Brief history of Machine translation

Brief history of MT

The use of mechanical dictionaries to overcome barriers of language was first

suggested in the 17th century. Both Descartes and Leibniz speculated on the

creation of dictionaries based on universal numerical codes. Actual examples

were published in the middle of the century by Cave Beck, Athanasius Kircher

and Johann Becher. The inspiration was the ‘universal language’ movement,

the idea of creating an unambiguous language based on logical principles and

iconic symbols (as the Chinese characters were believed to be), with which all

humanity could communicate without fear of misunderstanding. Most familiar is

the interlingua elaborated by John Wilkins in his ‘Essay towards a Real Character

and a Philosophical Language’ (1668).

In subsequent centuries there were many more proposals for international

languages (with Esperanto as the best known), but few attempts to mechanize

translation until the middle of this century. In 1933 two patents appeared

independently in France and Russia. A French-Armenian, George Artsrouni, had

designed a storage device on paper tape which could be used to find the equivalent

of any word in another language; a prototype was apparently demonstrated in

1937. The proposal by the Russian, Petr Smirnov-Troyanskii, was in retrospect

more significant. He envisaged three stages of mechanical translation: first, an

editor knowing only the source language was to undertake the ‘logical’ analysis

of words into their base forms and syntactic functions; secondly, a machine was

to transform sequences of base forms and functions into equivalent sequences in

the target language; finally, another editor knowing only the target language was

to convert this output into the normal forms of that language. Although his patent

referred only to the machine which would undertake the second stage, Troyanskii

believed that “the process of logical analysis could itself be mechanised”.

Troyanskii was ahead of his time and was unknown outside Russia when,

within a few years of their invention, the possibility of using computers for

translation was first discussed by Warren Weaver of the Rockefeller Foundation

and Andrew D. Booth, a British crystallographer. On his return to Birkbeck College

(London) Booth explored the mechanization of a bilingual dictionary and began

collaboration with Richard H. Richens (Cambridge), who had independently been

using punched cards to produce crude word-for-word translations of scientific

abstracts. However, it was a memorandum from Weaver in July 1949 which
brought the idea of MT to general notice and suggested methods: the use of wartime

cryptography techniques, statistical analysis, Shannon’s information theory,

and exploration of the underlying logic and universal features of language. Within

a few years research had begun at a number of US centres, and in 1951 the first

full-time researcher in MT was appointed: Yehoshua Bar-Hillel at MIT. A year later

he convened the first MT conference, where the outlines of future research were

already becoming clear. There were proposals for dealing with syntax, suggestions

that texts should be written in controlled languages, arguments for the construction

of sublanguage systems, and recognition of the need for human assistance (preand

post-editing) until fully automatic translation could be achieved. For some, the

first requirement was to demonstrate the technical feasibility of MT. Accordingly,

at Georgetown University Leon Dostert collaborated with IBM on a project which

resulted in the first public demonstration of a MT system in January 1954. A

carefully selected sample of Russian sentences was translated into English,

using a very restricted vocabulary of 250 words and just six grammar rules.

Although it had little scientific value, it was sufficiently impressive to stimulate

the large-scale funding of MT research in the United States and to inspire the

initiation of MT projects elsewhere in the world, notably in the Soviet Union.

For the next decade many groups were active: some adopting empirical

trial-and-error approaches, often statistics-based, with immediate working systems

as the goal; others took theoretical approaches, involving fundamental linguistic

research, aiming for long-term solutions. The contrasting methods were usually

described at the time as ‘brute-force’ and ‘perfectionist’ respectively. Examples

of the former were the lexicographic approach at the University of Washington

(Seattle), later continued by IBM in a Russian-English system completed for the

US Air Force, the statistical ‘engineering’ approach at the

and the methods adopted at the Institute of Precision Mechanics in the Soviet

Union, and the National Physical Laboratory in Great Britain. Largest of all was

the group at Georgetown University, whose successful Russian-English system

is now regarded as typical of this ‘first generation’ of MT research. Centres of

theoretical research were at MIT, Harvard University, the University of Texas, the

University of California at Berkeley, at the Institute of Linguistics in Moscow and

the University of Leningrad, at the Cambridge Language Research Unit (CLRU),

and at the universities of Milan and Grenoble. In contrast to the more pragmatically

oriented groups where the ‘direct translation’ approach was the norm, some of the

theoretical projects experimented with early versions of interlingua and transfer

systems (e.g. CLRU and MIT, respectively).

