example [Mel'cuk and Polguere1987,Mel'cuk and Zholkovsky1988].

A classic discussion of translation problems is [Vinay and Darbelnet1977]. This is concerned

with translation problems as faced by humans, rather than machines, but it points out several of
the problems mentioned here.
The discussion in this chapter touches on two issues of general linguistic and philosophical
interest: to what extent human languages really do carve the world up differently, and whether
there are some sentences in some languages which cannot be translated into other languages. As
regards the first question, it seems as though there are some limits. For example, though
languages carve the colour spectrum up rather differently, so there can be rather large
differences between colour words in terms of their extensions, there seems to be a high level of
agreement about `best instances'. That is, though the extension of English red, and Japanese
akai is different, nevertheless, the colour which is regarded as the best instance of red by
English speakers is the colour which is regarded as the best instance of akai by Japanese
speakers. The seminal work on this topic is [Berlin and Kay1969], and see the title essay of
[Pullum1991]. The second question is sometimes referred to as the question of effability, see
[Katz1978,Keenan1978] for relevant discussion.

Next: Representation and Processing Up: Translation Problems Previous: Summary

http://privatewww.essex.ac.uk/~doug/book/node57.html (1 de 2) [20/09/2000 18:03:52]

Further Reading

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node57.html (2 de 2) [20/09/2000 18:03:52]

Representation and Processing Revisited: Meaning

Next: Introduction Up: Machine Translation Previous: Further Reading

Representation and Processing

Revisited: Meaning

● Introduction
● Semantics
● Pragmatics
● Real World Knowledge
● Summary
● Further Reading

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node58.html [20/09/2000 18:03:52]

Introduction

Next: Semantics Up: Representation and Processing Previous: Representation and Processing

Introduction
The discussion in previous chapters reinforces the point made in Chapter about the value of
syntactic, and `shallow' semantic analysis, but it also shows why performing a syntactic
analysis alone is not sufficient for translation. As the discussion in Chapter indicates, there
are many cases where problems seem to require deeper, more meaning oriented representations,
and enrichment of the kind of knowledge systems are equipped with. In this chapter we will try
to give a flavour of what is involved in this.
It is useful to think of this knowledge as being of three kinds: (i) linguistic knowledge which is
independent of context , semantic knowledge; (ii) linguistic knowledge which relates to the
context (e.g. of earlier utterances), sometimes called pragmatic knowledge; and (iii) common
sense , general, non-linguistic knowledge about the real world , which we will call real world
knowledge. It should be stressed that the distinction between these different kinds of
knowledge is not always clear, and there are those who would dispute whether the distinction is
real. However, it is at least a convenient subdivision of the field, and we will examine each sort
of knowledge in turn, in Sections , , and . Discussing these different kinds of
knowledge will also allow us to describe some more general translation problems.
Apart from giving an overview and flavour of what is involved, the point we would like to
stress in this chapter is that though dealing with meaning in a general way poses many unsolved
problems, and in general one should not expect to find much in the way of real world ,
pragmatic , or even semantic processing in current commercial MT systems, such processing it
is not totally beyond the reach of current theory.

Next: Semantics Up: Representation and Processing Previous: Representation and Processing

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node59.html [20/09/2000 18:03:53]

Semantics

Next: Pragmatics Up: Representation and Processing Previous: Introduction

Semantics

Semantics is concerned with the meaning of words and how they combine to form sentence
meanings. It is useful to distinguish lexical semantics , and structural semantics --- the former
is to do with the meanings of words, the latter to do with the meanings of phrases, including
sentences. We will begin with the former.
There are many ways of thinking about and representing word meanings, but one that has
proved useful in the field of machine translation involves associating words with semantic
features which correspond to their sense components. For example, the words man, woman,
boy, and girl might be represented as:
man = (+HUMAN, +MASCULINE and +ADULT) woman = (+HUMAN,
-MASCULINE and +ADULT) boy = (+HUMAN, +MASCULINE and
-ADULT) girl = (+HUMAN, -MASCULINE and -ADULT)
Associating words with semantic features is useful because some words impose semantic
constraints on what other kinds of words they can occur with. For example, the verb eat
demands that its AGENT (the eater) is animate and that its PATIENT (that which is eaten) is
edible, --- concrete (rather than abstract, like sincerity, or beauty), and solid (rather than liquid,
so one cannot `eat' beer, coffee, etc.; soup is a borderline case). We can encode this constraint
in our grammar by associating the features HUMAN and EDIBLE with appropriate nouns in
our dictionary and describing our entry for eat as something like cat=verb,
AGENT=HUMAN, PATIENT=EDIBLE. The grammar will now only accept objects of eat that
have the feature EDIBLE. Thus these selectional restrictions , as they are called, act as a filter
on our grammar to rule out unwanted analyses. Consider sentence ( ):
● ate the game.

The English word game is ambiguous - it can mean several things, including a form of play or
sport or a wild animal hunted or fished for food. Using selectional restrictions of the sort
described above we can eliminate the `form of play or sport' meaning if the system is able to
infer that `food' is EDIBLE, but that forms of play are not.
Selectional restrictions have proved a very useful device and are found in most MT systems to
a greater or lesser extent. Unfortunately, however, exceptions to selectional restrictions abound,
especially in metaphorical speech . Thus we find sentences like This car eats money, used to
mean that the car is expensive to maintain, so, rather than use selectional restrictions to
eliminate interpretations, we should use them to state preferences between alternative
interpretations.
Notice that stating selectional restrictions in terms of semantic relations is easier than trying to
state them in terms of (surface) grammatical relations . Using grammatical relations we would

http://privatewww.essex.ac.uk/~doug/book/node60.html (1 de 6) [20/09/2000 18:03:57]

Semantics

have to say that eat prefers an animate SUBJECT in active sentences, and an animate NP in the
by phrase in passive sentences (and an edible OBJECT in actives, and an edible SUBJECT in
passives).
We will now look briefly at how semantic relations can help in one of the thorniest problems
for machine translation, namely the translation of prepositions.
Take, for example, the translation of the English preposition at into Spanish , and, for the sake
of exposition, make the simplifying assumption that it receives only two translations in
Spanish , namely a and en, as in the following:
● ❍ at midday

❍ mediodía
● ❍ at school

❍ la escuela
The choice of Spanish preposition depends on the type of noun that follows it. Roughly, where
the preposition is followed by a temporal noun, as in the first example, it translates as a, but
where the preposition is followed by a locational noun, as in the second example, it translates as
en.
We can pick out the correct translation of at by assigning it an appropriate Semantic Relation
(SR) during analysis. For example, the feature SR=TIME might be assigned to indicate that at
expresses a temporal relation, and the feature SR=PLACE might be used to mean that at
expresses a location relation. We could then have translation rules of the following form:

at, SR=TIME a at, SR=PLACE en

These semantic relations are assigned on the basis of the type of noun that follows the
preposition. This means that the noun midday must be marked in the dictionary with some
temporal feature (e.g. semtype=time), while nouns like school must be marked with some
locational feature (e.g. semtype=location).
We are assuming that semantic relations attach to prepositions. More properly, a semantic
relation describes the role which the whole prepositional phrase, not just the preposition, plays
in relation to its head, but it is convenient to allow the preposition to carry this feature too, in
order to formulate the above translation rules. A prepositional phrase marked with the semantic
relation TIME, for example, might indicate the time at which the action indicated by the verb
takes place, while a phrase marked with the semantic relation PLACE might indicate the
location at which it took place.
Although these features would solve many problems in translating prepositions, the semantic
relations expressed by PLACE and TIME are not always fine grained enough. We can, for
example, distinguish two different types of usage for locational at: `(to be) at school' indicates a
position, whereas `(to shoot) at the goal' indicates a movement towards a certain place. We
could decompose the semantic relation into two separate relations, say PLACE_POSITION for
the first phrase, and PLACE_PATH for the second phrase. Note that the calculation of these
new semantic relations will depend not only on the semantic features of the nouns that follow
them, but crucially on the type of verb.

http://privatewww.essex.ac.uk/~doug/book/node60.html (2 de 6) [20/09/2000 18:03:57]

Semantics

Our brief example illustrates some of the problems we face when trying to assign semantic
relations to prepositional phrases, or other categories. First, it is difficult to know what a
canonical set of semantic relations might look like, since the refinement or granularity required
(that is, the number of distinctions we want to make) depends to some extent on the type of
translation problem encountered. Secondly, the finer the granularity, the more elaborate the
feature system will have to be, in order to differentiate nouns, for example. Finally, the
calculation of semantic relations depends on a number of factors, including as we have seen the
type of verb and the type of the following noun.
We have described semantic features as more or less optional additions to representations ---
the addition of a semantic feature may serve to disambiguate a representation, by indicating
which sense of a word is involved, but the representation is still conceived of as a structure
consisting of lexical items (words). A more radical idea is to take the semantic features as
exhausting the meaning of words, and to replace the lexical items by the appropriate set of
features. Thus, one would have representations with (+HUMAN, +MASCULINE, +ADULT,
...) in place of the lexical item man. The idea is that the meanings of lexical items can be
decomposed into sets of semantic primitive s. Since such sets of semantic primitive s might
well be universal, one can in this way approach the goal of an interlingua . Here one cannot
manage satisfactorily simply with sets of features, however. Instead, one needs to produce
structures in which the predicates are semantic primitive s. For example, the representation of
kill might be along the following lines:
● CAUSE [ BECOME [ NOT [ ALIVE ]

As we have already noted in Chapter there are some doubts in general about the feasibility
and advisability of this process of lexical decomposition . For example, there is a small but
significant difference in meaning between kill and cause to become not alive --- in particular,
where a `killing' is a single event, a `causing to become not alive' involves at least two events (a
`causing', and a `dying'), and if the causal chain that links a particular event to dying is long
enough, one may admit that the event caused the dying, but not want to say there has been a
`killing'. Of course, these doubts depend on what one thinks the relation is between the
semantic primitive s like CAUSE, BECOME, etc., and English words like cause, become, etc.,
and also on the assumption that there is no semantic primitive KILL. Notice that, while a
collection of semantic primitives that includes KILL is going to be quite large (perhaps in the
order of a thousand primitives), this is still far less than the vocabulary one finds in normal use
--- so there may still be some value in semantic decomposition , even if the number of
primitive s that words decompose into is quite large.
So far we have concentrated our discussion of semantics on the meaning of words, but
semantics is also concerned with linguistic `systems' such as tense and aspect and
determination, all of which are of considerable importance in translation. Consider the problem
of how to translate the present tense in French into English, where there are at least three
possibilities, exemplified in the following:
● ❍ Elle vit à Londres.

❍ She lives in London.

● ❍ Elle vit à Londres depuis le mois dernier.

❍ She has lived in London since last month.

● ❍ Elle mange son dîner.

http://privatewww.essex.ac.uk/~doug/book/node60.html (3 de 6) [20/09/2000 18:03:57]

Semantics

❍ She is eating her dinner.

Of course, one could try to formulate rules which describe the conditions under which French
present tense is realized as English present, English present perfect, or present progressive, but
such rules would be very complex. A more attractive possibility is to try to find some more
abstract representation which directly describes the temporal and aspectual relations that these
sentences involve. Here we will outline one type of approach.
The English tense system is used to convey two different types of information. One is the time
of the event --- both the present simple I sing and the present progressive I am singing describe
an event in the present. The other is the nature of the event --- e.g. the progressive stresses that
the event is `in progress'. Henceforth we shall reserve the word tense to mean the time of an
event and use the word aspect to refer to the way the event is viewed (as an on-going or
completed process, a state, or a simple event, etc.). We will use the term time reference to
cover both tense and aspect.
We can think of tense as expressing a relation between the time of the event and the time of
speech. Thus, with the present ( I sing), the time of the event (which we could call E) overlaps
with the time of speech (which we could call S). Contrast the future ( I shall sing) where the
time of the event follows the time of speech ( E follows S), or the past, where E precedes S.
However, this is not sufficient to distinguish all the different temporal forms of the English
verb. There is a problem with the past, where our definition of tense does not allow us to
differentiate between the simple past ( I sang) and the pluperfect (or past-perfect --- I had
sung), since in both cases the time of the event is prior to the time of speech. One solution is to
define an additional point of time, called the reference time (R). Consider, for example, the
sentence:
● two o'clock Sam had already eaten.

At two o'clock specifies a moment in time which precedes the time of speech, but which is not
the time of event. Two o'clock is not the time at which John ate, but the time by which he had
already eaten. The temporal relations of this sentence can be expressed as follows, where <
means `precedes':
E < R, R < S
This indicates that the time of the event ( E) precedes the reference time ( R), and R precedes
the time of speech ( S).
We can now distinguish the pluperfect from the simple past by stipulating that in both cases the
time of the event precedes the time of speech ( E < S), but while in the pluperfect the time of
the event precedes the reference time (E < R), in the simple past the time of event and the
reference time coincide ( E = R).
We can do something similar to distinguish the present perfect ( ) from the other tenses. Here
too the event described precedes the speech time, but there is a sense in which sentences in the

present perfect are `about' the present (for example, ( ) would be appropriate only if
Sam's previous eating habits are still of current relevance). We can capture this by making
reference time and speech time coincide ( R=S).
● has eaten snails.

This gives the following picture:

http://privatewww.essex.ac.uk/~doug/book/node60.html (4 de 6) [20/09/2000 18:03:57]

Semantics

● ❍

We now have the apparatus to represent the difference in tense and aspect between the
examples above. Of course, having a way of representing tense and aspect values as above is
one thing, calculating the representations for particular inputs is another. This is no trivial task,
since the tense and aspect values of the verb will in general depend on many factors, including
the form of the verb, and whether it is modified by any time adverbials such as yesterday and
tomorrow.
However, let us assume that we have calculated the tense and aspect values of the following
sentence, and see how this helps translation.
● ❍ Elle vit à Londres depuis le mois dernier.

This sentence might receive a semantic representation along the lines of Figure . The feature
time-ref encodes the information about tense and aspect, in particular, the fact that the
reference time coincides with the time of speech, and the event time precedes the reference time
(and hence also the time of speech).

Figure: Representation Indicating Time Values after French Analysis

Since the information encoded by the time-ref feature is presumed to be preserved in
translation, this feature can treated as an interlingual feature, and thus can be mapped
unchanged onto the target language (in this case English), giving the representation in Figure
.

http://privatewww.essex.ac.uk/~doug/book/node60.html (5 de 6) [20/09/2000 18:03:57]

Semantics

Figure: Representation after Transfer but before English Synthesis

The verb form has lived can then be generated from this representation by English synthesis,

giving the translation ( ). Other time-ref values would be realized differently --- in
principle, the correct translations of the examples above can be obtained.
● has lived in London since last month.

This treatment of tense and aspect involves a lot of complicated machinery, and is not entirely
unproblematic. Nevertheless it gives some indication of how one might attempt to handle the
difficult problem of tense and aspect in MT.

Next: Pragmatics Up: Representation and Processing Previous: Introduction

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node60.html (6 de 6) [20/09/2000 18:03:57]

Pragmatics

Next: Real World Knowledge Up: Representation and Processing Previous: Semantics

Pragmatics

Recall that we made a distinction between semantics , or context-independent meaning , and pragmatics, or
context-dependent meaning. The term `context' is used ambiguously, to refer to the rest of the text in which
a sentence occurs (sometimes referred to as the discourse ), and to circumstances external to the text itself,
such as who the author of the text is, and the social setting in which it occurs, which also contribute to its
interpretation.
To see why the discourse is important, let us consider the translation of anaphoric pronouns . Anaphoric
pronouns are those which refer back to some antecedent earlier in the text, as the pronoun it in (

) refers back to its antecedent the cake.

● took the cake from the table. Then he ate it.

Take the translation of ( ) from English into French . We know that it must refer back to some singular
noun in the previous text or discourse . It has been shown that it is very often the case that the antecedent of
a pronoun is in the same sentence or in the immediately preceding sentence. Assuming that these are the first
sentences in our text, then it can potentially refer back to one of three NPs, namely Sam, the cake or the
table. The syntactic facts of English constrain the pronoun to agree in number and gender with its
antecedent, so it being a neuter pronoun cannot possibly refer to Sam, which is either masculine or feminine.
That leaves us with the choice of either cake or table. One might wonder at this stage whether we need to
decide between the two at all, or whether we can preserve the ambiguity of it in translation. It turns out that
French , like English, requires a pronoun to agree in number and gender with its antecedent. However, since
cake translates as the masculine noun gâteau in French and table as the feminine noun table, this means that
we do have to decide which noun the pronoun it refers back to, in order to translate it either as le (where it
would be interpreted as referring to le gâteau --- cake) or as la (where it would refer back to la table in the
translation of the first sentence). In the above example we can use selectional restrictions on the type of
object that eat can have (namely `edible' objects) to exclude, or at least `disprefer', table as an antecedent for
it. This leaves cake as the best candidate. Providing rules which allow this sort of process to be performed
automatically is not too difficult, but unfortunately resolving pronoun reference is not generally that simple.
First of all, let us consider cases where the pronoun antecedent is not in the current or preceding sentence.
An example might be the following dialogue between two speakers A and B, which appeared in Chapter .
● ❍ A: Now insert the cartridge at the back.

❍ B: Okay.

❍ A: By the way, did you order more toner today?

❍ B: Yes, I got some when I picked up the new paper.

❍ A: OK, how far have you got?

❍ A: Did you get fixed?

It in the last sentence of ( ) refers to the cartridge, although the cartridge was last mentioned in the first

http://privatewww.essex.ac.uk/~doug/book/node61.html (1 de 4) [20/09/2000 18:04:00]

Pragmatics

sentence. Looking for the pronoun's antecedent in the present or preceding sentence this time will not get us
the right result. To find the antecedent, we need to think of the previous discourse not as an unstructured
whole, or a simple sequence of sentences, but rather as a series of `segments', where a segment is a stretch of
discourse in which the (not necessarily adjacent) sentences address the same topic. Cue phrases such as by
the way, and next provide clues to where one segment ends and another one begins. We then constrain the
referent of an anaphor to belong to the same discourse segment as the anaphor.
In the example (

), there are three obvious referents for it: the cartridge ( a), toner ( c), and paper ( d). However,
sentences ( c) and ( d) which form a digression, that is, a discourse segment with a topic (namely
toner) distinct from the main discourse (and whose purpose is not directly related to the purpose of the main
discourse --- in this case the purpose of reassembling the printer). The start of the new segment is signalled
by by the way and the resumption of the old segment is signalled by OK. It is for this reason that the
expressions toner and new paper cannot provide referents for it. In fact, once discourse structure is taken
into account, it can be seen that the cartridge is the only possible antecedent, because it is the only possible
antecedent which is in the same discourse segment as the anaphor .

http://privatewww.essex.ac.uk/~doug/book/node61.html (2 de 4) [20/09/2000 18:04:00]

Pragmatics

Figure: Discourse Structure

Faced with two competing candidates for pronominal reference in a segment, there is another fact about
discourse that we can exploit to get at their resolution, and this is the notion of focus . At any time in a
discourse segment there is an object which is the prime candidate for pronominal reference, and this
element is called the focus . Different suggestions have been made as to how to identify the focus. Often,
there are syntactic signals. For example, in the following example, the focus is much more likely to be Kim,
than Sam, and Kim is more likely to be the antecedent of a pronoun in the following sentence.

● was Kim who Sam telephoned. was in the bath.

The focus of a sentence is also often the NP that has the THEME role in the previous sentence (the THEME
role includes what we have been calling the PATIENT role, but is slightly more general). This is the case
with Kim in ( ), which reinforces the structural cue. But even in the following sequence, where there are
no clear structural clues, key is the THEME and hence most likely to be the focus of the first sentence (and
therefore key is preferred to doormat as the referent of it in the second sentence).
● put the key under the doormat.

When she came home, she found that had been

stolen.
Thus, information about discourse structure is of some help in the resolution of pronoun-antecedent
relations. However, employing knowledge of discourse alone will not enable us to resolve the reference of
all pronouns , as we shall see below.
Let us first look briefly at the other side of pragmatics we mentioned, the context of use. It is obvious that
the identity of the speaker/writer and the addressee will affect the translation of indexical expressions such
as I and you since some languages make a distinction, for instance between you (singular) and you (plural).
Similarly, in languages where an adjective agrees in gender with its noun (as in French , for example), it
will be necessary to know not only the number of the speakers and the addressees, but also their gender in
translating an example like Are you happy?. In addition, knowing the relationship between the addresser and
addressee can be important for translation. The degree of formality between them will affect, for example,
the choice of either vous (formal) or tu (informal) as the translation of you when translating from English

http://privatewww.essex.ac.uk/~doug/book/node61.html (3 de 4) [20/09/2000 18:04:00]

Pragmatics

into French . In many languages, including Japanese , the social relation of speaker and hearer can determine
the form of verb, and even the choice of verb. There are, for example, different verbs for giving as from a
superior to an inferior, and for giving as an inferior to a superior.
We have said that a sentence has to be interpreted relative to both the previous discourse and to the situation
in which it is uttered. In addition, it seems that the meaning of a message is shaped by its producer's
intentions and beliefs. For example, how we interpret ( ) depends on whether the speaker intended it
as a command (to close the front cover), or as an statement (describing the state the cover is likely to be in).
● front cover should be closed.

Of course, the interpretation also depends on the hearer inferring correctly what the speaker's intentions are.
Whether the above sentence is interpreted as a command or statement will affect its translation in some
languages.

Next: Real World Knowledge Up: Representation and Processing Previous: Semantics

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node61.html (4 de 4) [20/09/2000 18:04:00]

Real World Knowledge

Next: Summary Up: Representation and Processing Previous: Pragmatics

Real World Knowledge

The above discussion may lead one to suspect that all the knowledge we need to extract the
meaning from texts and translate them can be got from the texts or their context s. This is,
however, clearly not the case, as the following classic examples show:
●

❍ Little Johnny was very upset. He had lost his toy train. Then he found it. It was in
his .
❍ I saw the soldiers aim at the women, and I saw several of fall.
❍ The council refused the women a permit because advocated violence.
❍ Sue went to put the key under the doormat. When she lifted up, a cockroach
quickly scampered across the path.
In the first example, the problem is the interpretation of pen --- it must be playpen, not writing
pen, because (roughly) for A to be in B, A must be smaller than B, and toy trains are smaller
than playpens, but not writing pens. In the second example, the question is who fell over ---
soldiers or women? In general, we reason that `aiming at' is often followed by firing at, and that
firing at is usually followed by those aimed at falling over, and only rarely followed by those
who do the aiming falling over. In the third case, most people understand that it is the women
who advocate violence --- this seems a normal enough ground for refusing a permit (of course,
it could be that the council advocated violence, and refused the women a permit so as to enrage
them, and incite them to violence). In the case of ( d), we exploit the fact that cockroaches are
more likely to hide under doormats than under keys to work out the most likely interpretation of
it.
In order to translate these examples one will often have to decide what the pronouns refer to,
because many languages use different forms, depending on properties of the antecedent. For
example, translating ( d) into German involves deciding what it refers to, since the possible
candidates are the key or the doormat, which have different gender s in German, which the
pronoun reflects. Similar issues are involved in translating ( b,c). The knowledge that is
deployed here appears to be non-linguistic knowledge, and the reasoning is more or less
`common sense' , perhaps with some small amount of specialist knowledge of the subject
matter. This is perhaps less obvious in the first case, where one may think that the meaning of
in is central, but it is surely clear for the others --- it is nothing to do with the meaning of aim at
that it is often followed by those aimed at falling over. However, even in the playpen -- writing
pen case, we can surely imagine a bizarre situation where little Johnny's playpen is in fact tiny,
and he has just been given a large fountain pen as a present. In such a situation, the
interpretation would be changed, but not because the meaning of the words had changed.

http://privatewww.essex.ac.uk/~doug/book/node62.html (1 de 5) [20/09/2000 18:04:03]

Real World Knowledge

The real world knowledge that is involved here includes common sense reasoning, as well as
general knowledge, and facts about certain more specialized domain s. Representing and
manipulating such knowledge automatically is one of the outstanding research questions of our
time, and the raison d' être of an entire discipline (Artificial Intelligence, AI). The problems of
representing and manipulating linguistic knowledge pale into insignificance compared to the
problems posed by real world knowledge.
One of the problems it raises is that (unlike most linguistic knowledge, in particular, most
knowledge of syntax and semantics ) such knowledge is generally `defeasible', that is, subject
to revision, and not guaranteed correct -- humans have little trouble assuming one thing
most of the time, but managing with a contradictory assumption on occasions (as in the small
playpen example above). This is extremely difficult to automate. A second problem is the huge
amount of such knowledge we seem to have (knowledge about the relative sizes of almost
everything, for example). However, there are some methods of representation that are useful for
some kinds of knowledge.
One particularly useful representation is the so called Semantic Net which can be used for
representing `is a' relations (such as `a dog is a mammal'). Figure gives a small part of such
a network.

Figure: A Fragment of a Semantic Net

Intuitively, the nodes in such a network stand for things, and the links between them are
relations. This means that it can easily be generalized for other sorts of relations. For example,
adding other objects, and using a `part of' relation, one could represent the fact that (say) a
printer is made up of various components, and the fact that these are in turn made up of
subcomponents, etc. Such information might be important in understanding sentences like the
following:
● Put the toner in the cartridge in the reservoir.

Knowing that the reservoir does not have a cartridge as a part would allow one to work out that
this is an instruction to put the toner which is in the cartridge in the reservoir, rather than an
instruction to put the toner in a particular cartridge (i.e. the one that is in the reservoir).
An alternative approach to general knowledge representation is to attempt to formulate it as
collections of `facts' and `rules'. Examples of facts might be the following, which indicate
individuals' departments:
dept(jones,sales). dept(brown,sales).
dept(smith,personnel). ...
The following rule might be used to indicate that two individuals are colleagues, if they work
for the same department (`A and B are colleagues if A works in department D, and B works in
department D'):
colleagues(A,B) <- dept(A,D), dept(B,D).
One problem with both the semantic net , and the `facts and rules' representations are that they
are both rather `small', or loosely organized collections of knowledge. This is not how at least

http://privatewww.essex.ac.uk/~doug/book/node62.html (2 de 5) [20/09/2000 18:04:03]

Real World Knowledge

some kinds of human knowledge seem to be. For example, what the reader knows about her
own home is probably not spread around as sets of unconnected facts. In some way, it seems to
be organized into a coherent, structured whole. (One way of seeing this is by describing your
home to someone -- what you will probably do is take them on a sort of mental tour, which
closely mirrors the physical organization of your home). Similarly, for many practical purposes,
such as eating in restaurants, one does not seem to have a collection of facts and rules, but a
structured `script' of things that typically happen. A great deal of effort has been devoted to the
issue of just what the right kinds of structure are for knowledge representation . The generic
name for such larger knowledge structures is frames. We will give an example of such a
representation in Chapter , but we will not pursue the idea here, because to a great extent
these larger knowledge structures can be built out of smaller ones, such as the ones we have
described.
We now have a way of representing at least some real world knowledge . The question is, how
can it be manipulated? This is a complex and not particularly well-understood matter, and we
will give only the barest outline. However, two points should be emphasised: (a) that as a
whole, the general problem of manipulating knowledge of the world in anything like the way
humans do is unsolved, and may even prove to be insoluble (this is something of a
philosophical question); but (b) under some restricted circumstances, something useful can be
done. The kind of restricted circumstances we have in mind are where there are relatively few
things to think about, and the ways they are related and can be organized and interact are very
limited. An example of this sort might be the internal workings of a printer -- it is possible to
list all the `things' (the printer parts), their relations, and relevant properties (cf. again Chapter
).
One thing that manipulating this knowledge means is using it to answer questions, and draw
inference s. For example, given that one knows that Smith works in the Finance Department,
and Jones works in the Finance Department, how can one work out that Smith and Jones are
colleagues? Given that Tweety is a bird, and that birds have wings, how can one work out that
Tweety has wings? Of course, given the representations above, these questions are not so hard
to answer. In the first case we have provided a rule, the only problem is to find the rule, and
follow it. In the other case, we have exemplified a datastructure (a semantic net), the only
problem is to define a procedure that allows one to use it.
In the first case, one could proceed as follows. In order to answer the question of whether
Brown and Jones are colleagues, one should look for appropriate facts and rules. Assuming
there are no appropriate facts, we have only the rule given above. This tells us that A and B are
colleagues if A works in department D, and B works in department D. We can treat these two
conditions as fresh questions, and answer them in the same way, except that now we have
relevant facts, which will tell us that Brown works in sales, and Jones works in sales. We have
now answered all the subsidiary questions in the affirmative. It follows that we have also
answered the initial question.
In the case of the semantic net s, we might define a procedure that answers questions in the
following way: to answer the question of whether an object has a property, first look to see if
the property is linked to the object by a HAS link. If it does, answer `yes'. If it does not, inspect
each of the IS-A links that end at the object, asking the same question at each one. Thus, though
it is not indicated that Tweety HAS wings, because Tweety IS-A bird, and bird HAS wings, we
can infer that Tweety HAS wings, and answer questions about whether Tweety has wings.

http://privatewww.essex.ac.uk/~doug/book/node62.html (3 de 5) [20/09/2000 18:04:03]

Real World Knowledge

This is a somewhat vague description. However, one may be able to see that some things are
possible, but also that this approach to representing and manipulating real world knowledge is
insufficient. These are some of the things that are lacking.
1. We have not provided a way of handling defeasible rules, or dealing with vague or
`fuzzy' predicates such as tall, hot, etc. For example, penguins are birds, but cannot fly.
Working on the principles just described, one would expect a system to assume that they
could fly. The rules we have given are interpreted as general or universal --- in fact, they
should only be interpreted as indicating defaults. Though there are some partial
techniques for dealing with this, how best to automate default reasoning remains an open
research question. Similarly, the categories we have mentioned in the discussion are
generally rather clear, in the sense that whether something is a bird, or a mammal seems
to be a question that can have a clear, yes or no answer. This is not the case with vague
predicates like hot, or tall. In these cases, not only is there usually some idea of a
standard of comparison (``Hot compared to what?''), which must be inferred from the
context , in some way, but the question of whether something is hot is one that often
lacks a clear answer --- rather than yes, or no, one may be inclined to answer a question
like Is it hot?, with a reply like `a little', or `somewhat'. Again, though there are some
interesting theories, it is mainly an open research question how to model the sort of
reasoning with fuzzy categories that humans can perform.
2. We have suggested how one can answer questions, once they are posed --- but not how
one can reason `forwards' independent of particular questions. For example, if someone
says The printer is broken, hearers may typically draw a whole range of conclusions
(such as ``I will not be able to print the next chapter'', or ``We will have to call an
engineer''), without particular questions being posed. The problem here is that while the
range of inference s drawn is large, it is not as large as it could be (it could be infinite,
since every conclusion will typically lead to new chains of inference s being started), and
it is not clear how to control this process.
3. We have not given any indication of how one would solve the actual problems raised by
the examples in ( ). One could, of course, simply record information about the relative
sizes of known objects as facts, and in the same way associate with other classes of
objects default sizes (e.g. sparrows are typically less than 10cms tall), but this does not
look very plausible as a model of how humans represent this knowledge.
4. We have not said anything about how one might reasonably set about encoding all the
knowledge that seems to be needed, even assuming that one had the `right' format. The
problem is that we cannot anticipate just what particular pieces of real world knowledge
a system may need in general. The amount of knowledge that human writers assume, and
readers supply without apparent effort or reflection is simply vast, and highly
unpredictable, and the effort involved in actually encoding it in this sort of way is
prohibitive. Far more feasible is the aim of equipping a computer with facts about a
specific domain . As we will describe in Chapter , some advanced so-called
Knowledge-Based systems are attempting to do just that.

