General introduction to mechine translation

The mechanization of translation has been one of humanity’s oldest dreams. In

the twentieth century it has become a reality, in the form of computer programs

capable of translating a wide variety of texts from one natural language into

another. But, as ever, reality is not perfect. There are no ‘translating machines’

which, at the touch of a few buttons, can take any text in any language and

produce a perfect translation in any other language without human intervention

or assistance. That is an ideal for the distant future, if it is even achievable in

principle, which many doubt.

What has been achieved is the development of programs which can produce

‘raw’ translations of texts in relatively well-defined subject domains, which can

be revised to give good-quality translated texts at an economically viable rate or

which in their unedited state can be read and understood by specialists in the

subject for information purposes. In some cases, with appropriate controls on the

language of the input texts, translations can be produced automatically that are of

higher quality needing little or no revision.

These are solid achievements by what is now traditionally called Machine

Translation (henceforth in this book, MT), but they have often been obscured

and misunderstood. The public perception of MT is distorted by two extreme

positions. On the one hand, there are those who are unconvinced that there is

anything difficult about analysing language, since even young children are able to

learn languages so easily; and who are convinced that anyone who knows a foreign

language must be able to translate with ease. Hence, they are unable to appreciate

the difficulties of the task or how much has been achieved. On the other hand,

there are those who believe that because automatic translation of Shakespeare,

Goethe, Tolstoy and lesser literary authors is not feasible there is no role for any

kind of computer-based translation. They are unable to evaluate the contribution

which less than perfect translation could make either in their own work or in the

general improvement of international communication.

Most translation in the world is not of texts which have high literary and

cultural status. The great majority of professional translators are employed to

satisfy the huge and growing demand for translations of scientific and technical

documents, commercial and business transactions, administrative memoranda, legal

documentation, instruction manuals, agricultural and medical text books, industrial

patents, publicity leaflets, newspaper reports, etc. Some of this work is challenging

and difficult. But much of it is tedious and repetitive, while at the same time

requiring accuracy and consistency. The demand for such translations is increasing

at a rate far beyond the capacity of the translation profession. The assistance of

a computer has clear and immediate attractions. The practical usefulness of an

MT system is determined ultimately by the quality of its output. But what counts

as a ‘good’ translation, whether produced by human or machine, is an extremely

difficult concept to define precisely. Much depends on the particular circumstances

in which it is made and the particular recipient for whom it is intended. Fidelity,

accuracy, intelligibility, appropriate style and register are all criteria which can be

applied, but they remain subjective judgements. What matters in practice, as far as

MT is concerned, is how much has to be changed in order to bring output up to a

standard acceptable to a human translator or reader. With such a slippery concept

as translation, researchers and developers of MT systems can ultimately aspire

only to producing translations which are ‘useful’ in particular situations — which

obliges them to define clear research objectives — or, alternatively, they seek

suitable applications of the ‘translations’ which in fact they are able to produce.

Nevertheless, there remains the higher ideal of equalling the best human

translation. MT is part of a wider sphere of ‘pure research’ in computerbased

natural language processing in Computational Linguistics and Artificial

Intelligence, which explore the basic mechanisms of language and mind by

modelling and simulation in computer programs. Research on MT is closely

related to these efforts, adopting and applying both theoretical perspectives and

operational techniques to translation processes, and in turn offering insights and

solutions from its particular problems. In addition, MT can provide a ‘test-bed’ on

a larger scale for theories and techniques developed by small-scale experiments in

computational linguistics and artificial intelligence.

The major obstacles to translating by computer are, as they have always been,

not computational but linguistic. They are the problems of lexical ambiguity, of

syntactic complexity, of vocabulary differences between languages, of elliptical and

‘ungrammatical’ constructions, of, in brief, extracting the ‘meaning’ of sentences

General introduction

____________and brief history

and texts from analysis of written signs and producing sentences and texts in

another set of linguistic symbols with an equivalent meaning. Consequently, MT

should expect to rely heavily on advances in linguistic research, particularly those

branches exhibiting high degrees of formalization, and indeed it has and will

continue to do so. But MT cannot apply linguistic theories directly: linguists

are concerned with explanations of the underlying ‘mechanisms’ of language

production and comprehension, they concentrate on crucial features and do not

attempt to describe or explain everything. MT systems, by contrast, must deal

with actual texts. They must confront the full range of linguistic phenomena, the

complexities of terminology, misspellings, neologisms, aspects of ‘performance’

which are not always the concern of abstract theoretical linguistics.

In brief, MT is not in itself an independent field of 'pure' research. It takes

from linguistics, computer science, artificial intelligence, translation theory, any

ideas, methods and techniques which may serve the development of improved

systems. It is essentially ‘applied’ research, but a field which nevertheless has

built up a substantial body of techniques and concepts which can, in turn, be

applied in other areas of computer-based language processing.