Much of the research of this period was of lasting importance, not only

for MT but also for computational linguistics and artificial intelligence — in

particular, the development of automated dictionaries and of techniques for

syntactic analysis — and many theoretical groups made significant contributions

to linguistic theory. However, the basic objective of building systems capable of

producing good translations was not achieved. Optimism had been high, there

were many predictions of imminent breakthroughs, but disillusionment grew as

the complexity of the linguistic problems became more and more apparent. In

a 1960 review of MT progress, Bar-Hillel criticized the prevailing assumption

that the goal of MT research should be the creation of fully automatic high

history of MT 7

RAND Corporation,Brief

quality translation (FAHQT) systems producing results indistinguishable from

those of human translators. He argued that the ‘semantic barriers’ to MT could

in principle only be overcome by the inclusion of vast amounts of encyclopaedic

knowledge about the ‘real world’. His recommendation was that MT should adopt

less ambitious goals, it should build systems which made cost-effective use of

human-machine interaction.

In 1964 the government sponsors of MT in the United States formed the

Automatic Language Processing Advisory Committee (ALPAC) to examine the

prospects. In its influential 1966 report it concluded that MT was slower, less

accurate and twice as expensive as human translation and stated that “there is no

immediate or predictable prospect of useful Machine Translation”. It saw no need

for further investment in MT research; instead it recommended the development of

machine aids for translators, such as automatic dictionaries, and continued support

of basic research in computational linguistics. The ALPAC report was widely

condemned as narrow, biased and shortsighted — it was certainly wrong to criticize

MT because output had to be post-edited, and it misjudged the economic factors

— but large-scale financial support of current approaches could not continue. Its

influence was profound, bringing a virtual end to MT research in the United States

for over a decade and damaging the public perception of MT for many years

afterwards.

In the following decade MT research took place largely outside the United

States, in Canada and in Western Europe, and virtually ignored by the scientific

community. American activity had concentrated on English translations of Russian

scientific and technical materials. In Canada and Europe the needs were quite

different: the Canadian bicultural policy created a demand for English-French (and

to a less extent French-English) translation beyond the capacity of the market,

and the European Economic Community (as it was then known) was demanding

translations of scientific, technical, administrative and legal documentation from

and into all the Community languages.

A research group was established at Montreal which, though ultimately

unsuccessful in building a large English-French system for translating aircraft

manuals, is now renowned for the creation in 1976 of the archetypal ‘sublanguage’

system Météo (Chapter 12) for translating weather reports for daily public

broadcasting. In 1976 the Commission of the European Communities decided

to install an English-French system called Systran, which had previously been

developed by Peter Toma (once a member of the Georgetown team) for

Russian-English translation for the US Air Force, and had been in operation since

1970 (see Chapter 10). In subsequent years, further systems for French-English,

English-Italian, English-German and other pairs have been developed for the

Commission. In the late 1970s, it was also decided to fund an ambitious research

project to develop a multilingual system for all the Community languages, based

on the latest advances in MT and in computational linguistics. This is the Eurotra

project, which involves research groups in all member states (see Chapter 14).