Next: Summary Up: Representation and Processing Previous: Pragmatics

http://privatewww.essex.ac.uk/~doug/book/node62.html (4 de 5) [20/09/2000 18:04:03]

Real World Knowledge

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node62.html (5 de 5) [20/09/2000 18:04:03]

Summary

Next: Further Reading Up: Representation and Processing Previous: Real World Knowledge

Summary
In this chapter we have looked at three kinds of knowledge that seem to be involved in solving
some sorts of translation problems, namely: semantic , pragmatic , and real world knowledge.
Particular problems we have looked at include the translation of prepositions, of tense and
aspect, and of pronouns. As we stated at the beginning of the chapter, the point to stress as
regards such knowledge is that its representation and manipulation pose many unsolved
problems, and one should not expect to find techniques which exploit it in existing commercial
MT systems (it follows that, for the most part, existing commercial MT systems may be
expected to lack adequate or general treatments of the sorts of problem which require such
knowledge). On the other hand, such processing is not totally beyond the reach of current
theory. In particular, within certain limits, and in restricted domain s, techniques of semantic ,
pragmatic , and `real world' knowledge processing can be exploited with some success.

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node63.html [20/09/2000 18:04:04]

Further Reading

Next: Input Up: Representation and Processing Previous: Summary

Further Reading
Introductions to linguistic semantics include [Hurford and Heasley1983,Kempson1977], and,
at a rather more advanced level [Cann1993,Chierchia and McConnell-Ginet1990].

The discussion of tense and aspect given here is inspired by that used in the EUROTRA project,
which is described in [Allegranza et al. 1991,Van Eynde1993a]. This, in its turn, is inspired by
the work of [Bruce1972], and ultimately [Reichenbach1947].

As regards pragmatics , [Levinson1983,Leech1983] are useful introductions. Relevant work

on discourse structure includes [Grosz and Sidner1986,Pustejovsky1987]. The treatment of
common sense inference and real world knowledge is the field of Artificial Intelligence, see
for example [Rich1983,Tennant1981,Barr and Fiegenbaum1981,Shapiro1987]. On semantic
net s, see [Sowa1984]. The perspective we have taken in this Chapter is rather that suggested
by the programming language Prolog. For an easy introduction to this see [Rogers1986]. For
more advanced material directed at AI applications, see [Bratko1986], for material focussed on
NLP applications, see [Gazdar and Mellish1989].

The play-pen -- writing-pen example is from Bar-Hillel [Bar-Hillel1951].

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node64.html [20/09/2000 18:04:04]

Input

Next: Introduction Up: Machine Translation Previous: Further Reading

Input

● Introduction
● The Electronic Document
❍ Basic Ideas
❍ SGML Markup and MT Input
● Controlled Languages
● Sublanguage MT
● Summary
● Further Reading

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node65.html [20/09/2000 18:04:05]

Introduction

Next: The Electronic Document Up: Input Previous: Input

Introduction
In the scenario we imagined in Chapter , the text was delivered in the form of a
machine-readable document, having been prepared in such a way as to facilitate translation.
This is an important time saver. In this chapter, we describe how the full potential of machine
readable texts can be exploited in three ways: first, by adopting the notion of an `electronic
document' and embedding an MT system in a complete document processing system; second,
by restricting the form of input by using simplified or controlled language ; and third, by
restricting both the form, and the subject matter of the input texts to those that fall within a
sublanguage --- it is here that the immediate prospects for MT are greatest. The common theme
of this chapter is how the successful application of MT can be enhanced by ensuring that the
input to the system is `appropriate'. Briefly, the message is this: having texts in machine
readable form is a prerequisite for sensible use of MT, but one can get much better results by (i)
adopting certain standard formats for the input, (ii) controlling the input, so that problematic
constructions, etc., are avoided, and (iii) where possible, tailoring the MT systems to the
language of particular domains.

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node66.html [20/09/2000 18:04:06]

The Electronic Document

Next: Basic Ideas Up: Input Previous: Introduction

The Electronic Document

● Basic Ideas
● SGML Markup and MT Input

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node67.html [20/09/2000 18:04:06]

Basic Ideas

Next: SGML Markup and Up: The Electronic Document Previous: The Electronic Document

Basic Ideas
Every text that is not delivered as an electronic document on a floppy disc, a magnetic tape, or via a computer
network will have to be put into the system manually. Re-typing a text into the computer solely to make it available
for MT is unlikely to be cost-effective --- it would often be quicker to have the text translated directly by a human
translator. In recent years it has become practicable to use an optical character reader (OCR) to input text available
only in printed form. Clearly this is much quicker than re-typing, but checking for and correcting scanning errors can
be time-consuming.

However, if as is the case with ETRANS as described in Chapter , the MT system fits into an overall document
production system (DPS), then text can be created, translated, re-edited and generally prepared for publication within
the same electronic environment. In the first part of this chapter we will explore this notion of an electronic document
in some detail.

Electronic text is simply text which is available in a machine readable form. For example, electronic text is produced
by ordinary office word processors. At its simplest, such a text is just a sequence of characters, and, for the characters
in use in general computing (i.e. the English alphabet, normal punctuation characters, plus characters such as the
`space' character, the `line-feed' character, etc.) there is a standard representation provided by the ASCII codes,
which associates each character with a seven or eight bit code (i.e. a number --- e.g. a is ASCII 97, b is ASCII 98, A is
ASCII 65, the `space' character is ASCII 32). Unfortunately, this standard is not sufficient for encoding the letters of
foreign alphabets and their accents, even those based on the Roman alphabet, let alone non-Roman alphabets, and
characters in non-alphabetic scripts, such as Japanese characters (Kanji). One approach to such alphabets is to extend
the ASCII codes beyond those needed by English. Another is to represent foreign accents and special characters by
sequences of standard ASCII characters. For example, a German u with umlaut ( ü) might be represented thus:
\"{u}.
One problem is that there is (as yet) no genuine accepted standard beyond basic ASCII, with the further complication
that many word processors use non-ASCII representations `internally', as a way of representing text format (e.g.
information about typeface, underlining, etc.) This lack of standards means that it is necessary to use special
conversion programs if one wants to freely import and export text from different languages and a variety of DPSs
(such as word processors). Even when such programs exist, they do not always preserve all the information (e.g. some
information about format may be lost).
Part of a general solution to these problems, however, is to distinguish two components of a printed document: the
text itself (a sequence of words and characters); and its rendition --- the form in which it appears on the page (or
screen). For example, consider a title or heading. There are the words which make up the title --- perhaps a noun
phrase such as `The Electronic Document' --- and the particular presentation or rendition of those words on the page.
In this book all section and chapter headings are aligned with the left margin and different levels of heading (chapter,
section, subsection) are printed in a distinctive typeface and separated by a standard space from the preceding and
following paragraphs of text.
If we think about this distinction between text and rendition in electronic terms, it is easy to see that we have to code
both the characters in the text, and indicate how we intend parts of that text to appear on screen or in printed form. In
the early days of electronic text handling, this problem was solved in a rather direct and obvious fashion: the author
would type in not only the substance of the text but also some special codes at appropriate places to tell the printer to
switch into the appropriate type faces and point size. For example, in typing in a title the author would carefully insert
an appropriate number of carriage returns (non-printing characters which start a newline) to get a nice spacing before
and after. She would also make sure the title was centred or left-aligned as required, and finally she would type in
special codes (say \[223\[-447) before and after the title string to switch the printer into a bold typeface with 24
`points' to the inch and back to its usual font and size immediately afterwards.
There are three evident problems with such a procedure:

http://privatewww.essex.ac.uk/~doug/book/node68.html (1 de 5) [20/09/2000 18:04:09]

Basic Ideas

1. The codes used are likely to be specific to particular printers or word processing set-ups and hence the
electronic document will not be directly portable to other systems for revision, integration with other
documents or printing.
2. The author is required to spend some of her time dealing with rendition problems --- a task that (prior to the
advent of electronic systems) had always been conveniently delegated to the compositor in a printing house.
3. If at some point it is decided that a different rendition of headings is required, someone has to go through the
entire document and replace all the codes and characters associated with the rendition of each heading.
The printer codes are a sort of tiny little program for a particular printer. The next development was to replace these
rather specific programs by some means of stating directly ``I want this in 24 point Roman boldface'' --- perhaps by a
`markup' like this: `\roman\24pt\bf '. Each printer or word processor can then be equipped with a special
program (a so-called `driver') which interprets this high-level code and sends the printer or screen appropriate specific
low-level codes. Providing everyone used exactly the same high-level codes in all systems, the problem of portability
would be solved.
However, there is another way of tackling the rendition problem. When one thinks about it abstractly, the only thing
that the author really needs to put into the text is some markup which says (in effect) `This is a heading', or `This is a
footnote' or `This is an item in an item list' and so on. Each piece of text is thus identified as being an instance of
some class of text elements. With such markup, the author no longer has to worry about how each such marked
document element is going to be printed or shown on screen --- that task can be delegated to the document designer
(the modern equivalent of a compositor). The document designer can specify an association between each type of
document element and the high-level rendition codes she wants it to have. In other words, she can say that she wants
all headings to be printed in 24 point boldface Roman. The document handling system then ensures that headings etc.
are displayed and printed as required.
This type of markup, where the author simply identifies particular pieces of text as being instances of particular
document elements, is known as descriptive or `intensional' (`intentional') markup. This notion is fundamental to all
modern document processing systems and techniques. Not only does this provide flexibility in how text is rendered,
provided that the way in which markup is made is consistent from system to system, the result is that electronic
documents can be freely passed between systems.
We can now be a little more precise about the notion of an electronic document: it contains electronic or
machine-readable text with descriptive markup codes which may be used to determine the rendition and other usages
of the document. Before we go on to give an idea of how this can be exploited for MT, it may be worth a brief
description of the standard descriptive markup: SGML (Standardised General Markup Language) which is specified
by the International Standards Organization. It is our belief that in the next few years no serious commercial MT
system will be supplied without some means of handling SGML.
SGML specifies that, ordinarily, text will be marked up in the way shown in the last example above, i.e. with
document elements surrounded by their names in angle brackets. An office memo marked up in SGML might look
like the example below. In addition to the actual text, various pairs of SGML tags delimiting the memo elements can
be seen here. The memo as a whole starts with <Memo> and ends with </Memo> (where / indicates the closing
delimiter). In between the Memo tag pair we find the sub-elements of the memo, also marked-up with paired tags (
<To>... </To>, <From> .. </From>, <Body>... <P>..</P>...</Body>).

The relationship between SGML tags, and the way text is actually rendered is given by an association table, such a
table might say, e.g. that the body of a memo should be separated from the previous part by a horizontal line. When

http://privatewww.essex.ac.uk/~doug/book/node68.html (2 de 5) [20/09/2000 18:04:09]

Basic Ideas

actually printed, this memo might look as in Figure :

Figure: How a Memo Marked Up in SGML Might Appear When Printed

The tagging principles of SGML are intended to extend to very complex and highly structured documents. Imposing
such a structure not only allows very fine, and flexible control of how documents are printed, it can also allow easy
access to and manipulation of information in documents, and straightforward consistency checking .
One thing the SGML standard does not do is try to specify a standard inventory of all possible document elements.
Users are perfectly free to define their own document types and to specify the elements in those documents. SGML
provides a special method of doing this known as a Document Type Definition (DTD) . A DTD is a sort of formal
grammar specifying all such relations in a particular type of document. For example, such a grammar might say that
all Memos (all our Memos at least) contain a To element followed by a From element followed by a Body element,
which itself contains at least one Paragraph followed by zero or more Paragraphs. This means that a Memo has the
following sort of DTD (grossly simplified):
Memo

http://privatewww.essex.ac.uk/~doug/book/node68.html (3 de 5) [20/09/2000 18:04:09]

Basic Ideas

To, From, Body Body Paragraph, Paragraph*

Using a DTD has several advantages:
1. The DTD makes sure that documents are truly portable between different SGML document systems; the
document system reads the accompanying DTD to find out what sort of elements will be in the document and
how they will be arranged with respect to each other. Thus, the document processing system knows what to
expect when it encounters a document which is an instance of a certain DTD.
2. It ensures that documents of a particular type (e.g. user manuals) are always structurally consistent with each
other. It suffices to define a DTD for the class of user manuals and then the SGML document processing
system will ensure that all documents produced by that DTD will indeed have the same overall structure. In
short, DTDs help to promote a certain rigour which is extremely desirable in technical documentation.
3. The use of DTDs in document preparation allows authors to deal directly with the content of texts whilst
having little or no direct contact with the actual markup used. What happens with the usual sort of SGML
system is that there is a window offering the author a choice of document entities appropriate for the document
she is preparing or revising. This list of document entities is obtained by reading the DTD for the document.
For example, in a memo, there will be a choice of To, From, and Body. The author clicks on the appropriate
element and the markup is entered into the text (perhaps invisibly). When actually typing in the Body, the
choice is narrowed down to Paragraph. Whilst this is not particularly interesting for simple documents like
memos, it is clear that it would be be immensely useful in constructing complex documents, and in document
retrieval.
With this general idea of Electronic Documents and markup, we can look at how an MT system can exploit the fact
that texts are represented in this way.

Next: SGML Markup and Up: The Electronic Document Previous: The Electronic Document

http://privatewww.essex.ac.uk/~doug/book/node68.html (4 de 5) [20/09/2000 18:04:09]

Basic Ideas

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node68.html (5 de 5) [20/09/2000 18:04:09]

SGML Markup and MT Input

Next: Controlled Languages Up: The Electronic Document Previous: Basic Ideas

SGML Markup and MT Input

An MT system should only attempt to translate things that are translatable. Suppose that some text contains the
acronym `MAT', which refers to a company called `Machine Aided Translation Ltd'. Clearly the correct translation
of this is either just MAT again or some new acronym that reflects the translation of the underlying name ---
perhaps TAO in French , being the acronym for Traduction Assistée par Ordinateur, which itself is the translation
of Machine Aided Translation. What is unquestionably incorrect is a translation of the form pallaison, this being
the sort of mat that a cat might sit on. The reader may think that the MT system ought to have spotted that MAT
cannot be a standard concrete noun because it is capitalised; but many MT systems routinely ignore capitals
because they need to recognise ordinary words which can appear with an initial capital letter at the start of a
sentence.
The way to deal with this sort of problem is to ensure that acronyms are recognised as a particular class of text
elements and marked up as such. This might be done (a) either by the author when the text is being created or (b) by
special tools used before translation which help translators to find acronyms and the like and mark them up
accordingly. For example, a specialised search and replace tool inside the document pre-editor could look for all
sequences of capitalised words and, after querying the translator to check whether a particular candidate sequence
really is an acronym, insert the appropriate markers in the text. The point is that once the text is marked up, the MT
system is in a much better situation to know that it is dealing with an untranslatable acronym and to treat it
accordingly.
Similarly, consider figures and diagrams in a document. These consist usually of pictorial material, which is
untranslatable, and a translatable text caption which characterises the pictorial material. Recognising the markup
tags which indicate that the following material in the document is pictorial, the MT system can simply ignore
everything until it encounters another tag telling it that it is about to see the caption, which it can translate as a
normal piece of text. Equally, it is easy to ask the MT system to translate (say) just a single chapter, because the
markup in the document will clearly identify the piece of text that constitutes the chapter. Markup is thus a
powerful tool in controlling the MT process.
DTDs are particularly useful in MT. Some MT systems keep a copy of each sentence they have already
encountered together with its translation (the post-edited version, if available). This habit is known in the industry
as Translation Memory. Over the years, MT vendors have found that in some organizations much of the
translation workload consists of entirely re-translating revised editions of technical manuals. These revised editions
may contain as much as 90% of the material that was already present in the previous edition --- and which was
already translated and post-edited. Hence automatically recognising sentences already translated and retrieving the
post-edited translation - as the Translation Memory technique allows --- results in a 90% reduction in post-editing
costs (and an enormous increase in the overall speed of the translation process). This is clearly very significant.
However, these sort of performance improvements are really the result of a defective documentation process. The
problem is that the organization paying to have the translation done is not keeping proper track of which parts of a
revised document really are different from the original version. Clearly only new or altered material really needs to
be even considered for translation.
Within the SGML standard it is possible to add features to text elements to record when they were last altered, by
whom and so on. This version control information can be maintained by the document system and it allows the user
to extract revised elements. Indeed, the principle can be extended so that earlier versions of a given revised element
are also kept, allowing the user to reconstruct any previous version of a document at any point.
The result of exercising proper version control in documentation is that only new elements for which there are no
existing translations will be submitted to the translation process. In this way, the document processing system takes
some of the burden otherwise carried by the MT system (viz, the `Translation Memory' facility).

http://privatewww.essex.ac.uk/~doug/book/node69.html (1 de 2) [20/09/2000 18:04:11]

SGML Markup and MT Input

Another advantage of using DTD s in MT involves generalizing the notion of a document slightly, to introduce the
notion of a `multilingual document'. In SGML, this is largely a matter of altering the DTDs of monolingual
document types. Take the Memo example: we can get a multilingual version by specifying that there is a copy of
each document element for each language. Here is a revised (and still simplified) Memo DTD for two languages:

Memo To, From, Body Body Paragraph, Paragraph*

Paragraph

Paragraph-L1, Paragraph-L2
There are now two types of Paragraph --- Paragraphs in language one and Paragraphs in language 2. Each
Paragraph element will contain one language 1 paragraph followed by one language 2 paragraph. (There are no
language specific To and From elements because it is assumed that they contain only proper names). This sort of
technique can be generalised to allow a document to carry text in arbitrarily many languages. Though this allows a
document to contain text for more than one language, it does not require it --- document elements can be empty ---
this would be the case for target language elements where the source element has not yet been translated.
The important thing to understand here is that just because the simple multilingual DTD we have described
`interleaves' the elements for different languages (we have a paragraph for L1 followed by the corresponding
paragraph for L2, etc.), this does not mean that we have to view the document that way. For example, a Memo in
English, French and German can be viewed on the screen of a document processing system with all the English
paragraphs, printed together, and the French paragraphs printed alongside, with the German paragraphs not shown
at all. Part of the flexibility in the rendition of a marked-up document is that the text content of classes of elements
can be hidden or shown at will. In practical terms, this means that a translator editing a multilingual document will
have considerable flexibility in choosing the way in which that document is presented (on screen or on paper) and
in choosing the type of element she wishes to see.

Turning back to the MT case, recall that in the scenario in Chapter , ETRANS takes the German text and then
makes available the English translation in the multilingual document. It should now be much clearer how this
works. Translatable elements from the source text are passed to the ETRANS system which then translates them.
The translated text is then placed under the corresponding target language text elements (which, up that point, have
been entirely empty of text). So far as is linguistically possible, the structure of the document is preserved.
In summary, it should be clear that the general idea of the Electronic Document is important within the context of
MT and can make a considerable contribution to the successful integration of MT within the office environment.

Next: Controlled Languages Up: The Electronic Document Previous: Basic Ideas

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node69.html (2 de 2) [20/09/2000 18:04:11]

Controlled Languages

Next: Sublanguage MT Up: Input Previous: SGML Markup and

Controlled Languages

The notion of controlled languages was introduced in Chapter where we described it as a

form of language usage restricted by grammar and vocabulary rules. The original idea arose
during the 1930s, when a number of influential linguists and scholars devoted considerable
effort to establishing a `minimal' variety of English, a variety specifically designed to make
English accessible to and usable by the largest possible number of people world wide. Basic
English , as it was called, differed from previous attempts to construct universal languages in
that it was a perfectly well-formed part of English, rather than some entirely artificial or hybrid
construction such as Esperanto. One of the central ideas of the Basic English movement was
that the number of general-purpose words needed for writing anything from a simple letter of
receipt through to a major speech on the world economic situation could be a few hundred
rather than the 75 000 upward available to skilled native speakers. This lexical economy was to
be achieved in part by using `operator verbs' with the set of nouns and adjectives to stand in for
the vast number of derived verbs which are frequently used. For example, whereas in ordinary
English we might write The disc controller design was perfected over numerous revisions,
Basic English would say ... was made perfect ..., where make is one of the operator verbs and
perfect one of the licensed Basic English adjectives.
The authors of Basic English explicitly recognised that the dictionary would need to be
extended with special terminology for scientific and technical writing. However, even if a text
contained terminology specific to a certain subject field, the general language component of the
text could perfectly well be accommodated within Basic English. The important point remains
that, for writing in a particular subject field, no more is needed than the Basic English
dictionary together with a (relatively small) technical vocabulary for that field.
The idea was later taken on by English-language based (predominantly North American)
corporations marketing capital goods on a world-wide basis. Rather than try to translate engine
manuals and the like into every possible language that might be required, it was assumed that if
they were written with sufficient care and attention, they could be read fairly easy by service
engineers and mechanics with limited English skills.
Although controlled languages were introduced partly to avoid or reduce human translation
costs, two important additional benefits were discovered. First, the readability and clarity of a
controlled language technical text often seems better than uncontrolled texts --- even for native
English readers. Second, controlled languages produce better results with MT than uncontrolled
languages.
The reasons for controlled languages' superior MT performance are easy to understand. First,
the restricted vocabulary means that fewer words need to be added to the MT system
dictionaries and more effort can be put into getting the entries which are required right. Second,

http://privatewww.essex.ac.uk/~doug/book/node70.html (1 de 3) [20/09/2000 18:04:13]

Controlled Languages

the grammar component of the system can be tailored to handle all and only those constructions
which are licensed by the controlled language specification, a specification which excludes the
most difficult and ambiguous constructions anyway.

A flavour of what is involved can be obtained by looking at the writing rules given above and
the dictionary excerpt on page , which are based on those of PACE, the controlled English
used by the UK Engineering company Perkins Engines. As will be clear from the dictionary
excerpt, the general principle is `one word, one meaning', for example, the only use of the verb
advise is `to give advice'. Thus, a usage such as Please advise us of the availability of parts at
your earliest convenience would not be allowed, since here it means `tell'. A useful
development of such a dictionary for MT purposes would be to add information about how
these words translate.
Using a restricted pool of words and terms also means that the system dictionaries can be
tailored (by the MT supplier or responsible translator) to cover exactly that set of words and
their translations. Being consistent about the use of terms will also help to improve the overall
consistency and quality of the texts being translated. After all, one of the simplest and most
direct benefits of MT for technical texts is that terms are always translated consistently because
they are simply looked up in an electronic bilingual term dictionary .

In general, it can be seen that the rules are mainly advice on constructions that should be
avoided, usually because they lead to ambiguity . The rules for controlled languages tend to be
stylistic guidelines rather than hard and fast grammar specifications. In general, much of the
success of controlled languages as corporate language tools stems from the emphasis placed on
critical analysis of the text and precise presentation of ideas. This is particularly apparent in the
first example on page , which illustrates the dramatic effect of using a controlled version of
English.
It is not particularly difficult to train people to write controlled language text i.e. text which
generally observes some set of fairly simple writing rules. For example, the Xerox corporation
currently offers its technical writers a one-day course in writing with MCE (Multinational
Customised English, a Xerox proprietary language). British Aerospace teaches the rudiments
of Simplified English (a general purpose technical English for the international aerospace
industry) in a few fairly short training sessions.

Next: Sublanguage MT Up: Input Previous: SGML Markup and

http://privatewww.essex.ac.uk/~doug/book/node70.html (2 de 3) [20/09/2000 18:04:13]

Controlled Languages

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node70.html (3 de 3) [20/09/2000 18:04:13]

Sublanguage MT

Next: Summary Up: Input Previous: Controlled Languages

Sublanguage MT

In the previous section, we looked at a method of controlling the input to an MT system, simplifying it by
avoiding certain uses of words, and avoiding potentially ambiguous constructions. Since the success of the
METEO MT system, which we mentioned briefly in Chapter , an important strand of MT has involved
concentrating on what we could loosely call `MT for Special Purpose Languages', or sublanguage MT. Here,
rather than imposing controls or simplifications on writers, one tries to exploit the restrictions in terms of
vocabulary and constructions that users of the language for specialized purposes normally accept, or simply
observe without reflection. The term sublanguage refers to the specialized language used (predominantly for
communication between experts) in certain fields of knowledge, for example, the language of weather reports,
stockmarket reports, the language of some kinds of medical discussion, the language of aeronautical
engineering. Specialized vocabulary is one characteristic of such `languages' (they typically contain words not
known to the non-specialist and also words used in different or more precise ways). However sublanguages are
also often characterised by special or restricted grammatical patterns. In MT, it is quite common to use the term
sublanguage rather loosely to refer not just to such a specialized language, but to its use in a particular type of
text (e.g. installation manuals, instruction booklets, diagnostic reports, learned articles), or with a particular
communicative purpose (communication between experts, giving instructions to non-experts, etc).
The chief attraction of sublanguage and text type restriction to MT researchers is the promise of improved
output, without the need to artificially restrict the input. Restricting the coverage to texts of particular types in
certain subject domains will allow one to profit from regularities and restrictions in syntactic form and lexical
content. This may be important enough to permit significant simplification of the architecture, and certainly
leads to a reduction in the overall coverage required. We reproduce an example from English to French output
from METEO :

Of course, the language of meteorological reports is special in happening to combine a rather small vocabulary
with a simple, telegraphic style of writing (notice in particular the complete absence of tenses from these
extracts --- the few verbs there are in non-finite forms). Nonetheless, a simplification of lexical and possibly
syntactic coverage can be expected in less extreme cases. To give an example with respect to lexical coverage,
it is reported that 114 of the 125 occurrences of the verb to match in a computer software manual translate into
the Japanese icchisuru-suru, which is listed as one of the less frequent of the 15 translations given in a
small-size English-Japanese dictionary. In the extract from a corpus of telecommunications text given below,
traffic always corresponds to the French trafic and never to circulation (which applies only to road traffic).
Moreover the dictionary writer can safely ignore the meaning of both trafic and traffic concerning dealings in
illegal merchandise (`drug traffic'). Also, for an increasing number of sublanguages one can rely on the
availability of a termbank (an on-line (multilingual) terminological dictionary ) defining and stating
equivalences for many of the technical terms that will be encountered. This greatly eases the job of dictionary
construction. Such examples can be multiplied almost at will.

As for syntactic coverage, examples of instruction manuals and other forms of

informative documentation typically share a number of common features. There will probably be no idioms ,
and a restricted set of sentential patterns. Another common feature is the relatively simple temporal dimension
of the text, e.g. predominant use of the simple present. There is also the common occurrence of enumeration as
a form of conjunction, usually either numbered or inset by dashes, etc. Some of these features can be seen by

http://privatewww.essex.ac.uk/~doug/book/node71.html (1 de 4) [20/09/2000 18:04:15]

Sublanguage MT

comparing the examples of English and French given below, which are drawn from a corpus of texts about
Telecommunications. All are of great benefit to the developer or user of an MT system. For the developer, they
mean that there are fewer problems of ambiguity , and development effort can be concentrated on a smaller
range of constructions. For the user, this should mean that better coverage is obtained, and that the system
performs better.
It is not, of course, the case that expository texts in different languages always exploit the same devices for a
particular communicative purpose. The following extracts from the same corpus show that English and French
differ in their use of impersonal constructions, with French favouring such constructions with the impersonal
subject pronoun il (`it') far more in this type of text than English does. But even in these cases, it is generally
easier to choose the correct translation, simply because the range of possibilities in such texts is smaller.
(Literal translations of the phrases we have picked out would be: `It is advisable to take account of...', It is
manifestly much more difficult...', and `It is advisable to take...'.)
● ❍ In this framework, the progressive evolution of the earth segment should be considered.

Dans ce contexte, l'évolution progressive du secteur terrien.

● ❍ Setting up a new satellite system, which may be either a regional system with the participation of a
group of countries, or a purely national
(domestic) system, is obviously much more difficult than using an existing system:
de mettre en place un nouveau système à satellites (système régional auquel participe un
groupe de pays ou système purement national) que d'utiliser un système existant:
● ❍ Simultaneously, arrangements should be made for recruitment and training of staff for installation,
operation and maintenance.
❍ En même temps, des dispositions pour le recrutement et la formation du personnel qui
sera chargé de l'installation, de l'exploitation et de la maintenance.
Text type can strongly influence translation, not just because certain syntactic constructions are favoured (e.g.
conjunction by enumeration), but also by giving special meanings to certain forms. An example of how the text
type can be useful in determining translational equivalents is the translation of infinitive verb forms from
French or German into English. Infinitives normally correspond to English infinitives, but are usually
translated as English imperatives in instructional texts. Thus, in a printer manual one would see (
b) as the translation of (

http://privatewww.essex.ac.uk/~doug/book/node71.html (2 de 4) [20/09/2000 18:04:15]

Sublanguage MT

a), rather than the literal translation.

● ❍ Richtige Spannung
`correct voltage to set'
❍ correct voltage
● ❍ les commandes
`to execute the commands'
❍ the commands
Thus, concentration on a sublanguage not only restricts the vocabulary and the number of source and target
language constructions to be considered, it can also restrict the number of possible target translations. Given the
potential that sublanguages provide for improvements in the quality of output of MT systems, and the fact that
most commercial institutions do in fact have their major translation needs in restricted areas, it is not surprising
that many research prototypes concentrate on restricted input in various ways, and that the design of tools and
resources supporting sublanguage analysis is a major area of research.

Next: Summary Up: Input Previous: Controlled Languages

Arnold D J

http://privatewww.essex.ac.uk/~doug/book/node71.html (3 de 4) [20/09/2000 18:04:15]

Sublanguage MT

Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node71.html (4 de 4) [20/09/2000 18:04:15]

Summary

Next: Further Reading Up: Input Previous: Sublanguage MT

Summary
In this chapter we have discussed three ways in which one can increase the likelihood of MT
being successful by taking care with the input to the system. We first concentrated on the
importance of integrating MT into the general document preparation environment, introducing
the notion of the electronic document . We stressed the importance of standards in the encoding
of texts, and showed how the process of MT can be aided by the adoption of the SGML
markup language. In the following section, we turned to the content of texts themselves and
introduced the notion of controlled languages, in which one adopts a simplified form of the
language in order to communicate simply and unambiguously with the reader. Using a
controlled language input greatly enhances the quality of output of MT systems. Finally we
discussed sublanguage MT, or MT in restricted domains, observing that the language used in
specialized technical domains is often quite different from and more restricted in style and
content than the `general language', and it is possible to take advantage of these characteristics
by tailoring an MT system to the language of particular domains.

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node72.html [20/09/2000 18:04:16]

Further Reading

Next: Evaluating MT Systems Up: Input Previous: Summary

Further Reading
SGML is defined in ISO 8879, 1986, which is extensively discussed in the standard reference
book on SGML[Goldfarb1986] . An excellent introduction to SGML is provided in [van
Herwijnen1990].

The examples of the use of controlled language that we give in the text are based on those in
[Pym1990]. See [Pym1990,Newton1992b] for discussion of the use of PACE as part of the
translation operation in Perkins.
A noteworthy example of a controlled language is Simplified English (SE), which is described
in the AECMA/AIA Simplified English Guide [AECMA1988]. This grew out of work done in
the late 1970s, on behalf of the Association of European Airlines (AECMA) into readability of
maintenance documentation within the civilian aircraft industry. As a result, an AECMA
working group researched the procedural texts in maintenance manuals. It contains a limited
general vocabulary of about 1500 words and a set of Writing Rules, similar to those we will
describe above.
On sublanguage , [Arnold1990] provides a short overview. [Lehrberger1982] and [Grishman
and Kittredge1986] are collections of articles on the subject. More detailed discussions can be
found in [Kittredge1982] , [Kittredge1987], [Sager1982], [Slocum1986], [Teller et al. 1988]
and [Hirschman1986].

Météo is described in [Hutchins and Somers1992, Chapter 12,], see also [Isabelle1987].
Recent developments are described in [Chandioux1976], [Chandioux1989a],
[Chandioux1989b], and [Grimaila and Chandioux1992].

The example concerning the English-Japanese translation of match in software manuals is

reported in [Tsujii et al. 1992].

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node73.html [20/09/2000 18:04:17]

Evaluating MT Systems

Next: Introduction Up: Machine Translation Previous: Further Reading

Evaluating MT Systems

● Introduction
● Some Central Issues
● Evaluation of Engine Performance
❍ Intelligibility
❍ Accuracy
❍ Error Analysis
❍ The Test Suite
● Operational Evaluation
● Summary
● Further Reading

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node74.html [20/09/2000 18:04:18]

Introduction

Next: Some Central Issues Up: Evaluating MT Systems Previous: Evaluating MT Systems

Introduction
How can you tell if an MT system is `good'? How can you tell which of two systems is `better'?
What do `good' and `better' mean in this context? These are the questions that this chapter tries
to answer.
In a practical domain like MT, such questions reduce to questions of suitability to users' needs:
what is the best and most economical way to deal with the user's translation requirements? In
the ideal case, it should be possible to give a simple and straightforward answer to this question
in a consumers' magazine. An article in such a magazine would discuss the most important
issues with a comparison table displaying the achievements of different MT systems on tests of
important aspects such as speed and quality . Unfortunately, the information necessary to make
informed judgements is not so readily available, partly because the methods for investigating
suitability are not well developed. In reality, MT users can spend quite a lot of money finding
out what a system can and cannot do for them. In this chapter we will look at the kind of thing
that should matter to potential users of MT systems, and then discuss some existing methods for
assessing MT system performance.

As we pointed out in the Introduction (Chapter ), we think that, in the short term, MT is
likely to be of most benefit to largish corporate organizations doing a lot of translation. So we
adopt this perspective here. However, most of the considerations apply to any potential user.

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node75.html [20/09/2000 18:04:18]

Some Central Issues

Next: Evaluation of Engine Up: Evaluating MT Systems Previous: Introduction

Some Central Issues

The evaluation of MT systems is a complex task. This is not only because many different
factors are involved, but because measuring translation performance is itself difficult. The first
important step for a potential buyer is to determine the translational needs of her organization.
Therefore she needs to draw up a complete overview of the translational process, in all its
different aspects. This involves establishing the size of the translation task, the text type of the
material and its form (is it machine readable and if so, according to which standards). It also
involves considering organizational issues, e.g. the tasks of each member of staff concerned in
some way with translation. With that information at hand she can start to investigate what the
consequences of the purchase of an MT system would be. These are some of the factors to keep
in mind:
Organizational Changes
Incorporating an MT system into the translation process will impact upon both the
process and the personnel involved. There will be consequences for system
administrators and support staff, but above all for the translators themselves, whose tasks
will change significantly. Whereas before they will probably have spent the major part of
their time actually translating or editing human translations, they will now find
themselves spending a lot of time updating the system's dictionaries and post-editing the
results of machine translation. There may also be a need to build automatic termbanks .
Translators will need to receive training in order to perform these new tasks adequately.
It is important that the personnel support the changeover to MT. They may not always be
aware of the fact that MT can lead to more job satisfaction among translators since MT
systems are particularly efficient at tedious, repetitive tasks whereas more challenging
translation work often still needs to be done by the human translators. If translators in an
organization have decided for some reason or other that they do not want to work with
MT, imposing it on them is guaranteed to produce poor results.
Technical environment
We have emphasised right from the start that success depends in part on MT being
effectively incorporated as part of a wider document preparation process inside an
organization. Smooth handling of text throughout the whole process will prevent
unnecessary delays. The MT engine and the document system may well come from
different suppliers but they must adhere to the same standards and formats for textual
material.
Bear in mind that good document preparation facilities in themselves can improve
translator productivity. A decade or so ago much of the productivity increase claimed by
some vendors of smaller MT systems could be attributed to their providing rather good
multi-lingual word processing facilities, at a time when many translators used only an
electric typewriter. Some MT vendors still supply a whole MT system package where the

http://privatewww.essex.ac.uk/~doug/book/node76.html (1 de 3) [20/09/2000 18:04:20]

Some Central Issues

engine is inextricably wrapped up with some specialised word processing and

text-handling tool unique to that particular system. This is undesirable on two counts:
first, if you are already familiar with a good multi-lingual word processor, little is gained
by having to learn another which does much the same things; second, it is likely that an
MT vendor's home-grown text-processing facilities will be inferior to the best
independent products, because most of the effort will have gone into developing the
translation engine.
Status of Vendor
Buying an MT system is a considerable investment, and the stability and future solvency
of the vendor is an important consideration. After all, contact with the vendor is ideally
not just limited to the initial purchase of the system. A solvent vendor can provide
installation support and training in the early stages, and general support and updates later,
which may improve performance considerably (e.g. specialized dictionaries, or new
language pairs which can be integrated into the existing MT set-up).

Engine Performance: Speed

In some circumstances, the speed at which the
engine churns out raw translated text won't actually be crucial. If the system requires
interaction with the translator whilst it is translating, then of course it should not amble
along so slowly as to to keep the translator waiting all the time. But if it is functioning
without direct interaction, it can proceed at its own pace in the background whilst the
translator gets on with other jobs such as post-editing or hand translation of difficult
material. This aspect also depends on the user's translational needs: if the user's material
requires 15 hours daily on a fast MT system and 20 on a slower one, no one will notice
the difference if the system is running overnight. Of course, there are situations where the
quick delivery of translation output is essential. (The agronomist in Chapter , who
wants to process very large quantities of material to a low level may be an example.) But
in general, slow speed is the one component of MT performance of which upgrading is
relatively easy: by buying some faster hardware for it to run on.
Engine Performance: Quality
This is a major determinant of success. Current general purpose commercial MT
systems cannot translate all texts reliably. Output can sometimes be of very poor quality
indeed. We have already mentioned that the post-editing task (and with it the cost)
increases as translation quality gets poorer. In the worst case, using MT could actually
increase translation costs by tying up translators in editing and maintenance tasks,
ultimately taking up more time than would have been required to produce translations
entirely by hand. Because of its enormous influence on the overall translation cost,
translation quality is a major aspect in MT evaluation.

http://privatewww.essex.ac.uk/~doug/book/node76.html (2 de 3) [20/09/2000 18:04:20]

Some Central Issues

Next: Evaluation of Engine Up: Evaluating MT Systems Previous: Introduction

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node76.html (3 de 3) [20/09/2000 18:04:20]

Evaluation of Engine Performance

Next: Intelligibility Up: Evaluating MT Systems Previous: Some Central Issues

Evaluation of Engine Performance

Substantial long-term experience with particular MT systems in particular circumstances shows

that productivity improvements and cost-savings actually achieved can be very variable. Not all
companies can apply MT as successfully as the following:

Different organizations experience different results with MT. The above examples indicate that
the kind of input text is one of the important factors for getting good results. A sound system
evaluation is therefore one which is executed within the company itself. An MT vendor might
provide you with translated material which shows what their system can do. There is, however,
no guarantee that the system will do the same in a different company setting, with different
texts. Only a company specific evaluation will provide the client with the feedback she
ultimately wants. Information provided by the MT vendor can be useful though, e.g. if system
specifications indicate what sort of text type it can or cannot handle or what sort of language
constructions are problematic for their system.
In evaluating MT systems one should also take into account the fact that system performance
will normally improve considerably during the first few months after its installation, as the
system is tuned to the source materials, as discussed in Chapter . It follows that performance
on an initial trial with a sample of the sort of material to be translated can only be broadly
indicative of the translation quality that might ultimately be achieved after several months or
years of work.
Something similar holds for those stages of the translation process which involve the translator,
like dictionary updating and post-editing of the output. Times needed for these tasks will
reduce as translators gain experience.
So how do we evaluate a system? Early evaluation studies were mainly concerned with the
quality of MT. Of course, assessing translation quality is not just a problem for MT: it is a
practical problem that human translators face, and one which translation theorists have puzzled
over. For human translators, the problem is that there are typically many possible translations,
some of them faithful to the original in some respects (e.g. literal meaning), while others try to
preserve other properties (e.g. style , or emotional impact).
In MT, the traditional transformer architecture introduces additional difficulties, since its
output sentences often display structures and grammar that are unknown to the target language.
It is the translator's task to find out what the correct equivalent is for the input sentence and its

http://privatewww.essex.ac.uk/~doug/book/node77.html (1 de 2) [20/09/2000 18:04:21]

Evaluation of Engine Performance

ill-formed translation. And, in turn, the evaluator's task is to find out how difficult the
translator's task is.
In the rest of this chapter we will describe the most common evaluation methods that have been
used to date and discuss their advantages and disadvantages.

● Intelligibility
● Accuracy
● Error Analysis
● The Test Suite

Next: Intelligibility Up: Evaluating MT Systems Previous: Some Central Issues

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node77.html (2 de 2) [20/09/2000 18:04:21]

Intelligibility

Next: Accuracy Up: Evaluation of Engine Previous: Evaluation of Engine

Intelligibility
A traditional way of assessing the quality of translation is to assign scores to output
sentences. A common aspect to score for is Intelligibility, where the intelligibility of a
translated sentence is affected by grammatical errors, mistranslations and untranslated words.
Some studies also take style into account, even though it does not really affect the intelligibility
of a sentence. Scoring scales reflect top marks for those sentences that look like perfect target
language sentences and bottom marks for those that are so badly degraded as to prevent the
average translator/evaluator from guessing what a reasonable sentence might be in the context.
In between these two extremes, output sentences are assigned higher or lower scores depending
on their degree of awfulness --- for example, slightly fluffed word order (`` ... in an interview
referred Major to the economic situation...'' will probably get a better score than something
where mistranslation of words has rendered a sentence almost uninterpretable (`` ...the peace
contract should take off the peace agreement....). Thus scoring for intelligibility reflects directly
the quality judgment of the user; the less she understands, the lower the intelligibility score.
Therefore it might seem a useful measure of translation quality.
Is there any principled way of constructing an intelligibility scoring system? Or rather is there
any generally agreed, and well motivated scoring system? We do not know of any. The major
MT evaluation studies which have been published report on different scoring systems; the
number of points on the scoring scales ranging from 2 (intelligible, unintelligible) to 9. The 9
point scale featured in the famous ALPAC Report and was not just used to score the
intelligibility of MT, but also of human translation. As a consequence the scale included
judgments on fairly subtle differences in e.g. style . This scale is relatively well-defined and
well-tested. Nevertheless we think that it is too fine-grained for MT evaluation and leads to an
undesirable dispersion of scoring results. Also, we think that style should not be included
because it does not affect the intelligibility of a text. On the other hand, a two point scale does
not give us enough information on the seriousness of those errors which affect the
intelligibility. (A two point scale would not allow a distinction to be drawn between the
examples in the previous paragraph, and complete garbage, (or something completely
untranslated) and a fully correct translation.) Perhaps a four point scale like the one below
would be more appropriate.

Once devised, scoring scales need to be tested, to make sure that scale descriptions are clear and
do not contain any expression that can be interpreted differently by different evaluators. The
test procedure should be repeated until the scale descriptions are uniformly interpreted by
evaluators.
A reasonable size group of evaluators/scorers must be used to score the MT output. Four
scorers is the minimum; a bigger group would make the results more reliable. The scorers
should be familiar with the subject area of the text they will score and their knowledge of the

http://privatewww.essex.ac.uk/~doug/book/node78.html (1 de 2) [20/09/2000 18:04:22]

Intelligibility

source language of the translation should also be good. Before an official scoring session is held
the scorers participate in a training session in which they can become acquainted with the scale
description. This training session should be similar for all scorers. During scoring it should be
impossible to refer to the source language text.

Next: Accuracy Up: Evaluation of Engine Previous: Evaluation of Engine

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node78.html (2 de 2) [20/09/2000 18:04:22]

Accuracy

Next: Error Analysis Up: Evaluation of Engine Previous: Intelligibility

Accuracy
By measuring intelligibility we get only a partial view of translation quality. A highly
intelligible output sentence need not be a correct translation of the source sentence. It is
important to check whether the meaning of the source language sentence is preserved in the
translation. This property is called Accuracy or Fidelity. Scoring for accuracy is normally done
in combination with (but after) scoring for intelligibility.
As with intelligibility , some sort of scoring scheme for accuracy must be devised. Whilst it
might initially seem tempting to just have simple `Accurate' and `Inaccurate' labels, this could
be somewhat unfair to an MT system which routinely produces translations which are only
slightly deviant in meaning. Such a system would be deemed just as inaccurate as an automated
`Monty Python' phrasebook which turns the innocent request Please line my pockets with
chamois into the target language statement My hovercraft is full of eels. Obviously enough,
if the output sentence is complete gobbledegook (deserving of the lowest score for
intelligibility) then it is impossible to assign a meaning, and so the question of whether the
translation means the same as the original cannot really be answered. (Hence accuracy testing
follows intelligibility rating).
The evaluation procedure is fairly similar to the one used for the scoring of intelligibility .
However the scorers obviously have to be able to refer to the source language text (or a high
quality translation of it in case they cannot speak the source language), so that they can
compare the meaning of input and output sentences.
As it happens, in the sort of evaluation considered here, accuracy scores are much less
interesting than intelligibility scores. This is because accuracy scores are often closely related
to the intelligibility scores; high intelligibility normally means high accuracy. Most of the time
most systems don't exhibit surreal or Monty Python properties. For some purposes it might be
worth dispensing with accuracy scoring altogether and simply counting cases where the output
looks silly (leading one to suppose something has gone wrong).
It should be apparent from the above that devising and assigning quality scores for MT output
--- what is sometimes called `Static' or `Declarative Evaluation' --- is not straightforward.
Interpreting the resultant scores is also problematic.
It is virtually impossible --- even for the evaluator --- to decide what a set of intelligibility and
accuracy scores for a single MT system might mean in terms of cost-effectiveness as a `gisting'
device or as a factor in producing high quality translation. To see this, consider the sort of
quality profile you might get as a result of evaluation (Figure ), which indicates that most
sentences received a score of 3 or 4, hence of middling intelligibility. Does that mean that you
can use the system to successfully gist agricultural reports? One cannot say.

http://privatewww.essex.ac.uk/~doug/book/node79.html (1 de 2) [20/09/2000 18:04:24]

Accuracy

Figure: Typical Quality Profile for an MT System

Turning to the high-quality translation case, it is clear that substantial post-editing will be
required. But it is not clear --- without further information about the relationship between
measured quality and post-editing times --- what effect on overall translator productivity the
system will have. Whilst it is presumably true that increasingly unintelligible sentences will
tend to be increasingly difficult to post-edit , the relationship may not be linear. For example, it
may be that sorting out minor problems (which don't affect intelligibility very much) is just as
much of an editing problem as correcting mistranslations of words (which affect intelligibility a
great deal). We could for example imagine the following two sentences to be part of our sample
text in Chapter 2. The first one is more intelligible than the second, yet more time will be
needed to fix the errors in it:
● ❍ The print page should be excell quality,
❍ The printed page should excellent quality.
It is true that a comparative evaluation of a number of different MT systems might demonstrate
that one system is in all respects better than the others. The information however does not tell
us whether buying the better MT system will improve the total translation process --- the
system could still be unprofitable. And even if two particular systems have different
performance profiles, it may not always be clear whether one profile is likely to be better
matched to the task in hand than the other. For example, look at the intelligibility ratings for
systems A and B in Figure . For system A the majority of sentences are neither very good
nor bad (rating 3 or 4). System B, by comparison, tends to do either quite well (scores of 7 are
common) or quite badly (scores 1, and 2 are frequent). Which system will be better in practice?
It is not possible to say.

Figure: Which Performance Curve is Better?

Next: Error Analysis Up: Evaluation of Engine Previous: Intelligibility

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node79.html (2 de 2) [20/09/2000 18:04:24]

Error Analysis

Next: The Test Suite Up: Evaluation of Engine Previous: Accuracy

Error Analysis
Rather than using broad indicators as guides to score assignments, you could focus on the
errors the MT system makes. The technique of error analysis tries to establish how seriously
errors affect the translation output.
The method is this. To start off, write down a large list of all the types of errors you think the
MT system might make. During the evaluation, all the errors in the translated text are counted
up. Because you consider some errors more serious than others, each type of error will be
multiplied by some weighting factor which you assign to it. The score then for each individual
sentence or the whole text will be the sum of all the weighted errors. So, if we take the raw
translation we were using in the scenario in Chapter as an example, error analysis might
work as follows.
For the example three sorts of error are identified. These three sorts are errors involving
selection of a vs one as the translation of German ein, errors in number agreement (e.g. * a
computers), and errors in the selection of prepositions. Using some short codes for each error
type, each error occurrence is marked up in the raw output. The resulting marked text is given
below.

To calculate the seriousness of the errors, weights in the range 0 to 1 are assigned to the three
error types. The weight for an error in preposition selection is higher than that for incorrect
number because the person responsible considers that incorrect number is relatively less
serious. This is summarized in the following table.

On the basis of this the total error score can be calculated. There are two errors in NUMber
agreement , two involving PREPositions, and one involving A/ONE selection, so the score is:
Although this method gives more direct information on the usefulness of an MT system, there
are immediate problems with using detailed error analysis. The first is practical: it will usually
require considerable time and effort to train scorers to identify instances of particular errors ---
and they will also need to spend more time analysing each output sentence. Second, is there any
good basis for choosing a particular weighting scheme? Not obviously. The weighting is in
some cases related to the consequences an error has for post-editing : how much time it will
take to correct that particular mistake. In some other cases it merely reflects how badly an error
affects the intelligibility of the sentence. Consequently, the result will either indicate the size of
the post-editing task or the intelligibility of the text, with its relative usefulness. In both cases
devising a weighting scheme will be a difficult task.

http://privatewww.essex.ac.uk/~doug/book/node80.html (1 de 2) [20/09/2000 18:04:26]

Error Analysis

There is, however, a third problem and perhaps this is the most serious one: for some MT
systems, many output sentences are so corrupted with respect to natural language correlates that
detailed analysis of errors is not meaningful. Error types are not independent of each other:
failure to supply any number inflection for a main verb will often mean that the subject and
verb do not agree in number as required. It will be difficult to specify where one error starts
and another ends and thus there is the risk of ending up with a general error scale of the form
one, two, .... lots. The assignment of a weighting to such complex errors is thus a tricky
business.

Next: The Test Suite Up: Evaluation of Engine Previous: Accuracy

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node80.html (2 de 2) [20/09/2000 18:04:26]

The Test Suite

Next: Operational Evaluation Up: Evaluation of Engine Previous: Error Analysis

The Test Suite

As we noted before, for some years the trend (at least in research circles) has been towards
Translation Engines with substantial linguistic knowledge in the form of grammars. LK
Engines have a different performance profile from Transformer Engines in that their output
will tend to contain rather fewer badly degraded sentences. (Perhaps at the price of failing to
produce anything in some cases).
Although the use of linguistic-knowledge based techniques tends to promote higher
Intelligibility (and Accuracy) output, it is possible that the linguistic knowledge embedded in
the system is defective or incomplete. Sometimes a certain grammar rule is too strict or too
general to apply correctly in all circumstances; sometimes the rules that handle one
phenomenon (e.g. modal verbs like may in The printer may fail) and the rules that handle
another phenomenon (eg. negation) fail to work correctly together when the two phenomena
co-occur or interact in a sentence. (For example, imagine the problems that will result if The
printer can not be cleaned (i.e. can be left uncleaned), and The printer cannot be cleaned (i.e.
must not be cleaned) are confused.)
Keeping track of these sorts of constructional errors and deficits has become rather a severe
problem for developers of MT systems and other large NLP systems. For example, while
running the system on a corpus of test texts will reveal many problems, many potential areas of
difficulty are hidden because the statistics are such that even quite large corpora will lack even
a single example of particular grammatical combinations of linguistic phenomena.
Rather than churning through increasingly large `natural' text corpora, developers have recently
turned their attention to the use of suites of specially constructed test sentences. Each sentence
in the suite contains either one linguistic construction of interest or a combination thereof. Thus
part of an English test suite might look as follows.

This fragment just churns through all combinations of modal verbs like can, may together with
optional not. In practice, one would expect test suites to run to very many thousands of
sentences, because of the many different combinations of grammatical phenomena that can
occur. Suites may include grammatically unacceptable sentences (e.g. * John not run) which
the parser should recognize as incorrect. In systems which use the same linguistic knowledge
for both analysing and synthesising text, the fact that an ill-formed sentence is rejected in
analysis suggests that it is unlikely to be constructed in synthesis either.
Nobody knows for sure how test suites should be constructed and used in MT. A bi-directional
system (a system that not only translates from German to English and from English to German)
will certainly need test suites for both languages. Thus success in correctly translating all the
sentences in a German test suite into English and all the sentences in an English test suite into
German would definitely be encouraging. However, standard test suites are rather blunt

http://privatewww.essex.ac.uk/~doug/book/node81.html (1 de 3) [20/09/2000 18:04:27]

The Test Suite

instruments for probing translation performance in the sense that they tend to ignore typical
differences between the languages involved in translation.
We can look at an example. In English the perfect tense is expressed with the auxiliary verb
have, like in He has phoned. In German however there are two auxiliary verbs for perfect tense:
haben and sein. Which verb is used depends on the main verb of the sentence: most require the
first, some require the second. So an English and a German test suite designed to check the
handling of perfect tense will look different.

The German test suite thus tests the perfect tense for verbs that take sein and verbs that take
haben and therefore have to test twice the number of sentences to test the same phenomenon.
However, if He has phoned is correctly translated into German Er hat angerufen, then we still
can not be sure that all perfect tenses are translated correctly. For testing of the English
grammar alone, there is no reason to include a sentence like He has gone into the English test
suite, since the perfect tense has already been tested. For translation into German however it
would be interesting to see whether the auxiliary verb sein is selected by the main verb gehen,
giving the correct translation Er ist gegangen.
Given this sort of problem, it is clear that monolingual test suites should be supplemented with
further sentences in each language designed to probe specific language pair differences. They
could probably be constructed by studying data which has traditionally been presented in books
on comparative grammar.
In a bi-directional system, we need test suites for both languages involved and test suites
probing known translational problems between the two languages. Constructing test suites is a
very complicated task, since they need to be complete with regard to the phenomena occurring
in the present and future input texts of the MT user. Thus one should first check whether there
are any existing test suites for the languages that need to be tested. (There are several
monolingual test suites around). Such a suite can be modified by adding material and removing
restrictions that are irrelevant in the texts for which the system is intended (eg. the texts to be
translated might not contain any questions). As far as we know there are no readily available
test suites for translational problems between two languages; to test for this, the evaluator will
have to adapt existing monolingual ones.
Once the test suites have been devised they are run through the system and an inventory of
errors is compiled. Clearly the test suite is an important tool in MT system development. How
useful will it be for a user of MT systems?
It is of course possible for the user to run an MT system on a test suite of her own devising and,
in some cases, this may be perfectly appropriate. It is especially useful to measure
improvements in a system when the MT vendor provides a system update. However, the test
suite approach does entail some drawbacks when used to assess system performance in
comparison with competing systems. The problem is familiar by now: how are the evaluation
results to be interpreted? Suppose System A and System B both produce acceptable translations
for 40% of the test sentences and that they actually fail on different, or only partially
overlapping, subsets of sentences. Which one is better? If System B (but not System A) fails on
test sentences which embody phenomena with very low frequencies in the user's type of text
material, then clearly System B is the better choice. But users typically do not have reliable
information on the relative frequencies of various types of constructions in their material, and it

http://privatewww.essex.ac.uk/~doug/book/node81.html (2 de 3) [20/09/2000 18:04:27]

The Test Suite

is a complex task to retrieve such information by going through texts manually (automated tools
to do the job are not yet widely available).
The same problem of interpretability holds when MT systems are evaluated by an independent
agency using some sort of standard set of test suites. Published test suite information certainly
gives a much better insight into expected performance than the vague promisory notes offered
with current systems; but it doesn't immediately translate into information about likely
performance in practice, or about cost effectiveness.
On top of this there is the problem of how to design a test suite, and the cost of actually
constructing it. Research is ongoing to determine what sort of sentences should go into a test
suite: which grammatical phenomena should be tested and to what extent should one include
co-occurrence of grammatical phenomena, should a test suite contain sentences to test semantic
phenomena and how does one test translation problems? These and additional problems might
be solved in the future, resulting in proper guidelines for test suite construction.

Next: Operational Evaluation Up: Evaluation of Engine Previous: Error Analysis

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node81.html (3 de 3) [20/09/2000 18:04:27]

Operational Evaluation

Next: Summary Up: Evaluating MT Systems Previous: The Test Suite

Operational Evaluation
In the previous sections we have discussed various types of quality assessment. One mayor
disadvantage of quality assessment for MT evaluation purposes, however, is the fact the overall
performance of an MT system has to be judged on more aspects than translation quality only.
The most complete and direct way to determine whether MT performs well in a given set of
circumstances is to carry out an operational evaluation on site comparing the combined MT and
post-editing costs with those associated with pure human translation. The requirement here is
that the vendor allows the potential buyer to test the MT system in her particular translation
environment. Because of the enormous investment that buying a system often represents,
vendors should allow a certain test period. During an operational evaluation a record is kept of
all the user's costs, the translation times and other relevant aspects. This evaluation technique is
ideal in the sense that it gives the user direct information on how MT would fit in and change
the existing translation environment and whether it would be profitable.
Before starting up the MT evaluation the user should have a clear picture of the costs that are
involved in the current set-up with human translation. When this information on the cost of the
current translation service is available the MT experiment can begin.
In an operational evaluation of MT time plays an important role. Translators need to be paid
and the more time they spend on post-editing MT output and updating the system's
dictionaries, the less profitable MT will be. In order to get a realistic idea of the time needed for
such translator tasks they need to receive proper training prior to the experiment. Also, the MT
system needs to be tuned towards the texts it is supposed to deal with.
During an evaluation period lasting several months it should be possible to fully cost the use of
MT, and at the end of the period, comparison with the costs of human translation should
indicate whether, in the particular circumstances, MT would be profitable in financial terms or
not.
One problem is that though one can compare cost in this way, one does not necessarily hold
quality constant. For example, it is sometimes suspected that post-edited MT translations tend
to be of inferior quality to pure human translations because there is some temptation to
post-edit only up to that point where a correct (rather than good) translation is realised. This
would mean that cost benefits of MT might have to be set against a fall in quality of translation.
There are several ways to deal with this. One could e.g. use the quality measurement scales
described above (Section ). In this case we would need a fine-grained scale like in the
ALPAC Report, since the differences between post-edited MT and HT will be small. But what
does this quality measurement mean in practice? Do we have to worry about slight differences
in quality if after all an `acceptable' translation is produced? Maybe a better solution would be
to ask an acceptability judgment from the customer. If the customer notices a quality decrease
which worries him, then clearly post-editing quality needs to be improved. In most cases,

http://privatewww.essex.ac.uk/~doug/book/node82.html (1 de 2) [20/09/2000 18:04:28]

Operational Evaluation

however, the experienced translator/post-editor is more critical towards translation quality than
the customer is.
In general it seems an operational evaluation conducted by a user will be extremely expensive,
requiring 12 personmonths or more of translator time. An attractive approach is to integrate the
evaluation process in the normal production process, the only difference being that records are
kept on the number of input words, the turnaround time and the costs in terms of time spent in
post-editing. The cost of such an integrated operational evaluation is obviously less. After all, if
the system is really good the translation costs will have been reduced and will compensate for
some of the costs of the evaluation method. (On the other hand, if the system is not an
improvement for the company, the money spent on its evaluation will be lost of course.)

Next: Summary Up: Evaluating MT Systems Previous: The Test Suite

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node82.html (2 de 2) [20/09/2000 18:04:28]

Summary

Next: Further Reading Up: Evaluating MT Systems Previous: Operational Evaluation

Summary
The purchase of an MT system is in many cases a costly affair and requires careful
consideration. It is important to understand the organizational consequences and to be aware of
the system's capacities. Unfortunately, it is not possible to draw up a comparison table for MT
systems on the basis of which MT buyers could choose their system. Although system
specifications can provide us with some useful information there are too many aspects which
influence the performance of MT that cannot be included in such a table. Furthermore, MT will
perform differently in different translation environments, depending mainly on the character of
the typical input texts. Without having the necessary information of the kind of input texts the
user has in mind, it is not possible to make a reliable prediction about the cost effectiveness of
an MT system. The consequences are that if we want information about an MT system we have
to evaluate it, and that this evaluation has to be specifically for the user's translational needs.
The evaluation strategies discussed in this chapter are strategies that a buyer might want to
pursue when considering the purchase of an MT system. Although they will provide the client
with a certain amount of useful information, each method has some drawbacks, which we have
tried to point out in our discussion.

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node83.html [20/09/2000 18:04:28]

Further Reading

Next: New Directions in Up: Evaluating MT Systems Previous: Summary

Further Reading
Useful discussion of evaluation methods can be found in [van Slype1982], and [Lehrberger
and Bourbeau1987]. Practical discussion of many different aspects of MT evaluation can be
found in [King and Falkedal1990], [Guida and MauriJuly 1986], and [Balkan et al. 1991].

A special issue of the Journal Machine Translation is dedicated to issues of evaluation of MT

(and other NLP) systems. The introduction to the issue, [Arnold et al. in press b], gives an
overview of the state of the issues involved, going into more detail about some issues glossed
over here. Several of the articles which appear in this issue report practical experience of
evaluation, and suggest techniques (for example, [Albisserin press,Flank et al. in press,Jordanin
press,Neal et al. in press].)

The problems of focusing evaluation on the MT engine itself (i.e. apart from surrounding
peripherals) are discussed in [Krauwerin press].

As things stand, evaluating an MT system (or other NLP system) involves a great deal of
human activity, in checking output, for example. A method for automating part of the
evaluation process is described in [Shiwenin press].

Some of the issues involved in construction of test suites are discussed in [Arnold et al. in
press a], and [Nerbonne et al. in press].

In this chapter, we have generally taken the users' perspective. However, evaluation is also an
essential for system developers (who have to be able to guage whether, and how much, their
efforts are improving a system). How evaluation technique can be applied so as to aid
developers discussed in [Minnisin press].