For its basic design, Eurotra owes much to research at Grenoble and at

Saarbrücken. During the 1960s the French group had built an ‘interlingua’ system

for Russian-French translation (not purely interlingual as lexical transfer was still
bilingual); however, the results were disappointing and in the 1970s it began to

develop the influential transfer-based Ariane system (Chapter 13). The Saarbrücken

group had also been building its multilingual ‘transfer’ system S

late 1960s (Chapter 11). It was now the general consensus in the MT research

community that the best prospects for significant advances lay in the development

of transfer-based systems. The researchers at the Linguistics Research Center

(LRC) at Austin, Texas (one of the few to continue after ALPAC) had come

to similar conclusions after experimenting with an interlingua system and was

now developing its transfer-based M

had begun at Kyoto University on the Mu transfer system for Japanese-English

translation. The Eurotra group adopted the same basic approach, although it

found subsequently that the demands of large-scale multilinguality led to the

incorporation of many interlingual features.

However, during the 1980s the transfer-based design has been joined by new

approaches to the interlingua idea. Most prominent is the research on knowledgebased

systems, notably at Carnegie Mellon University, Pittsburgh (see section 18.1),

which are founded on developments of natural language understanding systems

within the Artificial Intelligence (AI) community. The argument is that MT must go

beyond purely linguistic information (syntax and semantics); translation involves

‘understanding’ the content of texts and must refer to knowledge of the ‘real

world’. Such an approach implies translation via intermediate representations based

on (extra-linguistic) ‘universal’ elements. Essentially non-AI-oriented interlingua

approaches have also appeared in two Dutch projects: the DLT system at Utrecht

based on a modification of Esperanto (Chapter 17) and the Rosetta system at

Phillips (Eindhoven) which is experimenting with Montague semantics as the

basis for an interlingua (Chapter 16)

More recently, yet other alternatives have emerged. For many years, automatic

translation of speech was considered Utopian, but advances in speech recognition

and speech production have encouraged the foundation of projects in Great Britain

(British Telecom) and in Japan (Advanced Telecommunications Research, ATR):

see section 18.6. The sophistication of the statistical techniques developed by

speech research has revived interest in the application of such methods in MT

systems; the principal group at present is at the IBM laboratories at Yorktown

Heights, NY (see section 18.3)

The most significant development of the last decade, however, is the

appearance of commercial MT systems. The American products from A

Weidner and Logos were joined by many Japanese systems from computer

companies (Fujitsu, Hitachi, Mitsubishi, NEC, Oki, Sanyo, Sharp, Toshiba), and

in the later 1980s by Globalink, PC-Translator, Tovna and the M

developed by Siemens from earlier research at Austin, Texas. Many of these

systems, particularly those for microcomputers, are fairly crude in the linguistic

quality of their output but are capable of cost-effective operation in appropriate

circumstances (see Chapter 9). As well as these commercial systems, there

have been a number of in-house systems, e.g. the Spanish and English systems

developed at the Pan-American Health Organization (Washington, DC), and the

systems designed by the Smart Corporation for Citicorp, Ford, and the Canadian

Department of Employment and Immigration. Many of the Systran installations

USY since theETAL system (Chapter 15); and in Japan workLPSystems,ETAL system

are tailor-made for particular organisations (Aérospatiale, Dornier, NATO, General

Motors).

Nearly all these operational systems depend heavily on post-editing to produce

acceptable translations. But pre-editing is also widespread: in some systems, for

instance, operators are required, when inputting text, to mark word boundaries or

even indicate the scope of phrases and clauses. At Xerox, texts for translation by

Systran are composed in a controlled English vocabulary and syntax; and a major

feature of the Smart systems is the pre-translation editor of English input.

The revival of MT research in the 1980s and the emergence of MT systems

in the marketplace have led to growing public awareness of the importance of

translation tools. There may still be many misconceptions about what has been

achieved and what may be possible in the future, but the healthy state of MT

is reflected in the multiplicity of system types and of research designs which

are now being explored, many undreamt of when MT was first proposed in the

1940s. Further advances in computer technology, in Artificial Intelligence and in

theoretical linguistics suggest possible future lines of investigation (see Chapter

18), while different MT user profiles (e.g. the writer who wants to compose a

text in an unknown language) lead to new designs. But the most fundamental

problems of computer-based translation are concerned not with technology but

with language, meaning, understanding, and the social and cultural differences of

human communication.