One of the best examples of MT evaluation in terms of rigour was that which formed the basis
of the ALPAC report [Pierce and Carroll1966], which we mentioned in Chapter (it is
normal to be rude about the conclusions of the ALPAC report, but this should not reflect on the
evaluation on which the report was based: the evaluation itself was a model of care and rigour
--- it is the interpretation of the results for the potential of MT which was regrettable).
See [Nagao1986, page 59,] for more detailed scales and criteria for evaluating fidelity and
ease of understanding.
As usual, Hutchins and Somers [Hutchins and Somers1992] contains a useful discussion of
evaluation issues (Chapter 9).

http://privatewww.essex.ac.uk/~doug/book/node84.html (1 de 2) [20/09/2000 18:04:29]

Further Reading

Next: New Directions in Up: Evaluating MT Systems Previous: Summary

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node84.html (2 de 2) [20/09/2000 18:04:29]

New Directions in MT

Next: Introduction Up: Machine Translation Previous: Further Reading

New Directions in MT

● Introduction
● Rule-Based MT
❍ Flexible or Multi-level MT
❍ Knowledge-Based MT
❍ Feasibility of General Purpose Rule-Based MT Systems
● Resources for MT
● Empirical Approaches to MT
❍ Example-Based Translation
❍ Statistical MT
● Summary
● Further Reading

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node85.html [20/09/2000 18:04:30]

Introduction

Next: Rule-Based MT Up: New Directions in Previous: New Directions in

Introduction
In the previous chapters, we have tried to give an idea of what is currently possible in MT. In
this chapter, we look to the future. Our aim is to give a flavour of current research in MT,
indicating what issues are receiving attention and what techniques are thought to be promising.
Of course, not all the ideas that are currently important are really new ones. A great deal of
current research is directed at how familiar techniques can be improved --- for example, how
standard `Linguistic Knowledge' approaches can be improved by using better linguistic
analyses (analyses based on better linguistic theories, or a better understanding of existing
theories), and developing or adapting more efficient processing methods, and better tools for
use in constructing and modifying systems. Likewise, an important feature of current research
involves work on sublanguage MT (cf. Chapter ), but though the design of tools to aid
sublanguage analysis is an increasingly important area, it is hardly a new development. Other
currently important work is concerned with integration, which can relate either to the
integration of MT with other Natural Language Processing technologies, or to the (non-trivial)
problems of integration of MT into general document processing technology that arise as one
tries to make a practically and commercially usable system out of a research prototype MT
system. A particularly important example of the former is research on `speech-to-speech' MT
systems --- that is, systems that can take spoken input, and produce spoken output (e.g. for more
or less simultaneous interpreting of telephone conversations). Such work is clearly important,
and often throws up interesting differences of emphasis (for example, in speech-to-speech
work, there is an emphasis on speed, and on dealing with sentence fragments, since one would
like to be able to translate each utterance as it is spoken, without waiting for the end. This gives
importance to `bottom up' methods of analysis, and severe restrictions on the input in terms of
text-type , etc). However, there is an obvious sense in which such work it is `more of the same'
--- it involves improving one aspect of an existing idea, rather than presenting a genuinely new
direction, and would be accessible on the basis of the earlier chapters of this book. In this
chapter, we will concentrate on what we think may turn out to be more radical ideas.

The chapter has three main sections. In Section , we outline some current issues and trends
in the design of sets of linguistic rules for MT, that is, work within the established `Linguistic
Knowledge', or `Rule-Based' paradigm. The next section ( ) gives an overview of some of the
corpus and machine readable dictionary resources which have recently become available.
These resources have stimulated a great deal of research within the traditional LK/rule-based
paradigm, and have also been of key importance in the trend towards so-called empirical
approaches to MT, which are sketched in Section .

http://privatewww.essex.ac.uk/~doug/book/node86.html (1 de 2) [20/09/2000 18:04:31]

Introduction

Next: Rule-Based MT Up: New Directions in Previous: New Directions in

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node86.html (2 de 2) [20/09/2000 18:04:31]

Rule-Based MT

Next: Flexible or Multi-level Up: New Directions in Previous: Introduction

Rule-Based MT

● Flexible or Multi-level MT
● Knowledge-Based MT
● Feasibility of General Purpose Rule-Based MT Systems

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node87.html [20/09/2000 18:04:32]

Flexible or Multi-level MT

Next: Knowledge-Based MT Up: Rule-Based MT Previous: Rule-Based MT

Flexible or Multi-level MT
Most transfer or interlingual rule-based systems are based on the idea that success in
practical MT involves defining a level of representations for texts which is abstract enough to
make translation itself straightforward, but which is at the same time superficial enough to
permit sentences in the various source and target languages to be successfully mapped into that
level of representation. That is, successful MT involves a compromise between depth of
analysis or understanding of the source text, and the need to actually compute the abstract
representation. In this sense, transfer systems are less ambitious than interlingual systems,
because they accept the need for (often quite complex) mapping rules between the most abstract
representations of source and target sentences. As our linguistic knowledge increases, so too
MT systems based on linguistic rules encoding that knowledge should improve. This position is
based on the fundamental assumption that finding a sufficiently abstract level of representation
for MT is an attainable goal. However, some researchers have suggested that it is not always the
case that the deepest level of representation is necessarily the best level for translation.
This can be illustrated easily by thinking about translation between closely related languages
such as Norwegian and Swedish .
● ❍ Min nya bil är blå (Swedish)
`my new car is blue'
❍ Den nye bilen min er blå (Norwegian)
`the new car mine is blue'

● ❍ Var har du hittat en så ful slips? (Swedish)

`Where did you find a such ugly tie'
❍ Hvor har du funnet et så stygt slips? (Norwegian)
`Where did you find a such ugly tie'
In the second example here, both languages have exactly the same word order, although the
words themselves and their grammatical features differ. In the first example, we see that
Swedish (like English) does not allow the use of an article together with a possessive pronoun,
which Norwegian (like, say, Italian) does. These are certainly minimal differences, and it would
be a serious case of overkill to subject the source language sentences to `in depth' analysis ,
when essentially all that is required to deal with this structural difference is to express a
correspondence between the structures described by the following syntactic rules (here `Poss'
stands for `Possessive pronoun').
(Swedish) NP Poss Adj N (Norwegian) NP Det Adj N
Poss
Of course, it would be straightforward to design a special purpose MT system which was
equipped only with the sort of linguistic rules required to perform this type of superficial

http://privatewww.essex.ac.uk/~doug/book/node88.html (1 de 5) [20/09/2000 18:04:36]

Flexible or Multi-level MT

manipulation of syntactic structures. But a number of considerations, not least economic

considerations, militate against this. Instead one could conclude that what is required is an
approach to rule-based translation which is sufficiently flexible to carry out deep analysis only
when required, so that the same MT engine can be used for dealing with pairs of closely related
languages and pairs of languages which differ greatly. Such ideas lie behind attempts to design
flexible systems which can operate in a variety of modes, according to the depth of analysis
required for the language pair, or even the particular examples in hand.
There are other reasons for the current interest in flexible systems. In the example above, we
have tried to show that what is the `appropriate level' of analysis for one language pair might be
quite inappropriate for another pair. But some researchers have pointed out that a similar
situation obtains within one and the same language pair. Though really convincing arguments
are hard to find, the idea is that translation seems to depend on information about different
levels of linguistic information at the same time. For example, for most translation purposes, as
we have noted previously, a representation in terms of semantic relations (AGENT, PATIENT,
etc.) is attractive. However, such a representation will probably not distinguish between ( a), (
b) and ( c). This means they will be translated alike, if this is the representation that is
produced by analysis . But in many cases this would not produce a very good translation.
● ❍ Sam broke the printer.

❍ It was the printer that Sam broke.

❍ It was Sam that broke the printer

Ideally, what one wants is a semantic account of the differences between these examples. This
has to do with the difference between what is presupposed, and what is asserted, or what is
treated as `given', and what as new information (e.g. in ( b) it is presupposed that Sam broke
something, and stated that the thing in question was the printer). Producing such an account is
not impossible, and may indeed produce a better MT system in the long run. However, it is by
no means easy, and, at least in the short term, it would be nice if one could use information
about semantic relations where that is useful, and information about surface syntactic form
where that was useful. This would be possible if one had a way of allowing information from a
variety of levels to be referred to in transfer . Of course, the difficulty then would be to allow
this flexibility while still ensuring that the pieces of information can be correctly combined to
give a suitable target translation.
There are various proposals in the MT literature concerning flexible MT. Some researchers
working within the paradigm of example-based MT , which we discuss below, have proposed
architectures which are flexible with respect to the level at which translation occurs. Another
rather radical idea depends on the fact that several contemporary linguistic theories provide a
`multidimensional' characterisation of a linguistic string. One can get a flavour of what is
involved by looking at the following representation.

Figure: A Multidimensional Representation

This representation of the sentence Kim walks is multidimensional, in the sense that it contains
information about several levels, or dimensions, of structure at the same time: information
about ORTHography , SYNtax , SEMantics , and constituent structure (the DaughTeRs
feature). Such multidimensional representations are known as signs. Identity of values is
indicated by tags, boxed indices like , .

http://privatewww.essex.ac.uk/~doug/book/node88.html (2 de 5) [20/09/2000 18:04:36]

Flexible or Multi-level MT

If we look first at the DTRS value, we can see that there are two daughters, the first an NP (i.e.
whose SYNtax contains an attribute CAT with value NP), and the second a VP. The NP has no
daughters, and the VP has one daughter, whose category is V. The ORTHography of the whole
S is made up of , the ORTHography of the NP, i.e. mary, and the ORTHography of the VP,
which is identical to the ORTHography of the V, tagged . The TNS (TeNSe) of S, VP, and V
are identical, and the NP, VP, and V have the same NUMber value.
The semantics of the S indicates that the argument of the predicate is the value tagged , that
is, the semantics of the NP, .
We have seen that representation carries information about ORTHography, SYNTax,
SEMantics and daughters (DTRS) at the same time (a fuller representation would include
information about morphology too). Formally, it is just a collection of features (i.e. attributes
and values) of the kind we have seen before, with the difference that the value of some of the
attributes can be an entire structure (collection of features), and we allow different attributes to
have the same value (indicated by means of a tag, a number written in a box). This is
sometimes called a re-entrance.
The syntactic information is essentially equivalent to the sorts of category label we have seen
before, and the value of the DTRS attribute simply gives the values of the daughters a node
would have in a normal consituent structure tree of the kind that were given in Chapter . One
interesting point to note is that there is a value for SEMantics given for the mother sign, and for
every one of the daughter signs. (In fact, the SEM value of the S is identical to the SEM value
of the VP, and the V, and the SEM value of the AGENT of the S is identical to the SEM value
of the NP Kim.)
One way one could use such a structure would be just to take the value of the SEM attribute for
the mother sign in the output of analysis , and input this value to transfer (in a transfer system )
or synthesis (in an interlingual system ). This would involve only adapting the techniques we
described in earlier chapters for transfer and synthesis to deal with complex attribute-value
structures, rather than trees (this is not very difficult). Of course, this would mean that one was
losing any benefit of multidimensionality for translation (though one might be able to exploit it
in analysis ).
If one is to exploit multidimensionality in transfer or synthesis (which was the aim) the only
possible part of the sign to recurse through, applying rules, is the structure of the DTRS
attribute. However, as we noted, this is just the surface phrase structure, enhanced with some
information about semantics and orthography. If this is so, then one might wonder whether any
advantage has been gained at all.
The solution is not to think in terms of applying rules to representations or structures at all, but
to focus on the attribute-value structure as simply a convenient graphic representation of the
solution to a set of constraint s. For example, for the representation on page , one such
constraint would be that the CATegory value of the mother sign is S. More precisely, the value
of SYN on the mother sign is an attribute-value structure which contains an attribute CAT, with
value S. That is, if we give names like X0, X1, X2, etc. to the various attribute-value structures,
with X0 the name of the mother sign, then the value of SYN in X0 is a structure X1, and the
value of CAT in X1 is S:
X0:SYN = X1 X1:CAT = S

http://privatewww.essex.ac.uk/~doug/book/node88.html (3 de 5) [20/09/2000 18:04:36]

Flexible or Multi-level MT

If we name the attribute-value structure of the VP X4, and that of the V X5, we also have the
following, indicating that S, VP, and V all have the same SEM values.
X0:SEM = X4:SEM X4:SEM = X5:SEM
The value of the ORTHography attribute in X0 is the concatenation of the values in the NP
(X6) and the VP (X5):
X0:ORTH = concatenation(X6:ORTH, X5,ORTH)

One can think of a representation like that on page as simply a graphic representation of the
solution to a set of such equations, and one can use the equations as the basis for translation, in
the following way. First, it is the task of analysis to produce the equation set. This is not, in
fact, difficult --- we have already seen, in Chapter how one can add instructions to grammar
rules to create different kinds of representations. Using them to create sets of equations is a
simple extension of this idea. This set of constraints describes a source structure. The
translation problem is now to produce a set of constraints whose solution will yield a target
language structure. Ultimately, of course, one is interested in the ORTH value in such a
structure, but in the meantime, one can state constraints such as: ``the SEM of the source
structure, and the SEM of the target structure must be identical'' (this assumes that the SEM
values are `interlingual'), or ``the SEM of the target structure must be the result of applying
some `transfer' function to the SEM of the source structure''. But one can easily state
constraints in terms of other attributes, for example, ``in the case of proper nouns, the value of
ORTH in the source structure and the value of ORTH in the target structure must be the same''.
Similarly, if we add attributes and values giving information about grammatical relations such
as subject, etc. into the constraints, we can state constraints in terms of these.
Of course, we cannot, in this way, guarantee that we will deal with all of the source structure
(we may leave parts untranslated by failing to produce appropriate target language constraints),
or that solving the target language constraints will produce a single target structure, or even any
structure at all (the constraints may be inconsistent). Nor have we indicated how the constraints
are to be solved. Moreover, one will often not want such constraints to be observed absolutely,
but only by default. For example, proper names should only keep the same orthography form if
there is no constraint that says otherwise (in translating English into French , one would like to
ensure that London translates as Londres). There are a number of serious difficulties and open
research questions here. However, one can get a feeling for a partial solution to some of these
problems by considering the following rather simple approach.
Recall that the constraints we gave above made the SEMantics of the S equal to the SEMantics
of the VP, and the V. One may immediately think of this as involving the V contributing its
SEMantics to the S, but one can also see it the other way round, as putting the semantics of the
whole S `into' the V. What this means, of course, is that all the semantic information conveyed
by the sentence is represented (somewhat redundantly) in the representations of the words. Now
suppose that we have translation constraints which say, for example, that the translation of the
word walk must be the word marcher, with the same semantics, and that the translation of Sam
must be Sam, again with the same semantics. What we must do now is produce a target
structure. The problem we have is interestingly like the problem we have when we try to parse a
sentence: then we typically know what the words are, and what order they are in, but not what
the sentence as a whole means; here we know what the words are, and what the sentence as a
whole means (it is represented `in the words'), but not what the word order should be. One
possibility is simply to use the target grammar to parse Sam, and marcher in all possible orders.

http://privatewww.essex.ac.uk/~doug/book/node88.html (4 de 5) [20/09/2000 18:04:36]

Flexible or Multi-level MT

To take a slightly more interesting case, suppose the source sentence is ( ):

● sees London.

If the target language is French , the target grammar will be asked to parse the strings in ( ):
● ❍ *voit Sam Londres.

❍ ?Londres voit Sam.

❍ *Sam Londres voit.

❍ Sam voit Londres.

One can expect the target grammar to reject ( a), and ( c). It would accept ( b), but only
with the meaning that is different from that of the source sentence, which we have carried over
in the constraints linking see to voir. This leaves only the correct solution ( d).

Next: Knowledge-Based MT Up: Rule-Based MT Previous: Rule-Based MT

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node88.html (5 de 5) [20/09/2000 18:04:36]

Knowledge-Based MT

Next: Feasibility of General Up: Rule-Based MT Previous: Flexible or Multi-level

Knowledge-Based MT

The term knowledge-based MT has come to describe a rule-based system displaying extensive
semantic and pragmatic knowledge of a domain , including an ability to reason , to some
limited extent, about concepts in the domain (the components, installation and operation of a
particular brand of laser printer could constitute a domain). We noted the appeal of such an
approach as a way of solving some basic MT problems in earlier chapters. Essentially, the
premise is that high quality translation requires in-depth understanding of the text, and the
development of the domain model would seem to be necessary to that sort of deep
understanding. One of the important considerations driving this work is an appreciation that
post-editing is time-consuming and very expensive, and therefore that efforts made to produce
high quality output will pay off in the long run. Since this may well turn out to be of great
utility, in this section we concentrate on an approach which attempts some degree of text
understanding on the basis of detailed domain knowledge, developed at the Center for Machine
Translation at Carnegie Mellon University in Pittsburgh.
To give some idea of what is at stake here, the prototype systems developed for English
Japanese translation during the late 1980s at CMU, dealing with the translation of instruction
manuals for personal computers, contained the following components:
● an ontology of concepts

● analysis lexica and grammars for English and Japanese

● generation lexica and grammars for English and Japanese

● mapping rules between the Interlingua and English/Japanese syntax

For a small vocabulary (around 900 words), some 1500 concepts were defined in detail. The
ontology dealt solely with the interaction between personal computers and their users. Nouns in
the interlingua correspond to `object concepts' in the ontology, which also contains `event
concepts', such as the event remove, corresponding to the English verb remove and the
Japanese verb torinozoku (by no means are all mappings from the interlingua into natural
language as straightforward as this, for example, the concept to-press-button must be
divided into subevents corresponding to pressing, holding down and releasing the button).
Concepts are represented in a form of frame representation language, familiar from work in
Artificial Intelligence and Natural Language Processing, in which frames (providing an intrinsic
characterisation of concepts) are linked in a hierarchical network. To give an idea of the amount
of detailed knowledge about concepts that one might want to encode, Table gives by way of
example a frame for the concept computer.

Table: Example Frame for the concept computer

http://privatewww.essex.ac.uk/~doug/book/node89.html (1 de 2) [20/09/2000 18:04:37]

Knowledge-Based MT

Knowledge-based MT is still pursued today at CMU in the KANT system, but is much more
modest in terms of its goals for domain knowledge , which is limited to that which is necessary
for stylistically adequate, accurate translation, as opposed to deep textual understanding. Thus
the domain model simply represents all the concepts relevant in the domain, but does not
support any further reasoning or inference about the concepts in the domain, other than that
which is directly encoded (e.g. hierarchical information such as the fact that personal computers
and mainframes are types of computer). The essential role of the domain model is to support
full disambig uation of the text. An important part of this is specifying, for every event concept
in the domain, what restrictions it places on the object concepts which constitute its arguments
(e.g. only living things can die, only humans can think, in a literal sense) or the `fillers' of `slots'
in its (frame-based) representation.
Once you start adding detailed knowledge in the pursuit of high quality translation through text
understanding, it is tempting to add more and more sources of knowledge. It is quite clear that
anaphora resolution and the resolution of other referential ambiguities requires reference to a
level of structure above sentential syntax and semantics (see e.g. the examples in Chapter ).
Likewise, for stylistic reasons, to increase the cohesiveness of the text, one might need to keep
some working representation of the paragraph structure. Achieving a really high quality
translation, especially with some sorts of text, might require treatment of metaphor ,
metonymy , indirect speech act s, speaker/hearer attitudes and so on. Over the last few years a
variety of groups in different parts of the world have begun experimenting with prototypes
intended to work with explicit knowledge or rule components dealing with a wide variety of
different types of information. All of these approaches can be viewed as examples, of one form
or another, of knowledge-based MT.

Next: Feasibility of General Up: Rule-Based MT Previous: Flexible or Multi-level

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node89.html (2 de 2) [20/09/2000 18:04:37]

Feasibility of General Purpose Rule-Based MT Systems

Next: Resources for MT Up: Rule-Based MT Previous: Knowledge-Based MT

Feasibility of General Purpose Rule-Based MT

Systems
The approaches to MT that we have discussed so far in this chapter can be distinguished from
each other mainly in terms of the various knowledge sources which are used in translation.
They are all straightforward rule-based approaches, as most work in MT has been until the last
few years. However it is widely recognised that there are serious challenges in building a
robust, general purpose, high quality rule-based MT system, given the current state of
linguistic knowledge. As we shall see, these problems and the increasing availability of raw
materials in the form of on-line dictionaries , termbanks and corpus resources have led to a
number of new developments in recent years which rely on empirical methods of various sorts,
seeking to minimize or at least make more tractable the linguistic knowledge engineering
problem.
One of the most serious problems, and probably the most serious problem, for linguistic
knowledge MT is the development of appropriate large-scale grammatical and lexical
resources . There are really a number of closely related problems here. The first is simply the
scale of the undertaking, in terms of numbers of linguistic rules and lexical entries needed for
fully automatic, high quality MT for general purpose and specialised language usage. Even
assuming that our current state of linguistic knowledge is sophisticated enough, the effort
involved is awesome, if all such information must be manually coded. It is generally accepted,
then, that techniques must be adopted which favour the introduction of semi-automatic and
automatic acquisition of linguistic knowledge.
The second concerns the difficulties of manipulating and managing such knowledge within a
working system. The experience of linguists developing a wide variety of natural language
processing systems shows that it is all too easy to add ad hoc, specially crafted rules to deal
with problem cases, with the result that the system soon becomes difficult to understand,
upgrade and maintain. In the worst case, the addition of a new rule to bring about some
intended improvement, may cause the entire edifice to topple and performance to degrade. To a
certain extent, these familiar problems can be avoided by adopting up to date formalisms, and
restricting the use of special devices as much as possible. It is also very important to do
everything possible to ensure that different grammar writers adopt essentially the same or
consistent approaches and document everything they do in detail.
The third issue is one of quality and concerns the level of linguistic detail required to make the
various discriminations which are necessary to ensure high quality output, at least for general
texts. This problem shows up in a number of different areas, most notably in discriminating
between different senses of a word, but also in relating pronouns to their antecedents.
Some consider that this third aspect is so serious as to effectively undermine the possibility of
building large scale robust general purpose MT systems with a reasonably high quality output,
arguing that given the current state of our understanding of (especially) sense differences, we

http://privatewww.essex.ac.uk/~doug/book/node90.html (1 de 3) [20/09/2000 18:04:38]

Feasibility of General Purpose Rule-Based MT Systems

are at the limits of what is possible for the time being in terms of the explicit encoding of
linguistic distinctions. An extremely radical approach to this problem is to try to do away with
explicitly formulated linguistic knowledge completely. This extreme form of the `empirical'
approach to MT is found in the work carried out by an MT group at IBM Yorktown Heights
and will be discussed in the section below on Statistical Approaches.
One interesting development is now evident which receives its impetus from the appreciation of
the difficulty and costliness of linguistic knowledge engineering. This is the growth of research
into the reusability of resources (from application to application and from project to project)
and the eventual development of standards for common resources. One of the reasons why this
is happening now is that there is undoubtedly a set of core techniques and approaches which are
widely known and accepted within the Natural Language Processing research community. In
this sense a partial consensus is emerging on the treatment of some linguistic phenomena. A
second important motivation is a growing appreciation of the fact that sharing tools, techniques
and the grammatical and lexical resources between projects, for the
areas where there is a consensus, allows one to direct research more appropriately at those
issues which pose challenges.
As well as the various difficulties in developing linguistic resources, there are other issues
which must be addressed in the development of a working MT system. If a system is to be used
on free text, then it must be robust. That is, it must have mechanisms for dealing with unknown
words and ill-formed output (simply answering `no' and refusing to proceed would not be
cooperative behaviour). In a similar way, it must have a way of dealing with unresolved
ambiguities, that is, cases in which the grammar rules, in the light of all available information,
still permit a number of different analyses. This is likely to happen in terms of both lexical
choice (for example, where there are a number of alternatives for a given word in translation)
and structural choice. For example, taken in isolation (and in all likelihood, even in many
contexts) the following string is ambiguous as shown:
● ❍ Sam told Kim that Jo had died last week.

❍ Sam told Kim [that Jo had died] last week.

❍ Sam told Kim [that Jo had died last week].

Such attachment ambiguities with adverbial phrases (such as last week) and prepositional
phrases ( on Tuesday) occur quite frequently in a language like English in which PPs and
ADVPs typically occur at the end of phrases. In many cases, they are strictly structurally
ambiguous, but can be disambiguated in context by the hearer by using real-word knowledge.
For example, the following is ambiguous, but the hearer of such a sentence would have enough
shared knowledge with the speaker to chose the intended interpretation (and perhaps would not
even be aware of the ambiguity ):
● ❍ Joe bought the book that I had been trying to obtain for Susan.

❍ [Joe bought [the book that I had been trying to obtain for Susan]].

❍ [Joe bought [the book that I had been trying to obtain] for Susan].

Consideration of issues such as these underlies work in integrating core MT engines with
spelling checkers, fail-safe routines for what to do when a word in the input is not in the
dictionary and adding preference mechanisms which chose an analysis in cases of true
ambiguity , but an appreciation of the serious nature of these issues has also provided an
motivation for the current interest in empirical , corpus or statistical-based MT, to which we
return after discussing the question of resources for MT.

http://privatewww.essex.ac.uk/~doug/book/node90.html (2 de 3) [20/09/2000 18:04:38]

Feasibility of General Purpose Rule-Based MT Systems

Next: Resources for MT Up: Rule-Based MT Previous: Knowledge-Based MT

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node90.html (3 de 3) [20/09/2000 18:04:38]

Resources for MT

Next: Empirical Approaches to Up: New Directions in Previous: Feasibility of General

Resources for MT
As researchers begin to consider the implications of developing their systems beyond the
level of proof-of-concept research prototypes with very restricted coverage, considerable
attention is being paid to the role that existing bilingual and monolingual corpus and lexical
resources can play. A corpus is essentially a large collection of texts, but for our purposes we
are interested only in such texts stored on computers in a standard format (e.g. extended
ASCII). Such texts may often contain standard markup (e.g. in SGML ) and for most practical
purposes one needs a set of corpus access tools for retrieving data at will.
Various research centres throughout the world have been developing monolingual corpus
resources for many years, and there has been a growing awareness throughout the eighties of
their importance to linguistic and lexicographic work. A number of sites hold substantial corpus
resources (several millions of words), an example being the Unit for Computer Research on the
English Language at the University of Lancaster which currently holds in excess of 5 million
words of corpus material, of which 4M words have been tagged with part-of-speech
information. Such collections are a rich repository of information about actual language usage.
Efforts are underway at different centres to (automatically or semi-automatically) annotate
corpus resources with various types of linguistic information, in addition to grammatical (POS)
tagging, prosodic annotation (indicating features of stress and annotation), syntactic tagging
(indicating phrasal groups of words, i.e. parsing or partial (skeleton) parsing); semantic tagging
and discourse level tagging (indicating anaphoric and other similar links). To give some idea
of scale, the planned British National Corpus will contain around 100M words of
grammatically tagged corpus material, with standard SGML markup. The following example
text has been tagged with the CLAWS tagset developed at UCREL, University of Lancaster
--- in cases where multiple tags are possible, the tag chosen by the probabilistic tagger is shown
in square brackets, with the alternatives following after commas.

These tags, which it must be stressed are assigned completely automatically and with a high
level of accuracy, provide a detailed parts of speech analysis of the text, distinguishing between
some 40 different subcategories of Noun (the tags for Nouns begin with N for Noun or P for
pronoun) and some 30 different subcategories of Verb, and so on.
Over the last few years there has been an increasing awareness of the importance of corpus
resources in MT research. Tools for extracting information automatically from texts are being
increasingly used, and new techniques developed. At the simplest level, a monolingual corpus
is a crucial tool for the linguist in determining language usage in a given domain, and a
bilingual corpus for determining the facts of translation. In developing MT systems, bilingual
texts are an extremely important resource, and they are most useful if organized in such a way
that the user can view translation `chunks' or `units'. In bitext (or `multitext') the text is aligned
so that within each bilingual (or multilingual) chunk the texts are translations of each other. The

http://privatewww.essex.ac.uk/~doug/book/node91.html (1 de 3) [20/09/2000 18:04:40]

Resources for MT

most common form of alignment takes the sentence to be the organizing unit for chunking and
techniques exist for performing this alignment of bitext automatically with a high level of
accuracy (96% or higher). Of course alignment does not need to stop at the sentence level and it
is possible to apply simple probability measures to a sentence aligned bitext to extract
automatically the most probable word pair alignments, and given some skeleton or phrasal
parsing, to attempt to extract useful information about phrasal alignment. A caveat is of course
in order --- the success of techniques such as probabilistic word pair alignment depends on the
size and quality of the corpus resource, and minimum size is probably 2M words of clean text.
The availability of bilingual or multilingual corpus resources of a decent size is currently a
limiting factor. Despite the fact that many international institutions and companies have large
bilingual or multilingual resources in appropriate formats, they have been slow to appreciate the
value of releasing these to the research community, although there are indications that this
situation is now changing (the Canadian English-French Hansard record of parliamentary
proceedings is a notable exception, see the extract on page ).

Much of the interest in corpus resources and machine-readable dictionaries comes not from
their value as static knowledge banks, which the grammar writer can consult but in the
possibilities of using the information they contain directly in the MT system, thus providing
some solution to the knowledge acquisition problem we noted above. One way this can be
achieved is by investigating procedures for automatically or semi-automatically deriving
linguistic rules for the MT system from the various sources of information. Ideas currently
under investigation include the use of monolingual corpus of sufficient size for automatic sense
disambiguation in context. As a further example, a part of speech tagged sentence aligned
bilingual text together with some probabilistic model, could be used to automatically provide
equivalent terms in the two languages which could then be automatically compiled into the
relevant formalism for lexical entries in an MT system.
A further resource which is now beginning to be adequately exploited is the machine-readable
dictionary (cf. Chapter ). Monolingual lexical entries can be constructed semi-automatically
from machine-readable dictionaries, and research is underway into semi-automatically deriving
a bilingual lexicon from these monolingual lexica by statistical comparison of the lexical
structures associated with various word senses. Another possibility is that of automatically
deriving subcategorization and semantic selectional information for lexical entries and
grammatical rules from corpus resources and machine-readable dictionaries. In all of these
applications, the knowledge banks can be used to ease the formulation of large amounts of
detailed linguistic information in a rule-based system. A number of other approaches, to which
we now turn, attempt to use the information implicit in bilingual corpora, dictionaries and
thesauri much more directly, as a component in the MT system.

Next: Empirical Approaches to Up: New Directions in Previous: Feasibility of General

Arnold D J

http://privatewww.essex.ac.uk/~doug/book/node91.html (2 de 3) [20/09/2000 18:04:40]

Resources for MT

Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node91.html (3 de 3) [20/09/2000 18:04:40]

Empirical Approaches to MT

Next: Example-Based Translation Up: New Directions in Previous: Resources for MT

Empirical Approaches to MT

Given the questions that have been raised about the feasibility of `rule-based' approaches, the
increasing availability of large amounts of machine readable textual material has been seen by a
number of research groups as opening possibilities for rather different MT architectures --- in
particular, so called `empirical' architectures which apply relatively `low-level' statistical or
pattern matching techniques either directly to texts, or to texts that have been subject to only
rather superficial analysis. The reasoning behind the term empirical is that in such approaches,
whatever linguistic knowledge the system uses is derived empirically, by examination of real
texts, rather than being reasoned out by linguists. We will look at two such approaches: the so
called `example' or `analogy' based approach, and the `statistical' approach.

● Example-Based Translation
● Statistical MT

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node92.html [20/09/2000 18:04:40]

Example-Based Translation

Next: Statistical MT Up: Empirical Approaches to Previous: Empirical Approaches to

Example-Based Translation

Throughout most of this book, we have assumed a model of the translation machine which
involves explicit mapping rules of various sorts. In the `translation by analogy', or
`example-based' approach, such mapping rules are dispensed with in favour of a procedure
which involves matching against stored example translations. The basic idea is to collect a
bilingual corpus of translation pairs and then use a best match algorithm to find the closest
example to the source phrase in question. This gives a translation template, which can then be
filled in by word-for-word translation.
This idea is sometimes thought to be reminiscent of how human translators proceed when using
a bilingual dictionary: looking at the examples given to find the source language example that
best approximates what they are trying to translate, and constructing a translation on the basis
of the target language example that is given. For example, the bilingual dictionary entry for
printer which we discussed in Chapter gave the following as examples.
● ❍ 's error faute f d'impression, coquille f;
❍ 's reader correcteur m, -trice f (d'épreuves).
Given a sentence like ( ) to translate, a human translator would certainly choose faute
d'impression or coquille as the translation, on the basis that a mistake is much more like an
error than it is like a reader.
● seems to be .

The distance calculation, to find the best match for the source phrase, can involve calculating
the closeness of items in a hierarchy of terms and concepts provided by a thesaurus . To give a
flavour of the idea, and the sort of problem it addresses, consider the problem of translating
Japanese phrases of the form A no B ( no is a particle indicating the relation between A and B)
into English. Among the forms to choose from are AB, A's B, B of A, B on A, B in A, and B for
A, cf Table which gives English paraphrases of examples involving no, together with the
correct translations for these different patterns. The problem is certainly not an esoteric one,
since the expression is claimed to occur in around 50% of Japanese sentences.

Table: Alternative Translations for the Particle no

For a given input, the system will then calculate how close it is to various stored example
translations based on the distance of the input from the example in terms of the thesaurus
hierarchy (this involves finding the `Most Specific Common Abstraction' for the input and the
alternative translations --- i.e. `closest' concept in the thesaurus hierarchy) and how `likely' the
various translations are on the basis of frequency ratings for elements in the database of

http://privatewww.essex.ac.uk/~doug/book/node93.html (1 de 3) [20/09/2000 18:04:43]

Example-Based Translation

examples. (Notice this means we assume that the database of examples is representative of the
texts we intend to translate.)
The following is an extension to this basic idea: pairs of equivalent source and target language
expression are given, along with example translations, written in parentheses, and interpreted as
stating `conditions' under which the given equivalence holds. For example, the rule for the
Japanese word sochira (`this', or `this person' --- i.e. the addressee, you), given below, indicates
that sochira translates as this when the example involves desu, (translating as be), and as you,
when the input involves something like okuru (translating as send). In translating an input like
sochira ni tsutaeru, the English pronoun you would be selected as the translation of sochira,
because tsutaeru (convey) is closest to okuru (send) in the thesaurus.
sochira
this (( desu {be}),...)
you (( okuru {send}),...)
this (( miru {see}),...)
This rule uses only information about the surrounding string, but one could imagine other sorts
of example, where information is given in terms of patterns of strings, or of grammatical
information. An example involving string patterns is
given below, which would be involved in translating examples involving the expression
o-negaishimasu along the lines of ( ) ( o-negaishimasu (`please') is a general expression
indicating that a request is being made, or a favour requested, o indicates that the preceding
noun phrase is an OBJECT).
● ❍ jinjika o o-negaishimasu.
personnel section OBJ please
May I speak to the personnel section?
❍ daimei o o-negaishimasu.
title OBJ please
Please give me the title.
To deal with this, rules like the following use information about surrounding string patterns:
X o o-negaishimasu
May I speak to X' ((jimukyoku {office}),...)
Please give me X' ((bangou {number}),...)
It should be evident that the feasibility of the approach depends crucially on the collection of
good data. However, one of the advantages of the approach is that the quality of translation
will improve incrementally as the example set becomes more complete, without the need to
update and improve detailed grammatical and lexical descriptions . Moreover, the approach can
be (in principle) very efficient, since in the best case there is no complex rule application to
perform --- all one has to do is find the appropriate example and (sometimes) calculate
distances. However, there are some complications. For example, one problem arises when one
has a number of different examples each of which matches part of the string, but where the
parts they match overlap, and/or do not cover the whole string. In such cases, calculating the
best match can involve considering a large number of possibilities.
A pure example-based approach would use no grammar rules at all, only example phrases.
However, one could also imagine a role for some normal linguistic analysis, producing a
standard linguistic representation. If, instead of being given in simple `string' form, examples

http://privatewww.essex.ac.uk/~doug/book/node93.html (2 de 3) [20/09/2000 18:04:43]

Example-Based Translation

were stated in terms of such representations (i.e. given as fragments of linguistic

representations), one would expect to be able to deal with many more variations in sentence
pattern, and allow for a certain amount of restructuring in generation. In this way, one would
have something that looked more like a standard LK architecture. The chief difference would
be in the level of specificity of the rules. In particular, where in a traditional transfer system the
rules are stated in as general a form as possible, to cover entire classes of case, what one would
have here is a system where the rules are stated in highly particular forms (each one for
essentially one case), but there is a general procedure for estimating, for each case, which rule
is most appropriate (i.e. by estimating which example is closest). Of course, what this suggests
is that there is no radical incompatibility between example-based, and rule-based approaches,
so that the real challenge lies in finding the best combination of techniques from each. Here one
obvious possibility is to use traditional rule-based transfer as a fall back, to be used only if there
is no complete example-based translation.

Next: Statistical MT Up: Empirical Approaches to Previous: Empirical Approaches to

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node93.html (3 de 3) [20/09/2000 18:04:43]

Statistical MT

Next: Summary Up: Empirical Approaches to Previous: Example-Based Translation

Statistical MT

Over the last few years there has been a growing interest in the research community in
statistical approaches to Natural Language Processing. With respect to MT, the term `statistical
approaches' can be understood in a narrow sense to refer to approaches which try to do away
with explicitly formulating linguistic knowledge, or in a broad sense to denote the application
of statistically or probablistically based techniques to parts of the MT task (e.g. as a word sense
disambiguation component). We will give a flavour of this work by describing a pure
statistical-based approach to MT.
The approach can be thought of as trying to apply to MT techniques which have been highly
successful in Speech Recognition, and though the details require a reasonable amount of
statistical sophistication, the basic idea can be grasped quite simply. The two key notions
involved are those of the language model and the translation model. The language model
provides us with probabilities for strings of words (in fact sentences), which we can denote by
(for a source sentence ) and (for any given target sentence ). Intuitively, is the
probability of a string of source words S occurring, and likewise for . The translation model
also provides us with probabilities --- is the conditional probability that a target sentence
will occur in a target text which translates a text containing the source sentence . The product
of this and the probability of S itself, that is gives the the probability of source-target pairs of
sentences occurring, written .
One task, then, is to find out the probability of a source string (or sentence) occurring (i.e. ).
This can be decomposed into the probability of the first word, multiplied by the conditional
probabilities of the succeeding words, as follows.
, etc...
Intuitively, the conditional probability is the probability that s2 will occur, given that s1 has
occurred; for example, the probability that am and are occur in a text might be approximately
the same, but the probability of am occurring after I is quite high, while that of are is much
lower). To keep things within manageable limits, it is common practice to take into account
only the preceding one or two words in calculating these conditional probabilities (these are
known respectively as `bigram' and `trigram' models). In order to calculate these source
language probabilities (producing the source language model by estimating the parameters), a
large amount of monolingual data is required, since of course the validity, usefulness or
accuracy of the model will depend mainly on the size of the corpus.
The second task requiring large amounts of data is specifying the parameters of the translation
model, which requires a large bilingual aligned corpus. As we observed above, there are rather
few such resources, however, the research group at IBM which has been mainly responsible for
developing this approach had access to three million sentence pairs from the Canadian

http://privatewww.essex.ac.uk/~doug/book/node94.html (1 de 3) [20/09/2000 18:04:46]

Statistical MT

(French-English) Hansard --- the official record of proceedings in the Canadian Parliament (cf.
the extract given above), from which they have developed a (sentence-) aligned corpus, where
each source sentence is paired with its translation in the target language, as can be seen on page
.
It is worth noting in passing that the usefulness of corpus resources depends very much on the
state in which they are available to the researcher. Corpus clean-up and especially the
correction of errors is a time-consuming and expensive business, and some would argue that it
detracts from the `purity' of the data. But the extract given here illustrates a potential source of
problems if a corpus is not cleaned up in some ways --- the penultimate French sentence
contains a false start, followed by ..., while the English text (presumably produced by a human
translator) contains just a complete sentence. This sort of divergence could in principle effect
the statistics for word-level alignment.
In order to get some idea of how the translation model works, it is useful to introduce some
further notions. In a word-aligned sentence-pair, it is indicated which target words correspond
to each source word. An example of this (which takes French as the source language) is given
in the second extract.

The numbers after the source words indicate the string position of the corresponding target
word or words. If there is no target correspondence, then no bracketted numbers appear after
the source word (e.g. a in a demandé). If more than one word in the target corresponds, then
this is also indicated. The fertility of a source word is the number of words corresponding to it
in the target string. For example, the fertility of asked with English as source language is 2,
since it aligns with a demandé. A third notion is that of distortion which refers to the fact that
source words and their target correspondences do not necessarily appear in the same string
position (compare tout acheter and buy everything, for example).
The parameters which must be calculated from the bilingual sentence aligned corpus are then (i)
the fertility probabilities for each source word (i.e. the likelihood of it translating as one, two,
three, etc, words respectively), (ii) the word-pair or translation possibilities for each word in
each language and (iii) the set of distortion probabilities for each source and target position.
With this information (which is extracted automatically from the corpus), the translation model
can, for a given S, calculate (that is, the probability of T, given S). This is the essence of the
approach to statistically-based MT, although the procedure is itself slightly more complicated in
involving search through possible source language sentences for the one which maximises ,
translation being essentially viewed as the problem of finding the S that is most probable given
T --- i.e. one wants to maximise . Given that

then one just needs to choose S that maximizes the product of and .
It should be clear that in an approach such as this there is no role whatsoever for the explicit
encoding of linguistic information, and thus the knowledge acquisition problem is solved. On
the other hand, the general applicability of the method might be doubted, since as we observed
above, it is heavily dependent on the availability of good quality bilingual or multilingual data
in very large proportions, something which is currently lacking for most languages.

http://privatewww.essex.ac.uk/~doug/book/node94.html (2 de 3) [20/09/2000 18:04:46]

Statistical MT

Results to date in terms of accuracy have not been overly impressive, with a 39% rate of correct
translation reported on a set of 100 short test sentences. A defect of this approach is that
morphologically related words are treated as completely separate from each other, so that, for
example, distributional information about sees cannot contribute to the calculation of
parameters for see and saw, etc. In an attempt to remedy this defect, researchers at IBM have
started to add low level grammatical information piecemeal to their system, moving in essence
towards an analysis-transfer-synthesis model of statistically-based translation. The information
in question includes morphological information, the neutralisation of case distinctions (upper
and lower case) and minor transformations to input sentences (such as the movement of
adverbs) to create a more canonical form. The currently reported success rate with 100 test
sentences is a quite respectable 60%. A major criticism of this move is of course precisely that
linguistic information is being added piecemeal, without a real view of its appropriacy or
completeness, and there must be serious doubts about how far the approach can be extended
without further additions of explicit linguistic knowledge, i.e. a more systematic notion of
grammar. Putting the matter more positively, it seems clear that there is a useful role for
information about probabilities. However, the poor success rate for the `pure' approach without
any linguistic knowledge (less than 40%) suggests that the real question is how one can best
combine statistical and rule- based approaches.

Next: Summary Up: Empirical Approaches to Previous: Example-Based Translation

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node94.html (3 de 3) [20/09/2000 18:04:46]

Summary

Next: Further Reading Up: New Directions in Previous: Statistical MT

Summary
We have tried in this chapter to give a brief overview of some of the issues and techniques
which are being actively researched today in MT. Of course, there is not enough room in one
chapter to do justice to the field, and we have of necessity omitted much work that is of interest.
In particular, we have restricted our discussion to MT itself and have said nothing at all about
recent work in the development of translators aid s, multilingual authoring packages and
terminological systems of various sorts. Nonetheless we have identified three important trends
in current research in MT. The first is the exploitation of current techniques from computational
linguistics to permit a multidimensional view of the translational relation between two texts.
The second in the increasing orientation of the research community towards the use of existing
resources of various sorts, either to extract useful information or directly as components in
systems. The third, related, trend is towards statistical or empirical models of translation.
Though we have dwelt in some detail in this short survey on `pure' statistical and simple pattern
matching methods, in fact much recent work advocates a mixture of techniques, for example
with statistical methods supplementing rule-based methods in various ways.

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node95.html [20/09/2000 18:04:47]

Further Reading

Next: Useful Addresses Up: New Directions in Previous: Summary

Further Reading
Our discussion of flexible translation between Swedish and Norwegian is based on
unpublished work by [Dyvik1992]. The standard references on sign-based approaches to
linguistic representation are [Pollard and Sag1987,Pollard and Sag1993]. The view of
constraint based translation that we describe is loosely modelled on that used in `Shake and
Bake' [Whitelock1992,Beaven1992]. See [Kaplan et al. 1989], [Sadler1991] and
[Sadler1993] for a slightly different approach. General discussion of how multi-dimensional
representations can be used in MT can be found in [Sadler and Arnold1993].

On knowledge-based MT see [Goodman and Nirenburg1991], and the special issue of the
journal Machine Translation, [Goodman1989].

On the processing of corpora, and their use in linguistics generally, see [Garside et al. 1987],
and [Aijmer and Altenberg1991].

The idea of example-based MT was first discussed in a paper by Nagao [Nagao1984]. For a
review of more recent work along these lines, see [Somers1992].

The pure statistical approach to MT is based on the work of a team at IBM , see for example
[Brown et al. 1990]. As regards aligned, bilingual corpora, the most common form of
alignment takes the sentence to be the organizing unit for chunking, see [Brown et al. 1991] and
[Gale and Church1991b] for relevant discussion. On automatic extraction of word
correspondences across bitext, see [Gale and Church1991a]. Techniques involving the use of
corpus resources for automatic sense disambiguation have also been explored within the DLT
project, see [Sadler1989].

The translation of no, which was described around page above, is discussed by [Sumita et
al. 1990]. The discussion of o-negaishimasu is from [Furuse and Iida1992b], see also [Furuse
and Iida1992a], and [Sumita and Iida1991].

The frame for computer on page above is taken from [Goodman and Nirenburg1991,
page 25,].
For up to date reports on research in the field of MT, there are several journals, and several
major international conferences. The specialist Journal is Machine Translation, edited by Sergei
Nirenburg, from Carnegie Mellon University in Pittsburg, USA, and published by Kluwer
Academic Publishers. However, the journal Computational Linguistics, published by the MIT
Press for the Association for Computational Linguistics (ACL), also publishes research which
is directly about MT.

http://privatewww.essex.ac.uk/~doug/book/node96.html (1 de 2) [20/09/2000 18:04:48]

Further Reading

The specialist conference for research on MT is called TMI --- for `Theoretical and
Methodological Issues (in Machine Translation)'. This has been held every two years since
1986, and proceedings are published ([TMI11985],[TMI21988][TMI31990],[TMI41992]).
Many of the papers in the last of these are directly or indirectly about the issue of `rationalist'
(i.e. rule-based ) vs. empirical approaches to MT. The proceedings of the main Computational
Linguistics conferences, namely ( COLING), the conferences of the Association for
Computational Linguistics (ACL) and the conferences of the European Chapters of the ACL,
also contain a high percentage of papers about MT. ACL conferences are held annually in the
USA (for example, [ACL281990,ACL291991,ACL301992]). The EACL conferences are held
biennially, [EACL11983,EACL21985,EACL31987,EACL41989,EACL51991], as is COLING:
Coling 84 [Coling841984] was held in Stanford, California, COLING 86 [Coling861986] in
Bonn, Coling 88 [Coling881988] in
Budapest, Coling 90 [Coling901990] in Helsinki, and Coling 92 [Coling921992] was held in
Nantes.

Next: Useful Addresses Up: New Directions in Previous: Summary

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node96.html (2 de 2) [20/09/2000 18:04:48]

Useful Addresses

Next: Glossary Up: Machine Translation Previous: Further Reading

Useful Addresses
ASLIB,
The Association for Information Management,
Information House,
20-24, Old Street,
London, EC1V 9AP,
UK.
The Association for Computational Linguistics (ACL),
c/o Donald E. Walker,
Bellcore, MRE 2A379,
445 South Street, Box 1910,
Morristown, NJ 07960-1910,
USA.
The Association for Computational Linguistics (European Chapter),
c/o Michael Rosner,
IDSIA,
Corso Elvezia 36,
CH-6900 Lugano,
Switzerland.
Computational Linguistics,
MIT Press Journals,
55 Hayward Street,
Cambridge, MA 02142,
USA.
Language Industry Monitor,
Uitgeverij LIM,
Eerste Helmersstraat 183,
1054 DT Amsterdam,
The Netherlands.
Machine Translation,
Kluwer Academic Publishers,
PO Box 322,
3300 AH, Dordrecht,
The Netherlands.
American Association for Machine Translation,
Suite 310, 655, Fifteenth Street,

http://privatewww.essex.ac.uk/~doug/book/node97.html (1 de 2) [20/09/2000 18:04:49]

Useful Addresses

Washington, DC 20005,
USA.
Asia-Pacific Association for Machine Translation,
305 Akasaka Chuo Mansion, 2-17 Akasaka 7-chome,
Minato-ku, Tokyo 107,
Japan.
European Association for Machine Translation,
ISSCO,
54 route des Acacias,
CH-1227 Carouge (Geneva),
Switzerland.
International Association for Machine Translation,
c/o AMTA,
Suite 310, 655, Fifteenth Street, NW,
Washington, DC 20005,
USA.

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node97.html (2 de 2) [20/09/2000 18:04:49]

Glossary

Next: References Up: Machine Translation Previous: Useful Addresses

Glossary
adjective phrase (AP) a complete construction headed by an adjective. APs typically modify
nouns and occur as complements to verbs such as be, seem, become. For example: The man
guilty of this heinous crime was imprisoned. John seems rather stupid.
adjunct or modifier an optional or secondary element in a construction which can be removed
without affecting the structural status of the rest of the construction. For example, yesterday in:
John kicked the ball yesterday (Compare John kicked the ball where the ball is not an adjunct,
because * John kicked yesterday is ungrammatical).
affix morpheme placed at the beginning (prefix), middle (infix), or end ( suffix) of the root or
stem of a word, e.g. relegal ize.
agreement the process whereby the form of one word requires a corresponding form of another
- for example, the plural form boys requires a plural form of the demonstrative determiner
these/* this: these boys vs * this boys.
algorithm a prescribed set of well-defined rules or instructions for the solution of a problem.
analysis the phase in natural language processing systems (including MT systems) in which a
structure or representation is assigned to source language (input) sentences or the representation
itself or the name for the module of linguistic rules involved.
anaphor a word or phrase which refers back to some previously expressed word or phrase or
meaning (typically, pronouns such as herself, himself, he, she).
antecedent the word or phrase to which a later word or phrase (e.g. an anaphor) refers.
Artificial Intelligence (AI) the branch of Computing Science concerned with simulating
aspects of human intelligence such as language comprehension and production, vision,
planning, etc.
ASCII American Standard Code for Information Interchange - a standard set of codes used for
representing alphanumeric information in a computer.
aspect a property of verbs or sentences, which refers primarily to the duration or type of
activity described, e.g. the distinction between Sam sang and Sam was singing.
attribute value pair Many contemporary linguistic analyses use collections of features or
attribute value pairs to encode various properties of a linguistic entity. In the pair number
sing , number is the attribute and sing is the value.
auxiliary (AUX) in English, auxiliary verbs are those which carry distinctions of tense, aspect,
etc, such as do, be and have. The modal auxiliaries include can/could, may/might, shall/should,

http://privatewww.essex.ac.uk/~doug/book/node98.html (1 de 7) [20/09/2000 18:04:51]

Glossary

ought to, need and used to. Auxiliary verbs are opposed to main verbs ( walk, play, etc.)
batch (processing) as opposed to interactive processing. In batch processing, a computer does
not perform tasks as soon as requested, but groups similar jobs together into batches and carries
them out together at some later time (e.g. overnight). Interactive processing allows the user to
issue an instruction and have it carried out more or less instantly.
bitext a bilingual text which is aligned so that within each bilingual chunk the texts are
translations of each other. The use of the term does not necessarily commit one as to the level at
which a text is chunked and aligned, e.g. into sentences or paragraphs, but the chunks are very
often sentences.
case a property of words, primarily nouns, which varies according to their syntactic function.
English distinguishes three cases of pronouns, one used for pronouns which are the subject of
finite verbs ( he, I) one for possessive pronouns ( his, my) and one for pronouns elsewhere (
him, me). The case system of many other languages is much more extensive.
CD-Rom a compact disc used for the storage of data in read-only (ROM) format.
collocation phrases composed of words that co-occur for lexical rather than semantic reasons,
for example, a heavy smoker is one who smokes a great deal, but someone who writes a great
deal is not a heavy writer. This seems to be a lexical fact, not related to the meanings of smoker
or writer.
common sense reasoning reasoning on the basis of common knowledge, as opposed to purely
logical reasoning, or reasoning that depends solely on the meanings of words. A purely logical
inference might be from If it is Tuesday, Sam is in London and It is Tuesday to the conclusion
Sam is in London. An example of common sense reasoning might be the inference that if
someone asks for a phone book it is because they want to look up a number, and make a phone
call.
complement a term for all constituents of the sentence required by a verb except for the subject
(e.g. the object is a complement of the verb).
compound two or more words which function as one word (e.g. fireplace, video-tape, door
handle). Most common in English and closely related languages are noun-noun compounds
functioning as nouns. Because such compounds have the external behaviour and distribution of
a lexical item, they are often taken to be morphological structures.
constituent a linguistic unit which is a component of a larger construction. These units can, in
turn, be analysed into further constituents (e.g. a noun phrase can be analysed into a
determiner and a noun).
constituent structure the structure of an expression in terms of the constituent syntactic parts
and their categories (as opposed to analysis in terms of grammatical or semantic relations).
context all the factors which systematically determine the form, meaning, appropriateness or
translation of linguistic expressions. One can distinguish between linguistic context (provided
by the preceding utterances or text) and non-linguistic context (including shared assumptions
and information).
controlled language a specially simplified version of a language which is adopted (typically by
a company or a documentation section of a company) as a partial solution to a perceived

http://privatewww.essex.ac.uk/~doug/book/node98.html (2 de 7) [20/09/2000 18:04:51]

Glossary

communication problem. Both the vocabulary and the syntactic structures may be restricted.
corpus collection of linguistic data, either written texts or a transcription of recorded speech.
Typically, corpora have to be quite large to be of any linguistic use (upwards of 100,000
tokens).
critiquing system a computer program which analyses a text and indicates where it deviates
from the norms of language use.
database generally, any collection of information that can be created, accessed, and processed
automatically. Many sophisticated software packages exist for creating and accessing databases
of information.
dependency grammar a type of grammar which operates essentially in terms of types of
dependencies or grammatical relation between heads and dependent elements of a construction
rather than in terms of constituent structure.
derivational a term used in morphology to refer to one of the two main processes of
work-formation, the other being inflectional. Derivational processes result in words of a
different class. In English, the major derivational process is suffixation, e.g. derive - derivation,
happy - happiness, nation - national.
electronic dictionary dictionary which is stored on computer and can be accessed by programs,
e.g. so that definitions can be looked up and displayed on screen.
feature see attribute-value pair
finite a form of a verb that can occur as the head of a sentence. In Sam wants to leave, wants is
finite, leave is non-finite.
gender 2 types of gender are distinguished in linguistics --- natural gender, where items refer to
the sex of real world entities, and grammatical gender, which has nothing to do with sex, but
which signals grammatical relationships between words in a sentence and which is shown e.g.
by the form of the article or the noun.
generation (also synthesis) the phase in a natural language processing system (including MT
systems) in which a strings or sentences are produced from some sort of underlying
representation, typically a meaning representation of some sort or the name for the module of
linguistic rules which causes this to happen.
grammar the term is generally used to include syntax and morphology but may also be used
in a wider sense to include rules of phonology and semantics. A grammar is a collection of
linguistic rules which define a language.
grammatical relations the relations which hold between a head (such as a verb) and its
dependents. For example, subject and object are grammatical relations borne by constituents in
a sentence.
head the central or most important element in a construction which determines the external
distribution of the construction and places certain requirements on the words or constituents it
occurs with. For example, the verb saw is head of the sentence The big man saw Mary and of
the VP saw Mary. Nouns are heads of NPs, prepositions are heads of PPs, adjectives of APs,
etc. In lexicography, head is another term for headword.

http://privatewww.essex.ac.uk/~doug/book/node98.html (3 de 7) [20/09/2000 18:04:51]

Glossary

headword word forming the heading of an entry in a dictionary.

homographs words which have the same spelling but which differ in meaning, e.g. bank
(financial institution) and bank (of a river).
idiom a sequence of words which functions semantically as a unit and with an unpredictable
meaning (e.g. kick the bucket, meaning die). This is generally accompanied by a degree of
syntactic restriction.
imperative verb forms or sentence types that are used to express commands (e.g. Go away!)
indexical a word which depends on the context of utterance for its meaning (e.g. I, you, here).
indirect object (IOBJ) the constituent of a sentence most typically associated with the goal or
recipient role. In English indirect objects are often PPs with the preposition to, e.g. Lee gave the
book to his friend.
inflectional term in morphology assigned to affixes which encode grammatical properties such
as number, tense and do not change the part of speech of the stems to which they are
attached.
interlingual language independent, a linguistic knowledge based approach to MT where
translation proceeds in 2 stages - analysis (where input string is mapped onto a language
independence representation) and generation, cf. transfer.
intransitive a verb that does not take a direct object (e.g. die).
lexicon used synonymously with dictionary.
light verbs (also support verbs) verbs that are semantically empty or relatively empty (e.g.
take in take a walk).
markup codes in some (text formatting) description language which determine how text will
look when printed.
metaphor in metaphorical usage, expressions are used in a way that appears literally false. For
example, using the word boiling to describe water which is simply too hot for comfort.
mood a term applied to sentences and verbs to signal a wide range of meanings, especially
speaker's attitude to the factual content of utterances, e.g. certainty, possibility (e.g. Sam
must/may be at home). The distinction between active and passive sentences/verbs is also
sometimes considered a mood.
morphology the branch of grammar which studies the structure or forms of words. The main
branches are inflectional morphology, derivational morphology, and compounding.
natural language a term which denotes a (naturally occurring) human language as opposed to
computer languages and other artificial languages.
NLP (Natural Language Processing) the field of inquiry concerned with the study and
development of computer systems for processing natural (human) languages.
noun phrase (NP) a complete construction headed by a noun. It can be substituted by, or act
as antecedent for, a pronoun of the appropriate sort:
The man who I saw yesterday has just knocked at the door. Can you let him in?

http://privatewww.essex.ac.uk/~doug/book/node98.html (4 de 7) [20/09/2000 18:04:51]

Glossary

number the number of a noun or noun phrase generally corresponds to the number of real
world entities referred to (e.g. singular NPs denote single individuals ( a table), plural NPs
denote collections of individuals ( two tables). However the relationship between real number
and grammatical number is not always straightforward - trousers is plural in form yet denotes a
singular entity (as in the committee are considering that question this afternoon) and some
nouns do not have distinct singular and plural forms ( sheep, salmon).
object (OBJ) also direct object - the constituent of a sentence generally associated with the
entity which undergoes the action. In English, the direct object of a verb is a NP and normally
follows the verb, e.g. Peter saw Mary.
OCR Optical Character Reader. A device which scans printed textual material and converts it
into electronic form, storing it in a file on the computer or disc. OCR technology has improved
dramatically in recent years and is now a reasonably accurate way of making text available in
electronic form.
participle the term covers both a word derived from a verb and used as an adjective, as in a
singing woman, and the -ing and -en non-finite forms of the verb, as in was singing (present
participle), has given (past participle).
particle an element which occurs in a single form (like a preposition in English) and with a
function that does not easily fit into standard parts of speech classifications. Particles very often
occur in constructions with certain verbs in English with varying degrees of idiosyncratic
interpretation: John took off at great speed (i.e. left). May gave herself up (i.e. surrendered)
part of speech (category) the class of units used in the description of a language, e.g. noun,
verb, noun phrase, verb phrase.
phonology the branch of linguistics which studies the sound systems of languages.
Phonological rules describe the patterns of sounds used distinctively in a language, and
phonologists are interested in the question of what constitutes a possible sound system for a
natural language.
post-editing program that performs some operations on the output of another program,
typically formatting the output for some device or filtering out unwanted items.
predicate traditional and modern grammars often divide sentences so that constituents other
than the subject are considered together to form the predicate (e.g. John (
subject) kicked the ball ( predicate)).
prepositional phrase (PP) a phrase headed by a preposition, a word such as on, in, between.
Prepositions combine with other constituents (usually noun phrases) to form prepositional
phrases, as in The man sat on the bench.
probabilistic a term for approaches to natural language processing (including MT) which rely
to some extent on statistical methods.
pronoun a word that can substitute for a noun phrase (e.g. he can substitute for John).
prosodic indicating stress or intonation.
reading a sense of a word that can be distinguished from other senses or meanings of the same
word.

http://privatewww.essex.ac.uk/~doug/book/node98.html (5 de 7) [20/09/2000 18:04:51]

Glossary

relative clause a clause which qualifies or restricts the meaning of the noun in a noun phrase.
It may be introduced by words such as who, which and that in English: the man who I saw this
morning, the woman (that) I sent the letter to.
root that part of a word that is left when all affixes have been removed ( industry is the root of
preindustrial).
selectional restrictions selectional restrictions are essentially semantic restrictions on
combinations of words. For example, verbs place such restrictions on their subjects and objects
- the verb frighten generally requires as (active) subject something animate which can
experience fear.
semantics the branch of linguistics which studies meaning in language. One can distinguish
between the study of the meanings of words (lexical semantics) and the study of how the
meanings of larger constituents come about (structural semantics).
semantic role also called deep case, semantic relation or thematic role. A semantic role is a
description of the relationship that a constituent plays with respect to the verb in the sentence.
The subject of an active sentence is often the agent or experiencer. Other roles include
instrumental, benefactive, patient: Peter (experiencer) died. The cat (agent) chased the dog
(patient).
SGML Standard Generalized Markup Language. A generic language for marking various
formatting and other textual relationships in a text.
source language when translating, the language one is translating out of; in French to English
translation, French is the source language.
speech act a declarative sentence can be used to perform a number of different speech acts. In
uttering It's cold in here a speaker may perform an act of requesting the hearer to close the
window or turn up the heating.
stem that part of a word to which inflectional affixes are attached (it consists of the root plus
any derivational affixes.
subcategorization the pattern of complements selected by head, e.g. the verb put
subcategorizes for an NP and a PP. We put the car in the garage, but not * We put the car.
subject the constituent of an active sentence most typically associated with the `doer' or
`undergoer' of an action. The verb agrees with the subject in person and number in English.
sublanguage a language used to communicate in a specialized technical domain or for a
specialized purpose, for example, the language of weather reports, expert scientific polemic or
other modes of scientific discourse, user or maintenance manuals, drug interaction reports, etc.
Such language is characterised by the high frequency of specialized terminology and often also
by a restricted set of grammatical patterns. The interest is that these properties make
sublanguage texts easier to translate automatically.
suffix an affix that is added following a root or stem, for example the boldface parts of legal
ize, nation al.
syntax the rules of a grammar which govern the way words are combined to form sentences
and other phrases in a language.

http://privatewww.essex.ac.uk/~doug/book/node98.html (6 de 7) [20/09/2000 18:04:51]

Glossary

tag to tag a text is to annotate it with grammatical information. Usually tagging takes the form
of part-of-speech annotations but semantic tags or tags encoding other linguistic information
can be used. Tagging is usually performed automatically or semi-automatically.
target language when translating, the language one is translating into; in French to English
translation, English is the target language.
tense a property of verbs relating primarily to the time at which the action or event denoted by
the verb takes place. For example, past tense verbs, as in Sam left, describe events in the past.
testsuite a collection of sentences or sentence fragments collated to test the capabilities of a
translation system or other NLP application.
thesaurus a list of words arranged according to meaning, rather than alphabetically as in a
standard dictionary.
transfer the phase in MT where a source language representation is mapped onto a target
language representation, a linguistic knowledge based approach to MT where translation
proceeds in three stages --- analysis (where input string is mapped onto a source language
representation) transfer and generation.

Next: References Up: Machine Translation Previous: Useful Addresses

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node98.html (7 de 7) [20/09/2000 18:04:51]

References

Next: Index Up: Machine Translation Previous: Glossary

References
ACL281990
28th Annual Meeting of the Association for Computational Linguistics, Pittsburg, Pa.,
1990.
ACL291991
29th Annual Meeting of the Association for Computational Linguistics, Berkeley, Ca.,
1991.
ACL301992
30th Annual Meeting of the Association for Computational Linguistics, University of
Delaware, Newark, Delaware, 1992.
AECMA1988
AECMA/AIA Simplified English. AECMA, Paris, 1988.
Aijmer and Altenberg1991
Karin Aijmer and Bengt Altenberg, editors. English Corpus Linguistics. Longman,
London, 1991.
Albisserin press
D. Albisser. Evaluation of MT systems at Union Bank of Switzerland. Machine
Translation, 7, in press. Special Issue on Evaluation.
Allegranza et al. 1991
V. Allegranza, P. Bennett, J. Durand, F. Van Eynde, L. Humphreys, P. Schmidt, and E.
Steiner. Linguistics for Machine Translation: The Eurotra linguistic specifications. In C.
Copeland, J. Durand, S. Krauwer, and B. Maegaard, editors, The Eurotra Linguistic
Specifications, pages 15--124. CEC, Luxembourg, 1991.
Allen1987
J. Allen. Natural Language Understanding. Benjamin Cummings Pub. Co., Inc., Menlo
Park, Ca., 1987.
Allerton1984
D.J. Allerton. Three (or four) levels of word cooccurrence restriction. Lingua, 63:17--40,
1984.
Apresian et al. 1992
Jurij Apresian, Igor Boguslavskij, Leonid Iomdin, Alexandre Lazurskij, Vladimir
Sannikov, and Leonid Tsinman. ETAP-2: The linguistics of a Machine Translation
system. META, XXXVII(4):97--112, 1992.
Arnold and des Tombe1987
D.J. Arnold and Louis des Tombe. Basic theory and methodology in Eurotra. In Sergei
Nirenburg, editor, Machine Translation:

http://privatewww.essex.ac.uk/~doug/book/node99.html (1 de 16) [20/09/2000 18:04:58]

References

Theoretical and Methodological Issues, pages 114--135. Cambridge University Press,

Cambridge, 1987.
Arnold and Sadler1989
D.J. Arnold and L.G. Sadler. Non-compositionality and translation. In Jeremy Peckham,
editor, Recent Developments and Applications of Natural Language Processing, pages
23--55. Kogan Page, London, 1989.
Arnold and Sadler1990
D.J. Arnold and L.G. Sadler. The theoretical basis of MiMo. Machine Translation,
5:195--222, 1990.
Arnold et al. 1988
D.J. Arnold, S. Krauwer, L. des Tombe, and L.G. Sadler. Relaxed compositionality in
MT. Proceedings of the Second International Conference on Theoretical and
Methodological Issues in Machine Translation, 1988.
Arnold et al. in press a
D.J. Arnold, D. Moffat, L.G. Sadler, and A. Way. Automatic test suite generation.
Machine Translation, 7, in press . Special Issue on Evaluation.
Arnold et al. in press b
D.J. Arnold, L.G. Sadler, and R.L. Humphreys. Evaluation: An assessment. Machine
Translation, 7, in press . Special Issue on Evaluation.
Arnold1986
D.J. Arnold. Eurotra: A European perspective on MT. Proceedings of the IEEE,
74(7):979--992, 1986.
Arnold1990
D.J. Arnold. Text typology and Machine Translation: An overview. In Pamela Mayorcas,
editor, Translating and the Computer 10, pages 73--89. Aslib, London, 1990.
Arnold1993
D.J. Arnold. Sur la conception du transfert. In P. Bouillon and A. Clas, editors, Etudes et
Recherches en Traductique : Problemes de Traduction par Ordinateur. Presses de
Montreal/AUPELF, Montreal, 1993.
Balkan et al. 1991
L. Balkan, M. Jaeschke, L. Humphreys, S. Meijer, and A. Way. Declarative evaluation of
an MT system: Practical experiences. Applied Computer Translation, 1(3), pages 49--59,
1991.
Bar-Hillel1951
Y. Bar-Hillel. The state of translation in 1951. American Documentation, 2:229--237,
1951. (Reprinted in Y. Bar-Hillel, Language and Information, Addison Wesley, Reading,
Mass., 1964, pages 153-165).
Barr and Fiegenbaum1981
Avron Barr and Edward A. Fiegenbaum, editors. The Handbook of Artificial Intelligence,
Vol 1. William Kaufmann, Inc., Los Altos, California, 1981.
Bauer1983
L. Bauer. English Word-Formation. Cambridge University Press, Cambridge, 1983.

http://privatewww.essex.ac.uk/~doug/book/node99.html (2 de 16) [20/09/2000 18:04:58]

References

Beaven1992
John L. Beaven. Shake-and-bake Machine Translation. Proceedings of the 15th
International Conference on Computational Linguistics, volume II, pages 602--609,
1992.
Bennett and Slocum1988
W. Scott Bennett and J. Slocum. METAL: the LRC Machine Translation system. In J.
Slocum, editor, Machine Translation Systems, pages 111--140. CUP, Cambridge, 1988.
Bennett et al. 1986
P. Bennett, R.L. Johnson, J. McNaught, J.M. Pugh, J.C. Sager, and H.L. Somers, editors.
Multilingual Aspects of Information Technology. Gower Press, 1986.
Bennett1993
Paul Bennett. The interaction of syntax and morphology in Machine Translation. In F.
Van Eynde, editor, Linguistic Issues in Machine Translation, pages 72--104. Pinter,
London, 1993.
Benson et al. 1986a
Morton Benson, Evelyn Benson, and Robert Ilson. The BBI Combinatory Dictionary of
English: A Guide to Word Combinations. John Benjamins Publishing Co., Amsterdam,
1986.
Benson et al. 1986b
Morton Benson, Evelyn Benson, and Robert Ilson. Lexicographic Description of English.
Studies in Language Companion Series. John Benjamins Publishing Co., Amsterdam,
1986.
Berlin and Kay1969
B. Berlin and P. Kay. Basic Color Terms: Their Universality and Evolution. University
of California Press, Berkeley, Ca., 1969.
Boguraev and Briscoe1989
B. Boguraev and E. Briscoe, editors. Computational Lexicography for Natural Language
Processing. Longman, London, 1989.
Borsley1991
Robert D. Borsley. Syntactic Theory: A Unified Approach. Edward Arnold, London,
1991.
Bratko1986
I. Bratko. Prolog Programming for Artificial Intelligence. Addison Wesley, Wokingham,
1986.
Brown and Miller1991
E.K. Brown and J.E. Miller. Syntax: A Linguistic Introduction to Sentence Structure.
Harper Collins, London, 1991.
Brown et al. 1990
P.F. Brown, J. Cocke, S.A. Della Pietra, F. Jelinek, J.D. Lafferty, R.L. Mercer, and P.S.
Roossin. A statistical approach to Machine Translation. Computational Linguistics,
16:79--85, 1990.
Brown et al. 1991
P.F. Brown, J.C. Lai, and R.L. Mercer. Aligning sentences in parallel corpora. In
http://privatewww.essex.ac.uk/~doug/book/node99.html (3 de 16) [20/09/2000 18:04:58]
References

Proceedings of 29th ACL, pages 169--176, Berkeley, California, 1991.

Bruce1972
Bertram Bruce. A model for temporal reference and its application in a question
answering program. Artificial Intelligence, 3:1--25, 1972.
Buchmann1987
B. Buchmann. Early history of Machine Translation. In M. King, editor, Machine
Translation Today: The State of the Art, Proceedings of the Third Lugano Tutorial, 1984,
pages 3--21. Edinburgh University Press, Edinburgh, 1987.
Cann1993
Ronnie Cann. Formal Semantics. Cambridge University Press, Cambridge, 1993.
Chandioux1976
J. Chandioux. Météo: Un système opérationnel pour la traduction automatique des
bulletins météorologiques destinés au grand public. META, 21:127--133, 1976.
Chandioux1989a
J. Chandioux. 10 ans de Météo (MD). In A. Abbou, editor, 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, pages
169--172. Editions Daicadif, Paris, 1989.
Chandioux1989b
J. Chandioux. Météo: 100 million words later. In D.L. Hammond, editor, American
Translators Association Conference 1989: Coming of Age, pages 449--453. Learned
Information, Medford, NJ, 1989.
Charniak and McDermott1985
Eugene Charniak and Drew McDermott. An Introduction to Artificial Intelligence.
Addison Wesley, Reading, Mass, 1985.
Charniak and Wilks1976
E. Charniak and Y. Wilks, editors. Computational Semantics. North Holland Pub. Co.,,
Amsterdam, 1976.
Chierchia and McConnell-Ginet1990
G. Chierchia and S. McConnell-Ginet. Meaning and Grammar: An Introduction to
Semantics. MIT Press, Cambridge, Mass., 1990.
Coling841984
Proceedings of the Eleventh International Conference on Computational Linguistics,
Stanford University, California, 1984.
Coling861986
Proceedings of the Twelfth International Conference on Computational Linguistics,
University of Bonn, 1986.
Coling881988
Coling88. Proceedings of the Thirteenth International Conference on Computational
Linguistics, volume 1-2, Budapest, 1988. John von Neumann Society for Computing
Sciences.
Coling901990

http://privatewww.essex.ac.uk/~doug/book/node99.html (4 de 16) [20/09/2000 18:04:58]

References

Proceedings of the Fourteenth International Conference on Computational Linguistics,

volume 1-3, Helsinki University, 1990.
Coling921992
Proceedings of the Fifteenth International Conference on Computational Linguistics,
volume 1-4, Université de Nantes, 1992.
Copeland et al. 1991a
C. Copeland, J. Durand, S. Krauwer, and B. Maegaard, editors. The Eurotra Formal
Specifications, volume 2 of Studies in Machine Translation and Natural Language
Processing. Office for Official Publications of the Commission of the European
Community, Luxembourg, 1991.
Copeland et al. 1991b
C. Copeland, J. Durand, S. Krauwer, and B. Maegaard, editors. The Eurotra Linguistic
Specifications, volume 1 of Studies in Machine Translation and Natural Language
Processing. Office for Official Publications of the Commission of the European
Community, Luxembourg, 1991.
Danlos and Samvelian1992
L. Danlos and P. Samvelian. Translation of the predicative element of a sentence:
Category switching, aspect and diathesis. Fourth International Conference on
Theoretical and Methodological Issues in Machine Translation, pages 21--34, 1992.
Danlos1992
L. Danlos. Support verb constructions: Linguistic properties, representation, translation.
Journal of French Language Studies, 2(1), 1992.
Dyvik1992
Helge Dyvik. The PONS system: Translation by principled rules of thumb (partial
translation between closely related languages). Unpublished Ms., Dept. of Linguistics
and Phonetics, University of Bergen, 1992.
EACL11983
Proceedings of the First Conference of the European Chapter of the Association for
Computational Linguistics, Pisa, Italy, 1983.
EACL21985
Proceedings of the Second Conference of the European Chapter of the Association for
Computational Linguistics, University of Geneva, Switzerland, 1985.
EACL31987
Proceedings of the Third Conference of the European Chapter of the Association for
Computational Linguistics, University of Copenhagen, Denmark, 1987.
EACL41989
Proceedings of the Fourth Conference of the European Chapter of the Association for
Computational Linguistics, UMIST, UK, 1989.
EACL51991
Proceedings of the Fifth Conference of the European Chapter of the Association for
Computational Linguistics, Berlin, 1991.
Estival et al. 1990
Dominique Estival, Afzal Ballim, Graham Russell, and Susan Warwick. A syntax and

http://privatewww.essex.ac.uk/~doug/book/node99.html (5 de 16) [20/09/2000 18:04:58]

References

semantics for feature-structure transfer. In Proceedings of the 3rd International

Conference on Theoretical and Methodological Issues in Machine Translation of Natural
Language, pages 131--143, Linguistics Research Centre, University of Texas at Austin,
USA, 1990.
Evens1988
M.W. Evens, editor. Relational Models of the Lexicon: Representing Knowledge in
Semantic Networks. Cambridge University Press, Cambridge, 1988.
Flank et al. in press
S. Flank, A. Temin, H. Blejer, A. Kehler, and S. Greenstein. Module-level testing for
natural language understanding. Machine Translation, 7, in press. Special Issue on
Evaluation.
Furuse and Iida1992a
Osamu Furuse and Hitoshi Iida. Cooperation between transfer and analysis in
example-based framework. In Proceedings of the 15th International Conference on
Computational Linguistics, volume II, pages 645--651, Nantes, France, 1992.
Furuse and Iida1992b
Osamu Furuse and Hitoshi Iida. An example-based method for transfer-driven Machine
Translation. In Fourth International Conference on Theoretical and Methodological
Issues in Machine Translation, pages 139--150, Montreal, Canada, 1992.
Gale and Church1991a
W.A. Gale and K.W. Church. Identifying word correspondence and parallel text. In 4th
DARPA Workshop on Speech and Natural Language, Asilomar, 1991.
Gale and Church1991b
W.A. Gale and K.W. Church. A program for aligning sentences in bilingual corpora. In
Proceedings of the 29th ACL, pages 177--184, University of California, Berkeley, Ca.,
1991.
Garside et al. 1987
Roger Garside, Geoffrey Leech, and Geoffrey Sampson, editors. The Computational
Analysis of English: A Corpus-Based Approach. Longman, London, 1987.
Gazdar and Mellish1989
G. Gazdar and C. Mellish. Natural Language Processing in Prolog. Addison Wesley,
Wokingham, 1989.
Goldfarb1986
C.F. Goldfarb, editor. The Standard Generalized Markup Language (ISO 8879).
International Organisation for Standardization, Geneva, 1986.
Goodman and Nirenburg1991
Kenneth Goodman and Sergei Nirenburg, editors. The KBMT Project: A Case Study in
Knowledge Based Machine Translation. Morgan Kaufmann, San Mateo, California,
1991.
Goodman1989
K. Goodman, editor. Machine Translation, volume 4. 1989. numbers 1 and 2 : Special
Issues on Knowledge Based MT, Parts I and II.
Grimaila and Chandioux1992

http://privatewww.essex.ac.uk/~doug/book/node99.html (6 de 16) [20/09/2000 18:04:58]

References

Annette Grimaila and John Chandioux. Made to measure solutions. In John Newton,
editor, Computers in Translation: A Practical Appraisal, pages 33--45. Routledge,
London, 1992.
Grishman and Kittredge1986
R. Grishman and R.I. Kittredge, editors. Analyzing Language in Restricted Domains:
Sublanguage Description and Processing. Lawrence Erlbaum Associates, New Jersey:
Hillsdale, 1986.
Grishman1986
Ralph Grishman. Computational Linguistics: An Introduction. Cambridge University
Press, Cambridge, 1986.
Grosz and Sidner1986
B.J. Grosz and C.L. Sidner. Attention, intentions, and the structure of discourse.
Computational Linguistics, 12:175--204, 1986.
Guida and MauriJuly 1986
Giovanni Guida and Giancarlo Mauri. Evaluation of Natural Language processing
systems: Issues and approaches. Proceedings of the IEEE, 74(7):1026--1035, July 1986.
Hanks and Church1989
Patrick Hanks and Kenneth W. Church. Word association norms, mutual information and
lexicography. 27th Annual Meeting of the Association for Comutational Linguistics,
pages 22--29, 1989.
Hatim and Mason1990
Basil Hatim and Ian Mason. Discourse and the Translator. Longman, London, 1990.
Hirschman1986
L. Hirschman. Discovering sublanguage structures. In R. Grishman and R.I. Kittredge,
editors, Analyzing Language in Restricted Domains: Sublanguage Description and
Processing, pages 211--234. Lawrence Erlbaum Associates, New Jersey: Hillsdale, 1986.
Hornby et al. 1974
A.S. Hornby, A.P. Cowie, and A.C. Gimson. Oxford Advanced Learner's Dictionary of
Current English. Oxford University Press, Oxford, 1974.
Hurford and Heasley1983
J.R. Hurford and B. Heasley. Semantics: A Course Book. Cambridge University Press,
Cambridge, 1983.
Hutchins and Somers1992
W.J. Hutchins and H.L. Somers. An Introduction to Machine Translation. Academic
Press, London, 1992.
Hutchins1986
W.J. Hutchins. Machine Translation: Past, Present, Future. Ellis Horwood/Wiley,
Chichester/New York, 1986.
Isabelle1987
P. Isabelle. Machine Translation at the TAUM group. In M. King, editor, Machine
Translation Today: The State of the Art, Proceedings of the Third Lugano Tutorial, 1984,
pages 247--277. Edinburgh University Press, Edinburgh, 1987.
Jordanin press
http://privatewww.essex.ac.uk/~doug/book/node99.html (7 de 16) [20/09/2000 18:04:58]
References

P.W. Jordan. A first-pass approach for evaluating Machine Translation systems. Machine
Translation, 7, in press. Special Issue on Evaluation.
Kaplan et al. 1989
Ronald M. Kaplan, Klaus Netter, Jurgen Wedekind, and
Annie Zaenen. Translation by structural correspondences. In Proceedings of the Fourth
Conference of the European Chapter of the Association for Computational Linguistics,
pages 272--281, University of Manchester, Institute of Science and Technology,
Manchester, UK, 1989.
Katz1978
J.J. Katz. Effability and translation. In F. Guenthner and M. Guenthner-Reutter, editors,
Meaning and Translation, pages 191--234. Duckworth, London, 1978.
Keenan1978
E.L. Keenan. Some logical problems in translation. In F. Guenthner and M.
Guenthner-Reutter, editors, Meaning and Translation, pages 157--190.
Duckworth, London, 1978.
Kempson1977
Ruth Kempson. Semantic Theory. Cambridge University Press, Cambridge, 1977.
King and Falkedal1990
M. King and K. Falkedal. Using test suites in evaluation of MT systems. In COLING-90,
volume 2, pages 211--216, Helsinki, Finland, 1990.
King1987
Margaret King, editor. Machine Translation Today: The State of the Art. Edinburgh
University Press, Edinburgh, 1987.
Kittredge1982
R.I. Kittredge. Variation and homogeneity of sublanguages. In R.I. Kittredge and J.
Lehrberger, editors, Sublanguage, pages 107--137. Walter de Gruyter, Berlin, 1982.
Kittredge1987
R.I. Kittredge. The significance of sublanguage for Automatic Translation. In S.
Nirenburg, editor, Machine Translation Systems, pages 59--67. Cambridge University
Press, Cambridge, 1987.
Knowles1990
Francis E. Knowles. Machine aids for translators: what does the future betoken? In
Pamela Mayorcas, editor, Translating and the Computer 10, pages 125--135. Aslib,
London, 1990.
Krauwerin press
Steven Krauwer. Evaluation of MT systems: A programmatic view. Machine
Translation, 7, in press. Special Issue on Evaluation.
Landsbergen1987a
Jan S.P. Landsbergen. Isomorphic grammars and their use in the Rosetta translation
system. In Margaret King, editor, Machine Translation Today: The State of the Art,
Proceedings of the Third Lugano Tutorial, 1984, pages 351--372. Edinburgh University
Press, Edinburgh, 1987.
Landsbergen1987b

http://privatewww.essex.ac.uk/~doug/book/node99.html (8 de 16) [20/09/2000 18:04:58]

References

Jan S.P. Landsbergen. Montague grammar and Machine Translation. In Peter Whitelock,
Mary McGee Wood, Harold L. Somers, Rod L. Johnson, and Paul Bennett, editors,
Linguistic Theory and Computer Applications, pages 113--147. Academic Press, London,
1987.
Lawson1982a
Veronica Lawson, editor. Practical Experience of Machine Translation. North-Holland,
Amsterdam, 1982.
Lawson1982b
Veronica Lawson, editor. Translating and the Computer 5: Tools for the Trade.
North-Holland, Amsterdam, 1982.
Leech1983
Geoffrey N. Leech. Principles of Pragmatics. Longman Linguistics Library. Longman,
1983.
Lehrberger and Bourbeau1987
John Lehrberger and Laurent Bourbeau. Machine Translation: Linguistic Characteristics
of MT Systems and General Methodology of Evaluation. John Benjamins, Amsterdam,
1987.
Lehrberger1982
J. Lehrberger. Automatic translation and the concept of sublanguage. In R.I. Kittredge
and J. Lehrberger, editors, Sublanguage, pages 81--107. Walter de Gruyter, Berlin, 1982.
Levinson1983
Stephen C. Levinson. Pragmatics. Cambridge Textbooks in Linguistics. Cambridge
University Press, Cambridge, 1983.
Maas1987
Heinz-Dieter Maas. The MT system SUSY. In Margaret King, editor, Machine
Translation Today: The State of the Art, Proceedings of the Third Lugano Tutorial, 1984,
pages 209--246. Edinburgh University Press, Edinburgh, 1987.
Mayorcas1990
Pamela Mayorcas, editor. Translating and the Computer 10: The Translation
Environment 10 Years on. Aslib, London, 1990.
MayorcasForthcoming
Pamela Mayorcas, editor. Translating and the Computer 12. Aslib, London,
Forthcoming.
McCord1989
Michael C. McCord. Design of LMT: A Prolog-based Machine Translation system.
Computational Linguistics, 15(1):33--52, 1989.
McDonald1987
David D. McDonald. Natural Language generation: Complexities and techniques. In
Sergei Nirenberg, editor, Machine Translation: Theoretical and Methodological Issues,
pages 192--224. Cambridge University Press, Cambridge, 1987.
McNaught1988a
J. McNaught. Computational Lexicography and Computational Linguistics.
Lexicographica (Special Issue on Computational Lexicography), 4:19--33, 1988.

http://privatewww.essex.ac.uk/~doug/book/node99.html (9 de 16) [20/09/2000 18:04:58]

References

McNaught1988b
J. McNaught. A survey of term banks worldwide. In C. Picken, editor, Translating and
the Computer, pages 112--129. Aslib, London, 1988.
McNaughtforthcoming
J. McNaught. Toward the neutral lexical database. In D. Walker, A. Zampolli, and N.
Calzolari, editors, Towards a Polytheoretical Lexical Database. Istituto di Linguistica
Computazionale, Pisa, forthcoming.
Melby1987
A. Melby. On human machine interaction in Machine Translation. In Sergei Nirenburg,
editor, Machine Translation: Theoretical and Methodological Issues, pages 145--54.
Cambridge University Press, Cambridge, 1987.
Melby1992
A. Melby. The translator workstation. In John Newton, editor, Computers in Translation:
A Practical Appraisal, pages 147--165. Routledge, London, 1992.
Mel'cuk and Polguere1987
I.A. Mel'cuk and A. Polguere. A formal lexicon in the meaning-text theory.
Computational Linguistics, 13:261--275, 1987.
Mel'cuk and Zholkovsky1988
Igor Mel'cuk and Alexander Zholkovsky. The explanatory combinatory dictionary. In
M.W. Evens, editor, Relational Models of the Lexicon: Representing Knowledge in
Semantic Networks, pages 41--74. CUP, 1988.
Minnisin press
S. Minnis. Constructive Machine Translation evaluation. Machine Translation, 7, in
press. Special Issue on Evaluation.
Nagao et al. July 1986
Makoto Nagao, Jun-ichi Tsujii, and Jun-ichi Nakamura. Machine Translation from
Japanese into English. Proceedings of the IEEE, 74(7):993--1012, July 1986.
Nagao1984
M. Nagao. A framework of a mechanical translation between Japanese and English by
analogy principle. In A. Elithorn and R. Banerji, editors, Artificial and Human
Intelligence, pages 173--180. North Holland, Amsterdam, 1984.
Nagao1986
Makoto Nagao. Machine Translation: How Far Can it Go? Oxford University Press,
Oxford, 1986. Translated by Norman Cook.
Nagao1989
M. Nagao, editor. Machine Translation Summit. Ohmsha Ltd, Tokyo Japan, 1989.
Neal et al. in press
J.G. Neal, E.L. Feit, and A. Montgomery. Benchmark investigation/identification project,
phase 1. Machine Translation, 7, in press. Special Issue on Evaluation.
Nerbonne et al. in press
J. Nerbonne, K. Netter, A.K. Diagne, J. Klein, and L. Dickmann. A diagnostic tool for
German syntax. Machine Translation, 7, in press. Special Issue on Evaluation.

http://privatewww.essex.ac.uk/~doug/book/node99.html (10 de 16) [20/09/2000 18:04:58]

References

Newton1992a
John Newton, editor. Computers in Translation: A Practical Appraisal. Routledge,
London, 1992.
Newton1992b
John Newton. The Perkins experience. In John Newton, editor, Computers in
Translation: A Practical Appraisal, pages 46--57. Routledge,
London, 1992.
Nirenburg1993
S. Nirenburg. Interlangue et traitement du sens dans les systèmes de TA. In P.Bouillon
and A.Clas, editors, Etudes et Recherches en Traductique : Problemes de Traduction par
Ordinateur. Presses de Montreal/AUPELF, Montreal, 1993.
Paillet1990
Alain Paillet. User experience of Termbase. In Pamela Mayorcas, editor, Translating and
the Computer 10, pages 97--108. Aslib, London, 1990.
Patterson1982
B. Patterson. Multilingualism in the European Community. Multilingua, 1(1):4--15,
1982.
Picken1985
C. Picken, editor. Translating and the Computer 6: Translation and Communication.
Aslib, London, 1985.
Picken1986
C. Picken, editor. Translating and the Computer 7. Aslib, London, 1986.
Picken1987
C. Picken, editor. Translating and the Computer 8: A Profession on the Move. Aslib,
London, 1987.
Picken1988
C. Picken, editor. Translating and the Computer 9: Potential and Practice. Aslib,
London, 1988.
Picken1989
C. Picken, editor. The Translator's Handbook. Aslib, London, 2nd edition, 1989.
Picken1990
C. Picken, editor. Translating and the Computer 11: Preparing for the Next Decade.
Aslib, London, 1990.
Pierce and Carroll1966
John R. Pierce and John B. Carroll. Language and Machines -- Computers in Translation
and Linguistics (ALPAC Report). ALPAC, Washington D.C., 1966.
Pollard and Sag1987
Carl Pollard and Ivan Sag. Information Based Syntax and Semantics, volume 1:
Fundamentals of CSLI Lecture Notes, 13. Chicago University Press, Chicago, 1987.
Pollard and Sag1993
Carl Pollard and Ivan Sag. An Information Based Approach to Syntax and Semantics: Vol
2: Agreement, Binding and Control. Chicago University Press, Chicago, 1993.

http://privatewww.essex.ac.uk/~doug/book/node99.html (11 de 16) [20/09/2000 18:04:58]

References

Proctor1978
P. Proctor, editor. Longman's Dictionary of Contemporary English. Longman Group
Ltd., Harlow, UK, 1978.
Pugh1992
Jeanette Pugh. The story so far: An evaluation of Machine Translation in the world
today. In John Newton, editor, Computers in Translation: A Practical Appraisal, pages
14--32. Routledge, London, 1992.
Pullum1991
Geoffrey K. Pullum. The Great Eskimo Vocabulary Hoax. Chicago University Press,
Chicago, 1991.
Pustejovsky and Bergler1992
J. Pustejovsky and S. Bergler, editors. Lexical Semantics and Knowledge Representation.
Springer-Verlag, Berlin, 1992.
Pustejovsky1987
J. Pustejovsky. An integrated theory of discourse analysis. In Sergei Nirenburg, editor,
Machine Translation: Theoretical and Methodological Issues, pages 169--191.
Cambridge University Press, Cambridge, 1987.
Pym1990
D.J. Pym. Pre-editing and the use of simplified writing for MT: An engineer's experience
of operating an MT system. In Pamela Mayorcas, editor, Translating and the Computer
10, pages 80--96. Aslib, London, 1990.
Reichenbach1947
Hans Reichenbach. Elements of Symbolic Logic. University of California Press,
Berkeley, Ca., 1947.
Rich1983
Elaine Rich. Artificial Intelligence. McGraw-Hill Book Co., Singapore, 1983.
Ritchie et al. 1992
G.D. Ritchie, G.J. Russell, A.W. Black, and S.G. Pulman. Computational Morphology:
Practical Mechanisms for the English Lexicon. MIT Press, Cambridge, Mass., 1992.
Ritchie1987
G.D. Ritchie. The lexicon. In Peter Whitelock, Mary McGee Wood, Harold L. Somers,
Rod L. Johnson, and Paul Bennett, editors, Linguistic Theory and Computer
Applications, pages 225--256. Academic Press, London, 1987.
Rogers1986
J.B. Rogers. A Prolog Primer. Addison Wesley, Wokingham, 1986.
Sadler and Arnold1993
Louisa Sadler and D.J. Arnold. Unification and Machine Translation. META,
37(4):657--680, 1993.
Sadler1989
V. Sadler. Working with Analogical Semantics: Disambiguation Techniques in DLT.
Distributed Language Translation, 5. Foris Pub., Dordrecht, 1989.
Sadler1991

http://privatewww.essex.ac.uk/~doug/book/node99.html (12 de 16) [20/09/2000 18:04:59]

References

Louisa Sadler. Structural transfer and unification formalisms. Applied Computer

Translation, 1(4):5--21, 1991.
Sadler1993
Louisa Sadler. Co-description and translation. In F. Van Eynde, editor, Linguistic Issues
in Machine Translation, pages 44--71. Pinter, London, 1993.
Sager1982
J.C. Sager. Types of translation and text forms in the environment of Machine
Translation (MT). In V. Lawson, editor, Practical Experience of Machine Translation,
pages 11--19. North Holland Publishing Co, Dordrecht, 1982.
Sager1990
J.C. Sager. A Practical Course in Terminology Processing. John Benjamins, 1990.
Schenk1986
A. Schenk. Idioms in the Rosetta Machine Translation system. Proceedings of COLING
86, pages 319--324, 1986.
Schubert1992
Klaus Schubert. Esperanto as an intermediate language for Machine Translation. In John
Newton, editor, Computers in Translation: A Practical Appraisal, pages 78--95.
Routledge, London, 1992.
Shapiro1987
S.C. Shapiro, editor. Encyclopaedia of AI. Wiley, New York, 1987.
Sharp1988
Randall Sharp. CAT-2 -- implementing a formalism for multi-lingual MT. In
Proceedings of the 2nd International Conference on Theoretical and Methodological
Issues in Machine Translation of Natural Languages, pages 76--87, Carnegie Mellon
University, Centre for Machine Translation, Pittsburgh, USA, 1988.
Shiwenin press
Yu Shiwen. Automatic evaluation of output quality for machine translation systems.
Machine Translation, 7, in press. Special Issue on Evaluation.
Sinclair1987
J. M. Sinclair, editor. Looking Up: An Account of the COBUILD Project in Lexical
Computing. Collins ELT, London, 1987.
Slocum et al. 1987
J. Slocum, W.S. Bennett, J. Bear, M. Morgan, and R. Root. Metal: The LRC Machine
Translation system. In M. King, editor, Machine Translation Today, pages 319--350.
Edinburgh University Press, Edinburgh, 1987.
Slocum1986
J. Slocum. How one might identify and adapt to a sublanguage: An initial exploration. In
R. Grishman and R.I. Kittredge, editors, Analyzing Language in Restricted Domains:
Sublanguage Description and Processing, pages 195--210. Lawrence Erlbaum
Associates, New Jersey: Hillsdale, 1986.
Slocum1988
J. Slocum, editor. Machine Translation Systems. Cambridge University Press,
Cambridge, 1988.

http://privatewww.essex.ac.uk/~doug/book/node99.html (13 de 16) [20/09/2000 18:04:59]

References

Snell1979
Barbara Snell, editor. Machine Aids for Translators: Translating and the Computer 2.
North Holland, Amsterdam, 1979.
Snell1982
Barbara Snell, editor. Term Banks for Tomorrow's World: Translating and the Computer
4. Aslib, London, 1982.
Somers1992
Harold L. Somers. Current research in Machine Translation. In John Newton, editor,
Computers in Translation: A Practical Appraisal, pages 189--207. Routledge, London,
1992.
Sowa1984
John F. Sowa. Conceptual Structures: Information Processing in Mind and Machine.
Addison Wesley, Reading, Mass., 1984.
Spencer1991
A. Spencer. Morphological Theory. Basil Blackwell, Oxford, 1991.
Stoll1988
C.H. Stoll. Translation tools on PC. In C. Picken, editor, Translating and the Computer
9: Potential and Practice, pages 11--26. Aslib, 1988.
Sumita and Iida1991
E. Sumita and H. Iida. Experiments and prospects of example-based Machine
Translation. In Proceedings of 29th ACL, pages 185--192, University of California,
Berkeley, Ca., 1991.
Sumita et al. 1990
Eiichiro Sumita, Hitoshi Iida, and Hideo Kohyama. Translating with examples: A new
approach to Machine Translation. In Proceedings of the 3rd International Conference on
Theoretical and Methodological Issues in Machine Translation of Natural Language,
pages 203--212, Linguistics Research Centre, University of Texas at Austin, USA, 1990.
Teller et al. 1988
V. Teller, M. Kosaka, and R. Grishman. A comparative study of Japanese and English
sublanguage patterns. In S. Nirenburg, editor, Proceedings of the Second Conference on
Theoretical and Methodological Issues in MT, Pittsburg, Pa, 1988. Carnegie Mellon
University.
Tennant1981
H. Tennant. Natural Language Processing: An Introduction to an Emerging Technology.
Petrocelli Pub., New York, 1981.
Thomas1992
Patricia Thomas. Computerized term banks and translation. In John Newton, editor,
Computers in Translation: A Practical Appraisal, pages 131--146. Routledge, London,
1992.
TMI11985
Proceedings of the Conference on Theoretical and Methodological Issues in Machine
Translation of Natural Languages, Colgate University, Hamilton, New York, 1985.
TMI21988

http://privatewww.essex.ac.uk/~doug/book/node99.html (14 de 16) [20/09/2000 18:04:59]

References

Second International Conference on Theoretical and Methodological Issues in Machine

Translation of Natural Languages, Carnegie Mellon University, Pittsburgh, 1988.
TMI31990
Third International Conference on Theoretical and Methodological
Issues in Machine Translation of Natural Language, University of Texas, Austin, 1990.
TMI41992
Fourth International Conference on Theoretical and Methodological Issues in Machine
Translation, CCRIT - CWARC, Montré al, Canada, 1992.
Tsujii et al. 1992
Jun-ichi Tsujii, Sofia Ananiadou, Iris Arad, and Satoshi Sekine. Linguistic knowledge
acquisition from corpora. In Sofia Ananiadou, editor, International Workshop on
Fundamental Research for the Future Generation of Natural Language Processing
(FGNLP), pages 61--81. CEC, UMIST, Manchester, UK, 1992.
Van Eynde1993a
F. Van Eynde. Machine Translation and linguistic motivation. In F. Van Eynde, editor,
Linguistic Issues in Machine Translation, pages 1--43. Pinter, London, 1993.
Van Eynde1993b
Frank Van Eynde, editor. Linguistic Issues in Machine Translation. Pinter, London,
1993.
van Herwijnen1990
Eric van Herwijnen. Practical SGML. Kluwer Academic Publishers, Dordrecht, 1990.
van Noord et al. 1990
G. van Noord, J. Dorrepaal, P. van der Eijk, M. Florenza, and L. des Tombe. The MiMo2
research system. In Third International Conference on Theoretical and Methodological
Issues in Machine Transation 11-13 June 1990, pages 213--124, Linguistics Research
Center, Austin, Texas, 1990.
van Slype1982
G. van Slype. Conception d'une méthodologie générale d'Evaluation de la Traduction
Automatique. Multilingua, 1(4):221--237, 1982.
Vasconcellos1988
M. Vasconcellos, editor. Technology as Translation Strategy. State University of New
York, Binghampton, 1988.
Vauquois and Boitet1985
Bernard Vauquois and Christian Boitet. Automated translation at Grenoble University.
Computational Linguistics, 11(1):28--36, 1985.
Vinay and Darbelnet1977
J.P. Vinay and J. Darbelnet. Stylistique Comparé du Français et de l'Anglais. Didier,
Paris, 1977.
Warwick1987
Susan Warwick. An overview of post-ALPAC developments. In Margaret King, editor,
Machine Translation Today: The State of the Art, Proceedings of the Third Lugano
Tutorial, 1984, pages 22--37. Edinburgh University Press, Edinburgh, 1987.
Whitelock1992
http://privatewww.essex.ac.uk/~doug/book/node99.html (15 de 16) [20/09/2000 18:04:59]
References

P. Whitelock. Shake-and-bake translation. Proceedings of the 15th International

Conference on Computational Linguistics (COLING 92), volume II, pages 784--791,
1992.
Wilks1973
Y.A. Wilks. An Artificial Intelligence approach to Machine Translation. In R.C. Schank
and K.M. Colby, editors, Computer Models of Thought and Language, pages 114--151.
Freeman, San Francisco, 1973.
Wilks1992
Yorick Wilks. Systran: It obviously works but how much can it be improved? In John
Newton, editor, Computers in Translation: A Practical Appraisal, pages 166--188.
Routledge, London, 1992.
Winograd1983
T. Winograd. Language as a Cognitive Process. Addison Wesley, Reading, Mass., 1983.

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node99.html (16 de 16) [20/09/2000 18:04:59]

 Index

Next: About this document Up: Machine Translation Previous: References

Index
?????
Further Reading
accuracy
Popular Conceptions and , The Scenario, Linguistic Knowledge Architectures, Accuracy,
Accuracy, The Test Suite, Further Reading
acronym
SGML Markup and , SGML Markup and
agreement
Further Analysis: Grammatical , Further Analysis: Grammatical , Further Analysis:
Grammatical , Further Analysis: Grammatical , Ambiguity, Ambiguity, Error Analysis,
Error Analysis, Error Analysis
Aijmer, K.
Further Reading
Albisser, D.
Further Reading
Allegranza, V.
Further Reading
Allen, J.
Further Reading
Allerton, D.J.
Further Reading
ALPAC
A Bit of , A Bit of , A Bit of , A Bit of , Further Reading, Intelligibility, Operational
Evaluation, Further Reading
ALPS system
Popular Conceptions and , A Bit of
Altenberg, B.
Further Reading
ambiguity
Document Preparation: Authoring , Interaction in Translation, Derivation, Compounds,

http://privatewww.essex.ac.uk/~doug/book/node100.html (1 de 26) [20/09/2000 18:05:11]

 Compounds, Compounds, Terminology, Terminology, Terminology, Introduction,
Ambiguity, Ambiguity, Multiword units: Idioms , Summary, Further Reading,
Semantics, Semantics, Pragmatics, Controlled Languages, Controlled Languages,
Sublanguage MT, Sublanguage MT, Knowledge-Based MT, Feasibility of General ,
Feasibility of General , Feasibility of General , Resources for MT, Statistical MT, Further
Reading
analysis
Linguistic Knowledge Architectures, Linguistic Knowledge Architectures, Linguistic
Knowledge Architectures, Linguistic Knowledge Architectures, Linguistic Knowledge
Architectures, Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Interlinguas, Interlinguas, Interlinguas,
Interlinguas, Introduction, Types of Word , Terminology, Multiword units: Idioms ,
Multiword units: Idioms , Multiword units: Idioms , Flexible or Multi-level , Flexible or
Multi-level , Flexible or Multi-level , Flexible or Multi-level , Flexible or Multi-level ,
Flexible or Multi-level , Flexible or Multi-level , Flexible or Multi-level
anaphoric pronouns
Pragmatics, Pragmatics, Pragmatics, Pragmatics, Pragmatics, Real World Knowledge,
Knowledge-Based MT, Feasibility of General , Resources for MT
antonym
Terminology
Apresian, J.
Further Reading
Apresian, Y.
A Bit of
ARIANE
Further Reading
Arnold, D.J.
Further Reading, Further Reading, Further Reading, Further Reading, Further Reading,
Further Reading, Further Reading, Further Reading
Aslib
Further Reading
aspect
Lexical and Structural , Semantics, Semantics, Semantics, Summary
Balkan, L.
Further Reading
Bar-Hillel, Y.
A Bit of , Further Reading
Barr, A.

http://privatewww.essex.ac.uk/~doug/book/node100.html (2 de 26) [20/09/2000 18:05:11]

 Further Reading
Basic English
Controlled Languages, Controlled Languages, Controlled Languages
batch
The Scenario
Bauer, L.
Further Reading
Beaven, J.L.
Further Reading
Bennett, P.
Further Reading
Bennett,;tex2html_html_special_mark_quot;P.
Further Reading
Benson,;tex2html_html_special_mark_quot;M.
Further Reading
Bergler, S.
Further Reading
Berlin, B.
Further Reading
Boguraev, B.
Further Reading
Boitet, C.
Further Reading
Booth, A.D.
A Bit of
Borsley, R.D.
Further Reading
bottom-up parsing
Parsing, Parsing, Parsing, Introduction
Bourbeau, L.
Further Reading
bracket notation
Grammars and Constituent
Bratko, I.
Further Reading
Brigham Young University
A Bit of

http://privatewww.essex.ac.uk/~doug/book/node100.html (3 de 26) [20/09/2000 18:05:11]

 Briscoe, E.
Further Reading
British Aerospace
Controlled Languages
British National Corpus
Resources for MT
Brown, E.K.
Further Reading
Brown, P.F.
Further Reading
Bruce, B.
Further Reading
Buchmann,B.
Further Reading
Canadian Meteorological Center
Popular Conceptions and
Canadian Meteorological Center, Dorval
Popular Conceptions and
Cann, R.
Further Reading
Carnegie Mellon University, Pittsburgh
Knowledge-Based MT
Carroll, J.B.
Further Reading, Further Reading
CAT-2
Further Reading
Chandioux, J.
Further Reading
Charniak, E.
Further Reading, Further Reading
Chierchia, G.
Further Reading
Church, K.W.
Further Reading, Further Reading, Further Reading
CLAWS
Resources for MT
COLING

http://privatewww.essex.ac.uk/~doug/book/node100.html (4 de 26) [20/09/2000 18:05:11]

 Further Reading
Collins-Robert English-French dictionary
Paper Dictionaries
collocation
Introduction, Multiword units: Idioms , Multiword units: Idioms , Multiword units:
Idioms , Summary, Further Reading
common sense knowledge
Introduction, Meaning, Meaning, Meaning, Introduction, Real World Knowledge, Real
World Knowledge, Further Reading
compound
Paper Dictionaries, Paper Dictionaries, Paper Dictionaries, Types of Word , Dictionaries
and Morphology, Dictionaries and Morphology, Dictionaries and Morphology,
Compounds, Compounds, Compounds, Compounds, Compounds, Compounds,
Compounds, Compounds, Compounds, Compounds, Terminology, Further Reading
constituent structure
Representing Linguistic Knowledge, Representing Linguistic Knowledge,
Grammars\indexgrammar and Constituent Structure\indexconstituent , Grammars and
Constituent , Grammars and Constituent , Grammars and Constituent , Grammars and
Constituent , Grammars and Constituent , Grammars and Constituent , Grammars and
Constituent , Grammars and Constituent , Grammars and Constituent , Grammars and
Constituent , Grammars and Constituent , Further Analysis: Grammatical , Further
Analysis: Grammatical , Further Analysis: Grammatical , Further Analysis: Grammatical
, Further Analysis: Grammatical , Further Analysis: Grammatical , Further Analysis:
Grammatical , Further Analysis: Grammatical , Further Analysis: Grammatical , Further
Analysis: Grammatical , Further Analysis: Grammatical , Further Analysis: Grammatical
, Further Analysis: Grammatical , Processing, Parsing, Parsing, Parsing, Parsing,
Generation, Generation, Generation, Further Reading, Flexible or Multi-level
constituent;tex2html_html_special_mark_quot;structure
Grammars and Constituent , Generation, Generation
constraint-based MT
Flexible or Multi-level , Flexible or Multi-level , Further Reading
context
Introduction, Introduction, Pragmatics, Pragmatics, Real World Knowledge, Real World
Knowledge
controlled language
Introduction, Document Preparation: Authoring , Document Preparation: Authoring ,
Introduction, Controlled Languages, Controlled Languages, Further Reading
controlled languages
Document Preparation: Authoring , Document Preparation: Authoring , Document
Preparation: Authoring

http://privatewww.essex.ac.uk/~doug/book/node100.html (5 de 26) [20/09/2000 18:05:11]

 Copeland, C.
Further Reading
corpora
Introduction, Feasibility of General , Resources for MT
critiquing systems
Document Preparation: Authoring , Document Preparation: Authoring
Danlos,;tex2html_html_special_mark_quot;L.
Further Reading
Darbelnet, J.
Further Reading
database
Types of Word , Terminology, Further Reading
declarativity
Comparative Grammar and , Types of Word
des Tombe, L.
Further Reading
dictionaries
Introduction, Transformer Architectures, Summary, Feasibility of General , Feasibility of
General , Resources for MT, Resources for MT, Example-Based Translation
dictionaries(bilingual)
Transformer Architectures, Transformer Architectures, Transformer Architectures,
Transformer Architectures, Transformer Architectures, Transformer Architectures,
Transformer Architectures, Transformer Architectures, Comparative Grammar and ,
Comparative Grammar and , Comparative Grammar and , Comparative Grammar and ,
Comparative Grammar and , Comparative Grammar and , Introduction, Introduction,
Paper Dictionaries, Paper Dictionaries, Types of Word , Types of Word , Types of Word
, Terminology
dictionaries(electronic)
The Scenario, Document Preparation: Authoring , Dictionary-Based Translation Support
, Dictionary-Based Translation Support , Dictionary-Based Translation Support ,
Dictionary-Based Translation Support , Introduction
dictionaries(machine readable)
Introduction, Introduction, Resources for MT, Resources for MT
dictionaries(monolingual)
Transformer Architectures, Transformer Architectures, Transformer Architectures,
Transformer Architectures, Comparative Grammar and , Introduction, Introduction,
Paper Dictionaries, Paper Dictionaries, Paper Dictionaries, Paper Dictionaries, Types of
Word , Inflection, Inflection, Inflection, Terminology, Further Reading, Further Reading
dictionaries(on-line)

http://privatewww.essex.ac.uk/~doug/book/node100.html (6 de 26) [20/09/2000 18:05:11]

 The Scenario, Further Reading, Feasibility of General
dictionaries(term)
The Scenario, The Scenario, Dictionary-Based Translation Support , Dictionary-Based
Translation Support , Dictionary-Based Translation Support , Further Reading
dictionary(term)
Controlled Languages, Sublanguage MT
discourse
Ambiguity, Pragmatics, Pragmatics, Pragmatics, Pragmatics, Pragmatics, Pragmatics,
Pragmatics, Pragmatics, Pragmatics, Pragmatics, Pragmatics, Pragmatics, Pragmatics,
Pragmatics, Pragmatics, Further Reading, Resources for MT
DLT
Further Reading
Document Type Definition (DTD)
Basic Ideas, Basic Ideas, SGML Markup and , SGML Markup and , SGML Markup and
domain knowledge
Real World Knowledge, Real World Knowledge, Summary, Knowledge-Based MT,
Knowledge-Based MT, Knowledge-Based MT, Knowledge-Based MT
Dutch
Derivation, Lexical and Structural , Multiword units: Idioms , Multiword units: Idioms
Dyvik, H.
Further Reading
electronic document
Introduction, The Scenario, The Scenario, The Scenario, The Scenario, Introduction,
SGML Markup and , Summary
ELU
Further Reading
empirical approaches to MT
Introduction, Flexible or Multi-level , Feasibility of General , Feasibility of General ,
Empirical Approaches to , Statistical MT, Summary, Further Reading
ENGSPAN
Popular Conceptions and
error analysis
Document Revision, Error Analysis, Error Analysis
Esperanto
Interlinguas, Further Reading
Estival, D.
Further Reading
ETAP-2

http://privatewww.essex.ac.uk/~doug/book/node100.html (7 de 26) [20/09/2000 18:05:11]

 Further Reading
EURATOM
A Bit of
Eurodicautom
Terminology
European Communities/CEC
Why MT Matters, Why MT Matters, Popular Conceptions and , A Bit of , Terminology
EUROTRA
A Bit of , Further Reading
evaluation
Introduction
Evens, M.W.
Further Reading
example-based MT
Flexible or Multi-level , Example-Based Translation, Example-Based Translation,
Further Reading
Falkedal, K.
Further Reading
feature
Further Analysis: Grammatical , Further Analysis: Grammatical , Comparative Grammar
and , Comparative Grammar and , Comparative Grammar and , Comparative Grammar
and , Types of Word , Types of Word , Types of Word , Types of Word , Types of Word
, Types of Word , Types of Word , Inflection, Inflection, Inflection, Inflection, Inflection,
Inflection, Derivation, Terminology, Lexical and Structural , Multiword units: Idioms
Fiegenbaum, E.A.
Further Reading
Finnish
Inflection, Inflection
Flank, S.
Further Reading
flexible MT
Flexible or Multi-level , Flexible or Multi-level , Further Reading
focus
Pragmatics, Pragmatics, Pragmatics
frame
Real World Knowledge, Knowledge-Based MT, Knowledge-Based MT,
Knowledge-Based MT, Further Reading
Frantext

http://privatewww.essex.ac.uk/~doug/book/node100.html (8 de 26) [20/09/2000 18:05:11]

 Terminology
French
Why MT Matters, Popular Conceptions and , Popular Conceptions and , A Bit of , A Bit
of , Comparative Grammar and , Comparative Grammar and , Comparative Grammar and
, Comparative Grammar and , Comparative Grammar and , Comparative Grammar and ,
Interlinguas, Interlinguas, Interlinguas, Interlinguas, Paper Dictionaries, Types of Word ,
Types of Word , Derivation, Terminology, Terminology, Ambiguity, Ambiguity,
Ambiguity, Lexical and Structural , Lexical and Structural , Lexical and Structural ,
Lexical and Structural , Lexical and Structural , Lexical and Structural , Lexical and
Structural , Lexical and Structural , Lexical and Structural , Multiword units: Idioms ,
Multiword units: Idioms , Multiword units: Idioms , Multiword units: Idioms , Multiword
units: Idioms , Multiword units: Idioms , Multiword units: Idioms , Multiword units:
Idioms , Semantics, Semantics, Semantics, Semantics, Pragmatics, Pragmatics,
Pragmatics, Pragmatics, Pragmatics, SGML Markup and , Sublanguage MT,
Sublanguage MT, Sublanguage MT, Sublanguage MT, Sublanguage MT, Flexible or
Multi-level , Flexible or Multi-level , Resources for MT, Statistical MT
fully automatic high quality machine translation (FAHQMT)
A Bit of , A Bit of , A Bit of , A Bit of
Furuse, O.
Further Reading
Gale,;tex2html_html_special_mark_quot;W.A.
Further Reading
Garside, R.
Further Reading
Gazdar, G.
Further Reading, Further Reading
gender
Paper Dictionaries, Types of Word , Pragmatics, Pragmatics, Pragmatics, Real World
Knowledge
general knowledge
Introduction
generation
Processing, Generation\indexgeneration, Generation, Generation, Generation,
Generation, Generation, Generation, Summary, Further Reading, Further Reading
Georgetown University
A Bit of
German
The Scenario, The Scenario, Document Revision, Transformer Architectures,
Transformer Architectures, Transformer Architectures, Transformer Architectures,
Transformer Architectures, Transformer Architectures, Transformer Architectures,

http://privatewww.essex.ac.uk/~doug/book/node100.html (9 de 26) [20/09/2000 18:05:12]

 Comparative Grammar and , Comparative Grammar and , Comparative Grammar and ,
Comparative Grammar and , Comparative Grammar and , Interlinguas, Compounds,
Compounds, Compounds, Lexical and Structural , Multiword units: Idioms , Multiword
units: Idioms , Real World Knowledge, Basic Ideas, Sublanguage MT, Sublanguage MT,
Error Analysis, The Test Suite, The Test Suite, The Test Suite
Germanic languages
Dictionaries and Morphology
GETA
A Bit of , A Bit of
GLOBALINK
Popular Conceptions and
Goodman, K.
Further Reading, Further Reading
Goodman,;tex2html_html_special_mark_quot;K.
Further Reading
grammar
Popular Conceptions and , Grammars\indexgrammar and Constituent
Structure\indexconstituent , Meaning, Transformer Architectures, Transformer
Architectures, Transformer Architectures, Transformer Architectures, Linguistic
Knowledge Architectures, Linguistic Knowledge Architectures, Linguistic Knowledge
Architectures, Linguistic Knowledge Architectures, Linguistic Knowledge Architectures,
Linguistic Knowledge Architectures, Linguistic Knowledge Architectures, Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Interlinguas, LK Engines Summarised, LK Engines Summarised, LK
Engines Summarised, LK Engines Summarised, LK Engines Summarised
grammar checkers
Document Preparation: Authoring
grammatical relation
Representing Linguistic Knowledge, Further Analysis: Grammatical , Further Analysis:
Grammatical , Further Analysis: Grammatical , Further Analysis: Grammatical , Further
Analysis: Grammatical , Further Analysis: Grammatical , Further Analysis: Grammatical
, Further Analysis: Grammatical , Further Analysis: Grammatical , Further Analysis:
Grammatical , Further Analysis: Grammatical , Further Analysis: Grammatical , Further
Analysis: Grammatical , Further Analysis: Grammatical , Meaning, Parsing, Parsing,
Parsing, Generation, Generation, Generation, Comparative Grammar and , Comparative
Grammar and , Types of Word , Types of Word , Types of Word , Lexical and Structural
, Lexical and Structural , Lexical and Structural , Multiword units: Idioms , Multiword
units: Idioms , Semantics, Flexible or Multi-level
grammatical;tex2html_html_special_mark_quot;relation

http://privatewww.essex.ac.uk/~doug/book/node100.html (10 de 26) [20/09/2000 18:05:12]

 Further Analysis: Grammatical Relations\indexgrammatical , Parsing
Grimaila, A.
Further Reading
Grishman, R.
Further Reading, Further Reading
Grosz, B.J.
Further Reading
Guida, G.
Further Reading
Hanks,;tex2html_html_special_mark_quot;P.
Further Reading
Hansard
Resources for MT, Statistical MT
Hatim, B.
Introduction, Evaluation of Engine
Heasley, B.
Further Reading, Further Reading
Hirschman, L.
Further Reading
homography
Types of Word
human aided machine;tex2html_html_special_mark_quot;translation (HAMT)
Further Reading
Hurford, J.R.
Further Reading, Further Reading
Hutchins, W.J.
Further Reading, Further Reading, Further Reading, Further Reading, Further Reading
Hutchins,;tex2html_html_special_mark_quot;W.J.
Further Reading, Further Reading, Further Reading
hypertext
The Scenario
hyponyms
Terminology
IBM
Feasibility of General , Statistical MT, Statistical MT, Further Reading
idiom
Paper Dictionaries, Paper Dictionaries, Paper Dictionaries, Paper Dictionaries, Types of

http://privatewww.essex.ac.uk/~doug/book/node100.html (11 de 26) [20/09/2000 18:05:12]

 Word , Types of Word , Types of Word , Introduction, Introduction, Lexical and
Structural , Lexical and Structural , Multiword units: Idioms , Further Reading,
Sublanguage MT
Iida, H.
Further Reading
indexical expression
Pragmatics
inference
Real World Knowledge, Real World Knowledge, Real World Knowledge, Further
Reading, Knowledge-Based MT, Knowledge-Based MT
inheritance
Types of Word , Types of Word
intelligibility
The Scenario, Document Revision, Transformer Architectures, Transformer
Architectures, Intelligibility, Intelligibility, Intelligibility, Accuracy, Accuracy,
Accuracy, Accuracy, Accuracy, Accuracy, Accuracy, Accuracy, Error Analysis, The Test
Suite, Operational Evaluation
interactive MT
Interaction\indexinteractive MT in Translation, Interaction in Translation, Summary,
Further Reading
interlingua-based systems
Introduction, Interlinguas, Interlinguas, Interlinguas, Interlinguas, Interlinguas,
Interlinguas, Interlinguas, Interlinguas, Interlinguas, Interlinguas, Interlinguas,
Interlinguas, Interlinguas, Interlinguas, Interlinguas, Interlinguas, Interlinguas, Summary,
Further Reading, Introduction, Introduction, Terminology, Terminology, Multiword
units: Idioms , Multiword units: Idioms , Semantics, Semantics, Flexible or Multi-level ,
Flexible or Multi-level , Knowledge-Based MT
Irish
Further Analysis: Grammatical
Isabelle, P.
Further Reading, Further Reading
Italian
A Bit of , Lexical and Structural
Japanese
Further Analysis: Grammatical , Interlinguas, Interlinguas, Lexical and Structural ,
Lexical and Structural , Lexical and Structural , Lexical and Structural , Further Reading,
Pragmatics, Basic Ideas, Sublanguage MT, Knowledge-Based MT, Knowledge-Based
MT, Example-Based Translation, Example-Based Translation
Jordan, P.W.

http://privatewww.essex.ac.uk/~doug/book/node100.html (12 de 26) [20/09/2000 18:05:12]

 Further Reading
KANT
Knowledge-Based MT
Kaplan, R.M.
Further Reading
Katz, J.J.
Further Reading
Kay, P.
Further Reading
KBMT
Further Reading
Keenan, E.L.
Further Reading
Kempson, R.
Further Reading, Further Reading
King, M.
Further Reading, Further Reading
Kittredge, R.I.
Further Reading, Further Reading
knowledge representation
Real World Knowledge, Real World Knowledge
knowledge-based systems
Real World Knowledge, Knowledge-Based MT, Knowledge-Based MT,
Knowledge-Based MT, Further Reading
Knowles, F. E.
Further Reading
Krauwer, S.
Further Reading
Landsbergen,;tex2html_html_special_mark_quot;J.S.P.
Further Reading
Lawson, V.
Further Reading, Further Reading
LDOCE
Further Reading
Leech, G.N.
Further Reading
Lehrberger, J.

http://privatewww.essex.ac.uk/~doug/book/node100.html (13 de 26) [20/09/2000 18:05:12]

 Further Reading, Further Reading
Levinson, S.C.
Further Reading
lexical decomposition
Semantics, Semantics, Semantics
lexical function
Multiword units: Idioms , Multiword units: Idioms , Further Reading
lexical hole
Types of Word , Lexical and Structural , Lexical and Structural , Multiword units: Idioms
, Multiword units: Idioms , Multiword units: Idioms
linguistic knowledge systems
Introduction, Introduction, Introduction, Linguistic Knowledge Architectures, Linguistic
Knowledge Architectures, Linguistic Knowledge Architectures, Linguistic Knowledge
Architectures, Linguistic Knowledge Architectures, Linguistic Knowledge Architectures,
Comparative Grammar and , Comparative Grammar and , Interlinguas, LK Engines
Summarised, LK Engines Summarised, LK Engines Summarised, LK Engines
Summarised, LK Engines Summarised, Summary, Summary, The Test Suite, The Test
Suite, Introduction, Feasibility of General
LMT
Further Reading
LOGOS
Popular Conceptions and
Maas, H.D.
Further Reading
machine aided human translation;tex2html_html_special_mark_quot;(MAHT)
Further Reading
Mason, I.
Introduction, Evaluation of Engine
Mauri, G.
Further Reading
Mayorcas, P.
Further Reading
McConnell-Ginet, S.
Further Reading
McCord, M.C.
Further Reading
McDermott, D.
Further Reading

http://privatewww.essex.ac.uk/~doug/book/node100.html (14 de 26) [20/09/2000 18:05:12]

 McDonald, D.D.
Further Reading
McNaught, J.
Further Reading
Mel'cuk, I.A.
A Bit of
Mel'cuk,;tex2html_html_special_mark_quot;I.A.
Further Reading
Melby, A.
Further Reading
Mellish, C.
Further Reading, Further Reading
METAL
Popular Conceptions and , A Bit of , Further Reading
metaphor
Semantics, Knowledge-Based MT
METEO
Popular Conceptions and , Popular Conceptions and , Popular Conceptions and , A Bit of
, A Bit of , Further Reading, Further Reading, Sublanguage MT, Sublanguage MT
metonymy
Knowledge-Based MT
Miller, J.E.
Further Reading
MIMO
Further Reading
MIMO-2
Further Reading
Minnis, S.
Further Reading
mismatch(lexical)
Terminology, Introduction, Lexical and Structural , Lexical and Structural , Summary,
Further Reading
mismatch(structural)
Introduction, Lexical and Structural , Lexical and Structural , Summary, Further Reading
modularity
LK Engines Summarised
morphology
Introduction, Transformer Architectures, Transformer Architectures, Transformer

http://privatewww.essex.ac.uk/~doug/book/node100.html (15 de 26) [20/09/2000 18:05:12]

 Architectures, Dictionaries and Morphology\indexmorphology, Inflection, Summary,
Further Reading, Further Reading, Multiword units: Idioms , Multiword units: Idioms ,
Flexible or Multi-level , Statistical MT, Statistical MT
morphology(derivational)
Dictionaries and Morphology, Dictionaries and Morphology, Derivation, Derivation,
Derivation, Compounds, Compounds, Compounds
morphology(inflectional)
Dictionary-Based Translation Support , Transformer Architectures, Transformer
Architectures, Transformer Architectures, Transformer Architectures, Dictionaries and
Morphology, Dictionaries and Morphology, Inflection, Inflection, Inflection, Inflection,
Inflection, Inflection, Inflection, Inflection, Derivation, Derivation
MU
Further Reading
multilevel MT
Flexible or Multi-level , Flexible or Multi-level
Multinational Customised English (MCE)
Controlled Languages
Nagao, M.
Further Reading, Further Reading, Further Reading, Further Reading
NASA
A Bit of
Neal, J.G.
Further Reading
Nerbonne, J.
Further Reading
Newton, J.
Further Reading, Further Reading, Further Reading
Nirenburg, S.
Further Reading, Further Reading, Further Reading, Further Reading
Normaterm
Terminology
Norwegian
Flexible or Multi-level , Flexible or Multi-level , Flexible or Multi-level , Further
Reading
number
Types of Word , Types of Word , Types of Word , Types of Word , Inflection, Inflection,
Pragmatics
OALD

http://privatewww.essex.ac.uk/~doug/book/node100.html (16 de 26) [20/09/2000 18:05:12]

 Paper Dictionaries, Types of Word , Further Reading
orthography
The Scenario, Introduction, Inflection, Derivation, Compounds, Flexible or Multi-level ,
Flexible or Multi-level
Oxford English Dictionary
Paper Dictionaries
Paillet, A.
Further Reading
Pan American Health Organization (PAHO)
A Bit of
parsing
Processing, Parsing, Parsing, Generation, Generation, Generation, Summary,
Transformer Architectures, Linguistic Knowledge Architectures, Linguistic Knowledge
Architectures, Comparative Grammar and , Comparative Grammar and , LK Engines
Summarised
parsing(partial)
Transformer Architectures, Transformer Architectures, Transformer Architectures,
Transformer Architectures, Transformer Architectures, Transformer Architectures,
Transformer Architectures
Patterson, B.
Why MT Matters
Perkins Approved Clear English (PACE)
Controlled Languages, Controlled Languages, Controlled Languages, Further Reading
Perkins Engines
Popular Conceptions and , Controlled Languages
phonetics
Paper Dictionaries
phonology
Introduction, Terminology, Terminology
Picken, C.
Further Reading, Further Reading
Pierce, J.R.
Further Reading, Further Reading
Polguere, A.
Further Reading
Pollard, C.
Further Reading, Further Reading
Pollard,;tex2html_html_special_mark_quot;C.

http://privatewww.essex.ac.uk/~doug/book/node100.html (17 de 26) [20/09/2000 18:05:12]

 Further Reading
post-editing
Popular Conceptions and , Popular Conceptions and , The Scenario, The Scenario, The
Scenario, The Scenario, The Scenario, The Scenario, The Scenario, The Scenario,
Document Preparation: Authoring , Document Preparation: Authoring , Document
Revision, Document Revision, Document Revision, Document Revision, Summary,
Further Reading, SGML Markup and , Some Central Issues, Some Central Issues, Some
Central Issues, Evaluation of Engine , Accuracy, Accuracy, Error Analysis, Operational
Evaluation, Operational Evaluation, Operational Evaluation
pragmatics
Meaning, Meaning, Meaning, Meaning, Meaning, Summary, Ambiguity, Ambiguity,
Ambiguity, Introduction, Introduction, Introduction, Pragmatics, Pragmatics, Pragmatics,
Summary, Summary, Further Reading, Knowledge-Based MT
pre-editing
The Scenario, Document Preparation: Authoring and , Further Reading, Further Reading
Prolog
Why MT Matters, Further Reading
pronunciation
Paper Dictionaries, Paper Dictionaries, Types of Word , Terminology
Pugh, J.
Further Reading
Pullum, G.K.
Further Reading
Pustejovsky, J.
Further Reading, Further Reading
Pym, D.J.
Further Reading, Further Reading, Evaluation of Engine
quality
Popular Conceptions and , Popular Conceptions and , Popular Conceptions and , A Bit of
, A Bit of , The Scenario, Document Revision, Document Revision, Document Revision,
Document Revision, Document Revision, Summary, Summary, Further Reading,
Transformer Architectures, Linguistic Knowledge Architectures, LK Engines
Summarised, Sublanguage MT, Summary, Introduction, Some Central Issues, Some
Central Issues, Evaluation of Engine , Evaluation of Engine , Accuracy, Operational
Evaluation, Operational Evaluation, Operational Evaluation, Operational Evaluation,
Knowledge-Based MT, Knowledge-Based MT, Feasibility of General , Feasibility of
General , Feasibility of General , Example-Based Translation
real world knowledge
Popular Conceptions and , Meaning, Meaning, Meaning, Meaning, Summary,
Ambiguity, Ambiguity, Ambiguity, Introduction, Introduction, Introduction,

http://privatewww.essex.ac.uk/~doug/book/node100.html (18 de 26) [20/09/2000 18:05:12]

 Introduction, Real World Knowledge, Real World Knowledge, Real World Knowledge,
Real World Knowledge, Real World Knowledge, Real World Knowledge, Real World
Knowledge, Summary, Summary, Further Reading
Reichenbach, H.
Further Reading
reversibility
Further Analysis: Grammatical , Transformer Architectures, Transformer Architectures,
Linguistic Knowledge Architectures, Linguistic Knowledge Architectures, Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Types of Word , Types of Word
rewrite rules
Grammars and Constituent
rewrite;tex2html_html_special_mark_quot;rules
Grammars and Constituent
Rich, E.
Further Reading, Further Reading
Ritchie, G.D.
Further Reading, Further Reading
Rockefeller Foundation
A Bit of
Rogers, J.B.
Further Reading
Romance languages
Compounds, Lexical and Structural
Rosetta
Further Reading
rule-based systems
Rule-Based MT, Feasibility of General , Resources for MT, Empirical Approaches to ,
Example-Based Translation, Statistical MT, Summary, Further Reading
Russian
Popular Conceptions and , A Bit of , Further Analysis: Grammatical , Lexical and
Structural
Sadler, L.
Further Reading, Further Reading, Further Reading, Further Reading
Sadler, V.
Further Reading
Sag, I.
Further Reading, Further Reading, Further Reading

http://privatewww.essex.ac.uk/~doug/book/node100.html (19 de 26) [20/09/2000 18:05:12]

 Sager, C.J.
Further Reading
Sager, J.C.
Further Reading
Samvelian, P.
Further Reading
Schenk, A.
Further Reading
Schubert, K.
Further Reading
script
Real World Knowledge
selectional restriction
Types of Word , Types of Word , Types of Word , Types of Word , Semantics,
Semantics, Semantics, Semantics, Semantics, Pragmatics, Real World Knowledge,
Resources for MT
semantic feature
Semantics, Semantics, Semantics, Semantics, Semantics
semantic net
Real World Knowledge, Real World Knowledge, Real World Knowledge, Real World
Knowledge, Further Reading
semantic primitive
Semantics, Semantics, Semantics, Semantics, Semantics, Semantics
semantic relation
Meaning, Meaning, Meaning, Meaning, Parsing, Parsing, Generation, Generation,
Generation, Generation, Comparative Grammar and , Comparative Grammar and ,
Interlinguas, Types of Word , Semantics, Semantics, Semantics, Semantics, Semantics,
Semantics, Semantics, Semantics, Semantics, Semantics, Semantics, Semantics, Flexible
or Multi-level , Flexible or Multi-level
semantics
Introduction, Introduction, Meaning, Meaning, Meaning, Parsing, Generation, Summary,
Introduction, Introduction, Semantics\indexsemantics, Pragmatics, Real World
Knowledge, Summary, Summary, Further Reading, Flexible or Multi-level , Flexible or
Multi-level , Knowledge-Based MT, Knowledge-Based MT
semantics(lexical)
Introduction, Interlinguas, Semantics, Semantics
semantics(structural)
Semantics, Semantics
SGML

http://privatewww.essex.ac.uk/~doug/book/node100.html (20 de 26) [20/09/2000 18:05:12]

 Basic Ideas, Basic Ideas, Basic Ideas, Basic Ideas, Basic Ideas, Basic Ideas, Basic Ideas,
SGML Markup and , SGML Markup and , Summary, Further Reading, Further Reading,
Resources for MT, Resources for MT
Shapiro, S.C.
Further Reading
Sharp, R.
Further Reading
Shiwen, Y.
Further Reading
Sidner, C.L.
Further Reading
Simplified English
Controlled Languages, Further Reading
Sinclair, J.M.
Further Reading
Slocum, J.
Further Reading, Further Reading, Further Reading, Further Reading, Further Reading
Snell, B.
Further Reading
Somers, H.L
Further Reading
Somers, H.L.
Further Reading, Further Reading, Further Reading, Further Reading, Further Reading,
Further Reading, Further Reading, Further Reading
Sowa, J.F.
Further Reading
SPANAM
Popular Conceptions and , A Bit of
Spanish
Interlinguas, Interlinguas, Interlinguas, Inflection, Inflection, Inflection, Compounds,
Compounds, Compounds, Semantics, Semantics, Semantics
speech act
Ambiguity, Ambiguity, Knowledge-Based MT
spelling checkers
Document Preparation: Authoring
Spencer, A.
Further Reading
statistical-based MT

http://privatewww.essex.ac.uk/~doug/book/node100.html (21 de 26) [20/09/2000 18:05:12]

 Feasibility of General , Empirical Approaches to , Statistical MT, Summary, Further
Reading
Stoll,;tex2html_html_special_mark_quot;C.H.
Further Reading
style
Evaluation of Engine , Intelligibility, Intelligibility, Knowledge-Based MT,
Knowledge-Based MT
subcategorization
Types of Word , Types of Word , Types of Word , Types of Word , Types of Word ,
Types of Word , Types of Word , Types of Word , Derivation, Resources for MT
sublanguage
Introduction, Introduction, Sublanguage MT, Sublanguage MT, Summary, Further
Reading, Introduction
Sumita, E.
Further Reading, Further Reading
support verb
Multiword units: Idioms , Multiword units: Idioms
support;tex2html_html_special_mark_quot;verb
Further Reading
SUSY
A Bit of , A Bit of , Further Reading
Swedish
Flexible or Multi-level , Flexible or Multi-level , Flexible or Multi-level , Flexible or
Multi-level , Further Reading
synonymy
Terminology, Terminology
syntax
Introduction, Introduction, Introduction, Further Analysis: Grammatical , Meaning,
Generation, Summary, Summary, Further Reading, Ambiguity, Ambiguity, Ambiguity,
Ambiguity, Multiword units: Idioms , Multiword units: Idioms , Multiword units: Idioms
, Multiword units: Idioms , Introduction, Pragmatics, Pragmatics, Real World
Knowledge, Sublanguage MT, Sublanguage MT, Sublanguage MT, Sublanguage MT,
Flexible or Multi-level , Flexible or Multi-level , Flexible or Multi-level ,
Knowledge-Based MT
synthesis
Linguistic Knowledge Architectures, Linguistic Knowledge Architectures, Linguistic
Knowledge Architectures, Linguistic Knowledge Architectures, Comparative Grammar
and , Comparative Grammar and , Comparative Grammar and , Comparative Grammar
and , Comparative Grammar and , Comparative Grammar and , Comparative Grammar

http://privatewww.essex.ac.uk/~doug/book/node100.html (22 de 26) [20/09/2000 18:05:13]

 and , Comparative Grammar and , Interlinguas, Interlinguas, Interlinguas, Introduction,
Types of Word , Inflection, Terminology, Lexical and Structural , Multiword units:
Idioms , Multiword units: Idioms , Multiword units: Idioms , Flexible or Multi-level ,
Flexible or Multi-level
SYSTRAN
Popular Conceptions and , A Bit of , Further Reading
tagging
Resources for MT, Resources for MT, Resources for MT
TAUM
A Bit of , A Bit of , Further Reading
TAUM-AVIATION
Further Reading
Teller, V.
Further Reading
Tennant, H.
Further Reading, Further Reading
tense
Types of Word , Types of Word , Types of Word , Inflection, Inflection, Inflection,
Inflection, Inflection, Inflection, Inflection, Inflection, Inflection, Inflection, Inflection,
Lexical and Structural , Lexical and Structural , Multiword units: Idioms , Semantics,
Semantics, Semantics, Summary
termbank
Terminology, Terminology, Terminology, Terminology, Further Reading, Sublanguage
MT, Some Central Issues, Feasibility of General
terminology
Interlinguas, Introduction, Introduction, Types of Word , Terminology, Further Reading
Termium
Terminology
test suite
The Test Suite, The Test Suite, Further Reading
text type
Document Revision, Transformer Architectures, Transformer Architectures, Sublanguage
MT, Sublanguage MT, Some Central Issues, Evaluation of Engine , Introduction
thesauri
Resources for MT, Example-Based Translation, Example-Based Translation,
Example-Based Translation
Thomas, P.
Further Reading
top-down parsing
http://privatewww.essex.ac.uk/~doug/book/node100.html (23 de 26) [20/09/2000 18:05:13]
 Parsing, Parsing, Parsing, Parsing
topic
Lexical and Structural
transfer
Linguistic Knowledge Architectures, Linguistic Knowledge Architectures, Linguistic
Knowledge Architectures, Linguistic Knowledge Architectures, Linguistic Knowledge
Architectures, Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Interlinguas,
Interlinguas, Interlinguas, Interlinguas, Interlinguas, Interlinguas, Interlinguas,
Interlinguas, Interlinguas, LK Engines Summarised, Types of Word , Flexible or
Multi-level , Flexible or Multi-level , Flexible or Multi-level
transfer(complex)
Comparative Grammar and , Comparative Grammar and , Lexical and Structural ,
Lexical and Structural , Further Reading
transfer(lexical)
Linguistic Knowledge Architectures, Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Introduction,
Types of Word , Inflection
transfer(structural)
Linguistic Knowledge Architectures, Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and
transfer-based systems
Introduction, Linguistic Knowledge Architectures, Linguistic Knowledge Architectures,
Linguistic Knowledge Architectures, Linguistic Knowledge Architectures, Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Interlinguas, Interlinguas, Interlinguas,
Interlinguas, Interlinguas, Interlinguas, Interlinguas, Interlinguas, Interlinguas,
Interlinguas, Interlinguas, Interlinguas, Interlinguas, Interlinguas, Interlinguas, Summary,
Further Reading, Further Reading, Introduction, Types of Word , Terminology,

http://privatewww.essex.ac.uk/~doug/book/node100.html (24 de 26) [20/09/2000 18:05:13]

 Multiword units: Idioms , Flexible or Multi-level , Flexible or Multi-level , Flexible or
Multi-level , Example-Based Translation, Statistical MT
transformer systems
Introduction, Introduction, Transformer Architectures, Transformer Architectures,
Transformer Architectures, Transformer Architectures, Transformer Architectures,
Transformer Architectures, Transformer Architectures, Transformer Architectures,
Linguistic Knowledge Architectures, Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Interlinguas, LK Engines Summarised, LK
Engines Summarised, Summary, Further Reading, Introduction, Types of Word ,
Multiword units: Idioms , Evaluation of Engine , The Test Suite
transformer;tex2html_html_special_mark_quot;systems
Multiword units: Idioms
translation tools
Dictionary-Based Translation Support Tools\indextranslation , Dictionary-Based
Translation Support , Document Revision, Document Revision, Further Reading,
Summary
tree representation
Grammars and Constituent , Grammars and Constituent , Grammars and Constituent ,
Grammars and Constituent , Grammars and Constituent , Grammars and Constituent ,
Grammars and Constituent , Further Analysis: Grammatical , Further Analysis:
Grammatical , Further Analysis: Grammatical , Processing, Parsing, Parsing, Parsing,
Parsing, Parsing, Parsing, Generation, Generation, Generation, Generation, Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Comparative
Grammar and , Comparative Grammar and , Comparative Grammar and , Lexical and
Structural , Lexical and Structural , Lexical and Structural , Multiword units: Idioms ,
Multiword units: Idioms , Semantics, Semantics
Tsujii, J.
Further Reading
University of Lancaster
Resources for MT, Resources for MT
University of Texas, Austin
A Bit of
US Atomic Energy Commission
A Bit of
USAF
A Bit of , A Bit of
Van Eynde, F.
Further Reading, Further Reading
van Noord, G.

http://privatewww.essex.ac.uk/~doug/book/node100.html (25 de 26) [20/09/2000 18:05:13]

 Further Reading
van Slype, G.
Further Reading
Vasconcellos, M.
Further Reading
Vasconcellos,;tex2html_html_special_mark_quot;M.
Further Reading
Vauquois,;tex2html_html_special_mark_quot;B.
Further Reading
Vinay, J.P.
Further Reading
Warwick, S.
Further Reading
Weaver, W.
A Bit of , A Bit of
WEIDNER
Popular Conceptions and , A Bit of
Welsh
Further Analysis: Grammatical
Whitelock,;tex2html_html_special_mark_quot;P.
Further Reading
Wilks, Y.
Further Reading
Wilks, Y.A.
Further Reading, Further Reading
Winograd, T.
Further Reading
Xerox Corporation
Controlled Languages
Zholkovsky, A.
Further Reading

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node100.html (26 de 26) [20/09/2000 18:05:13]

About this document ...

Up: MACHINE TRANSLATION Previous: Index

About this document ...

MACHINE TRANSLATION: An Introductory Guide
This document was generated using the LaTeX2HTML translator Version 95 (Thu Jan 19 1995)
Copyright © 1993, 1994, Nikos Drakos, Computer Based Learning Unit, University of Leeds.

The command line arguments were:

latex2html book.tex.
The translation was initiated by Arnold D J on Thu Dec 21 10:52:49 GMT 1995

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/node101.html [20/09/2000 18:05:14]

Footnotes

...done.
These estimates of CEC translation costs are from [Patterson1982].

http://privatewww.essex.ac.uk/~doug/book/footnode.html (1 de 40) [20/09/2000 18:05:33]

Footnotes

...translations.
In fact, one can get perfect translations from one kind of system, but at the cost of
radically restricting what an author can say, so one should perhaps think of such systems
as (multilingual) text creation aids, rather than MT systems. The basic idea is similar to
that of a phrase book, which provides the user with a collection of `canned' phrases to
use. This is fine, provided the canned text contains what the user wants to say.
Fortunately, there are some situations where this is the case.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (2 de 40) [20/09/2000 18:05:33]

Footnotes

...process.
Of course, the sorts of errors one finds in draft translations produced by a human
translator will be rather different from those that one finds in translations produced by
machine.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (3 de 40) [20/09/2000 18:05:33]

Footnotes

...structure).
Of course, some languages have larger vocabularies than others, but this is mainly a
matter of how many things the language is used to talk about (not surprisingly, the
vocabulary which Shakespeare's contemporaries had for discussing high-energy physics
was rather impoverished), but all languages have ways of forming new words, and this
has nothing to do with logical perfection.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (4 de 40) [20/09/2000 18:05:33]

Footnotes

...views,
Weaver described an analogy of individuals in tall closed towers who communicate
(badly) by shouting to each other. However, the towers have a common foundation and
basement. Here communication is easy: ``Thus it may be true that the way to translate ...
is not to attempt the direct route, shouting from tower to tower. Perhaps the way is to
descend, from each language, down to the common base of human communication --- the
real but as yet undiscovered universal language.''

http://privatewww.essex.ac.uk/~doug/book/footnode.html (5 de 40) [20/09/2000 18:05:33]

Footnotes

...text.
Hatim and Mason [Hatim and Mason1990] give a number of very good examples where
translation requires this sort of cultural mediation.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (6 de 40) [20/09/2000 18:05:33]

Footnotes

...top,
For some reason, linguists' trees are always written upside down, with the `root' at the
top, and the leaves (the actual words) at the bottom.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (7 de 40) [20/09/2000 18:05:33]

Footnotes

...result.
In English, SUBJECTs can only be omitted in imperative sentences, for example orders,
such as Clean the printer regularly, and in some embedded sentences, e.g. the boxed part
of It is essential 122#122

http://privatewww.essex.ac.uk/~doug/book/footnode.html (8 de 40) [20/09/2000 18:05:33]

Footnotes

...reference.
We have not specified the time-reference information: see Chapter .

http://privatewww.essex.ac.uk/~doug/book/footnode.html (9 de 40) [20/09/2000 18:05:33]

Footnotes

...better
Another possibility would be to have another rule which put the translated preposition
immediately after the verb object, giving Turn the button back a position.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (10 de 40) [20/09/2000 18:05:33]

Footnotes

...THEME.
The names of these particular Semantic Relations should not be taken too seriously. In
fact, of course, it does not much matter what the relations are called, so long as they are

http://privatewww.essex.ac.uk/~doug/book/footnode.html (11 de 40) [20/09/2000 18:05:33]

Footnotes

the same in the source and target grammars.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (12 de 40) [20/09/2000 18:05:33]

Footnotes

...dictionary.
`Paper' here is intended to convey `intended for human readers', as opposed to
`electronic' meaning `intended for use by computers'. Of course, it is possible for a paper
dictionary to be stored on a computer like any other document, and our use of `paper'
here is not supposed to exclude this. If one were being precise, one should distinguish
`paper' dictionaries, `machine readable' dictionaries (conventional dictionaries which are
stored on, and can therefore be accessed automatically by computer), and `machine
usable dictionaries'.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (13 de 40) [20/09/2000 18:05:33]

Footnotes

...#bilingualentry#3293>).
The form of the monolingual entry is based on that used in the Oxford Advanced
Learner's Dictionary (OALD); the bilingual entry is similar to what one finds in
Collins-Robert English-French dictionary.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (14 de 40) [20/09/2000 18:05:33]

Footnotes

...usage.
One can also get some idea of the cost of dictionary construction from this. Even if one
were able to write four entries an hour, and keep this up for 8 hours a day every working
day, it would still take over three years to construct even a small size dictionary. Of
course, the time it takes to write a dictionary entry is very variable, depending on how
much of the work has already been done by other lexicographers.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (15 de 40) [20/09/2000 18:05:33]

Footnotes

...processes.
In fact, it is arguable that the vocabulary of a language like English, with relatively
productive morphological processes, is infinite, in the sense that there is no longest word
of the language. Even the supposedly longest word antidisestablishmentarianism can be
made longer by adding a prefix such as crypto-, or a suffix such as -ist. The result may
not be pretty, but it is arguably a possible word of English. The point is even clearer
when one considers compound words (see Section .

http://privatewww.essex.ac.uk/~doug/book/footnode.html (16 de 40) [20/09/2000 18:05:33]

Footnotes

...animate.
The restriction applying on the OBJECT of the verb actually concerns the thing which is
buttoned whether that appears as the OBJECT of a active sentence or the SUBJECT of a
passive sentence.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (17 de 40) [20/09/2000 18:05:33]

Footnotes

...form:
In this rule we write +finite for finite=+. We also ignore some issues about
datatypes, in particular, the fact that on the right-hand-side V stands for a string of
characters, while on the lefthand (lexical) side it stands for the value of an attribute,
which is probably an atom, rather than a string.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (18 de 40) [20/09/2000 18:05:33]

Footnotes

...themselves,
More precisely, the rule is that the third person singular form is the base form plus s,
except (i) when the base form ends in s, ch, sh, o, x, z, in which case +es is added (for
example, poach- poaches, push- pushes), and (ii) when the base form ends in y, when ies
is added to the base minus y.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (19 de 40) [20/09/2000 18:05:33]

Footnotes

...items.
Notice, however, that we still cannot expect morphological analysis and lexical lookup to
come up with a single right answer straight away. Apart from anything else, a form like
affects could be a noun rather than a verb. For another thing, just looking at the word

http://privatewww.essex.ac.uk/~doug/book/footnode.html (20 de 40) [20/09/2000 18:05:33]

Footnotes

form in isolation will not tell us which of several readings of a word is involved.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (21 de 40) [20/09/2000 18:05:33]

Footnotes

...happy)
Note that the category of the stem word is important, since there is another prefix un
which combines with verbs to give verbs which mean `perform the reverse action to X'
--- to unbutton is to reverse the effect of buttoning.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (22 de 40) [20/09/2000 18:05:33]

Footnotes

...students):
Where words have been fused together to form a compound , as is prototypically the case
in German , an additional problem presents itself in the analysis of the compound ,
namely to decide exactly which words the compound consists of. The German word
Wachtraum, for example, could have been formed by joining Wach and Traum giving a
composite meaning of day-dream. On the other hand, it could have been formed by
joining Wacht to Raum, in which case the compound would mean guard-room.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (23 de 40) [20/09/2000 18:05:33]

Footnotes

...act
Creative in the sense of `genuine invention which is not governed by rules', rather than
the sense of `creating new things by following rules' --- computers have no problem with
creating new things by following rules, of course.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (24 de 40) [20/09/2000 18:05:33]

Footnotes

...AGENT
This discussion of the Japanese passive is a slight simplification. The construction does
sometimes occur without the adversive sense, but this is usually regarded as a
`europeanism', showing the influence of European languages.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (25 de 40) [20/09/2000 18:05:33]

Footnotes

...anaphor.
This is a simplification, of course. For one thing, it could be used to refer to something
outside the discourse, to some entity which is not mentioned, but pointed at, for example.
For another thing, there are some other potential antencedents, such as the back in
(411#411 a), and it could be that Speaker A is returning to the digression in sentence
(412#412 f). Though the discourse structure can helps to resolve pronoun-antecedent
relations, discovering the discourse structure poses serious problems.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (26 de 40) [20/09/2000 18:05:33]

Footnotes

...superior.
Politeness dictates that giving by the hearer to the speaker is normally giving
`downwards' ( kureru), so this is the verb used to describe requests, and giving by the
speaker to the hearer is normally giving `upwards' ( ageru), so this is the verb used to
describe offers, etc.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (27 de 40) [20/09/2000 18:05:33]

Footnotes

...correct
As noted above, knowledge about selectional restrictions is unusual in being defeasible
in just this way: the restriction that the AGENT of eat is ANIMATE is only a preference,
or default, and can be overridden. This leads some to think that it is not strictly speaking
linguistic knowledge at all. In general, the distinction between linguistic and real world
knowledge is not always very clear.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (28 de 40) [20/09/2000 18:05:33]

Footnotes

...ASCII
ASCII stands for American Standard Code for Information Interchange.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (29 de 40) [20/09/2000 18:05:33]

Footnotes

...checking
For example, suppose one has a printer manual marked up in this way, with special
markup used for the names of printer components wherever they occur. It would be very
easy to extract a list of printer parts automatically, together with surrounding text. This
text might be a useful addition to a parts database. As regards consistency, it would be
easy to check that each section conforms to a required pattern --- e.g. that it contains a list

http://privatewww.essex.ac.uk/~doug/book/footnode.html (30 de 40) [20/09/2000 18:05:33]

Footnotes

of all parts mentioned in the section.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (31 de 40) [20/09/2000 18:05:33]

Footnotes

...translated.
Although most elements of the structure are exactly matched, there may sometimes be
differences. For example, if the document element Paragraph is composed of document
element Sentence(s), it is perhaps unwise to insist that each Sentence in each language is
paired exactly with a single corresponding Sentence in every other language, since
frequently there is a tendency to distribute information across sentences slightly
differently in different languages. However, at least for technical purposes, it is usually
perfectly safe to assume that the languages are paired Paragraph by Paragraph, even
though these units may contain slightly different numbers of sentences for each language.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (32 de 40) [20/09/2000 18:05:33]

Footnotes

...Engines.
`PACE' stands for `Perkins Approved Clear English'.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (33 de 40) [20/09/2000 18:05:33]

Footnotes

...impact).
For an excellent discussion of the range of aspects that a good translation may need to
take into account, see Hatim and Mason [Hatim and Mason1990].

http://privatewww.essex.ac.uk/~doug/book/footnode.html (34 de 40) [20/09/2000 18:05:33]

Footnotes

...chamois
This comes from the section on `Talking to the Tailor' in an English-Italian phrasebook
of the 1920s.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (35 de 40) [20/09/2000 18:05:33]

Footnotes

...Evaluation'
`Declarative' here is to be contrasted with `procedural'. A declarative specification of a
program states what the program should do, without considering the order in which it
must be done. A procedural specification would specify both what is to be done, and
when. Properties like Accuracy and Intelligibility are properties of a system which are
independent of the dynamics of the system, or the way the system operates at all ---
hence `non-procedural', or `declarative'.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (36 de 40) [20/09/2000 18:05:33]

Footnotes

...grammar.
It would be nice to try to find possible problem areas by some sort of automatic scanning
of bilingual texts but the tools and techniques are not available to date.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (37 de 40) [20/09/2000 18:05:33]

Footnotes

...re-entrance.
Here `same value' is to be interpreted strongly, as token identity --- in a sentence with two
nouns, there would be two objects with the `same' category value, namely, the two nouns.
This is often called `type' identity. In everyday usage, when we speak of two people
having the `same' shirt, we normally mean type identity. Token identity would involve
them sharing one piece of clothing. On the other hand, when we speak of people having
the same father, we mean token identity.

http://privatewww.essex.ac.uk/~doug/book/footnode.html (38 de 40) [20/09/2000 18:05:33]

Footnotes

...context.
This may use the measure of Mutual Information, taking into account (roughly) the
amount of mutual context elements share

http://privatewww.essex.ac.uk/~doug/book/footnode.html (39 de 40) [20/09/2000 18:05:33]

Footnotes

Arnold D J
Thu Dec 21 10:52:49 GMT 1995

http://privatewww.essex.ac.uk/~doug/book/footnode.html (40 de 40) [20/09/2000 18:05:33]

Untitled

PostScript versions of `the MT book'

Doug Arnold
September 6, 2000
Here you can access a PostScript version of D.J. Arnold, Lorna Balkan, Siety Meijer, R.Lee
Humphreys and Louisa Sadler Machine Translation: an Introductory Guide, Blackwells-NCC,
London, 1994, ISBN: 1855542-17x.
1. The title page, including the copyright notice.
2. The book itself: printed with two a4pages on each actual a4 sheet.

You can also down load individual chapters (with two a4 pages on each actual a4 sheet):
1. The title page, including the copyright notice.
2. Preface, Contents, Chapter 1: Introduction
3. Chapter 2: Machine Translation in Practice
4. Chapter 3: Representation and Processing
5. Chapter 4: Machine Translation Engines
6. Chapter 5: Dictionaries
7. Chapter 6: Translation Problems
8. Chapter 7: Representation and Processing revisited: Meaning
9. Chapter 8: Input
10. Chapter 9: Evaluating MT Systems
11. Chapter 10: New Directions in MT
12. Useful Addresses and Glossary
13. Bibliography and index

A web browsable (html) version is also available.

Please note, this book remains copyright the authors. You are free to download all or part of it
for person use. If you want to use it for more than this, you need permission: you can get it by
writing to me: [email protected].
[email protected]

http://privatewww.essex.ac.uk/~doug/MTbook/ [20/09/2000 18:06:26]