Malmkjær K (2003). ‘On a pseudo-subversive use ofcorpora in translator training.’ In Zanettin F et al.(eds.). 119–134.

Mauranen A (2000). ‘Strange strings in translated language:a study on corpora.’ In Olohan M (ed.) Interculturalfaultlines. Manchester: St. Jerome. 119–141.

Mauranen A (2002). ‘Where’s cultural adaptation?’ Intra-linea 5.

Mauranen A (2004). ‘Corpora, universals and interference.’In Mauranen A & Kujamaki P (eds.). 65–82.

Mauranen A & Kujamaki P (eds.) (2004). Translationuniversals: do they exist? Amsterdam: Benjamins.

Olohan M (2001). ‘Spelling out the optionals in translation:a corpus study.’ UCREL Technical Papers 13, 423–432.

Olohan M & Baker M (2000). ‘Reporting that in trans-lated English: evidence for subconscious processes ofexplicitation?’ Across Languages and Cultures 1(2),141–158.

Øveras L (1998). ‘In search of the third code: an investigationof norms in literary translation.’ Meta 43(4), 571–588.

Papai V (2004). ‘Explicitation: a universal of translatedtext.’ In Mauranen A & Kujamakin P (eds.). 143–164.

Pearson J (2000). Terms in context. Amsterdam: Benjamins.Pearson J (2003). ‘Using parallel texts in the translator

training environment.’ In Zanettin et al. (eds.). 15–24.Puurtinen T (1998). ‘Syntax, readability and ideology in

children’s literature.’ Meta 43(4), 524–533.Puurtinen T (2003). ‘Genre-specific features of transla-

tionese? Linguistic differences between translated andnon-translated Finnish children’s literature.’ Literaryand Linguistic Computing 18(4), 389–406.

Puurtinen T (2004). ‘Explicitation of clausal relations: acorpus-based analysis of clause connectives in translatedand non-translated Finnish children’s literature.’ InMauranen A & Kujamaki P (eds.). 165–176.

Schmied J (2002). ‘A translation corpus as a resource fortranslators: the case of English and German preposi-tions.’ In Maia B, Haller J & Ulrych M (eds.) Trainingthe language services provider for the new millennium.Porto: Universidade do Porto. 251–269.

Schmied J & Hildegard S (1996). ‘Explicitness as a univer-sal feature of translation.’ In Ljung M (ed.) Corpus-basedstudies in English: papers from ICAME 17. Amsterdam/Atlanta: Rodopi. 21–36.

Scott M (1996). Wordsmith Tools 3.0. Oxford: OxfordUniversity Press.

Teich E (2003). Cross-linguistic variation in system andtext. The Hague: Mouton de Gruyter.

Tirkkonen-Condit S (2004). ‘Unique items – over- or under-represented in translated language?’ In Mauranen A &Kujamaki P (eds.). 177–184.

Tognini-Bonelli E (2001). Corpus linguistics at work.Amsterdam: Benjamins.

Tognini-Bonelli E & Manca E (2002). ‘Welcoming children,pets and guests: a problem of nonequivalence in thelanguages of ‘‘agriturismo’’ and ‘‘farmhouse holidays’’.’Textus 15(2), 317–334.

Toury G (1980). In search of a theory of translation. TelAviv: Tel Aviv University.

Toury G (1995). Descriptive translation studies andbeyond. Amsterdam: Benjamins.

Toury G (2004). ‘Probabilistic explanations in translationstudies: welcome as they are, would they qualify as uni-versals?’ In Mauranen A & Kujamaki P (eds.). 15–32.

Vanderauwera R (1985). Dutch novels translated intoEnglish: the transformation of a ‘minority’ literature.Amsterdam: Rodopi.

Van Halteren H (ed.) (1999). Syntactic Wordclass Tagging.Dordrecht: Kluwer.

Varantola K (2003). ‘Translators and disposable corpora.’In Zanettin F et al. (eds.). 55–70.

Zanettin F (1998). ‘Bilingual comparable corpora and thetraining of translators.’ Meta 43(4), 616–630.

Zanettin F, Bernardini S & Stewart D (eds.) (2003).Corpora in translator education. Manchester: St. Jerome.

Relevant Websites

http://www.wikipedia.org.

Machine Translation: History 375

Machine Translation: History

J Hutchins, Norwich, UK

� 2006 Elsevier Ltd. All rights reserved.

Precursors and Pioneers, 1933–1954

Although we might trace the origins of ideas relatedto machine translation (MT) to 17th-century specula-tions about universal languages and mechanical dic-tionaries, it was not until the 20th century that thefirst practical suggestions could be made, in 1933

with two patents issued in France and Russia toGeorges Artsrouni and Petr Trojanskij, respectively.Artsrouni’s patent was for a general-purpose machinethat could also function as a mechanical multilingualdictionary. Trojanskij’s patent, also basically for amechanical dictionary, went further with detailedproposals for coding and interpreting grammaticalfunctions using ‘universal’ (Esperanto-based) symbolsin a multilingual translation device.

Neither of these precursors was known to AndrewBooth (a British crystallographer) and Warren Weaver

376 Machine Translation: History

(a director at the Rockefeller Foundation) when theymet in 1946 and 1947 and discussed tentative ideasfor using the newly invented computers for translat-ing natural languages. In July 1949, a memorandumby Weaver, which stimulated the start of MT researchin the United States, suggested various methodsbased on his knowledge of cryptography, statistics,information theory, logic, and language universals(Hutchins, 1997).

In June 1952, the first MT conference wasconvened at MIT by Yehoshua Bar-Hillel, who hadbeen appointed to survey the field. It was alreadyclear that full automation of good-quality transla-tion was a virtual impossibility and that human in-tervention either before or after computer processes(known from the beginning as pre- and post-editing,respectively) would be essential. Ideas were putforward for normalizing input texts and for micro-glossaries to reduce ambiguity problems and forsome kind of syntactic structure analysis. Sugges-tions for future activity were proposed; in particu-lar, Leon Dostert from Georgetown Universityargued for a public demonstration of the feasibilityof MT.

Accordingly, he collaborated with IBM on a simpleMT system demonstrated in New York on January 7,1954, with a great deal of media attention. A carefullyselected sample of Russian sentences was translatedinto English using a very restricted vocabulary of 250words and just six grammar rules. Although the sys-tem had little scientific value, its output was suffi-ciently impressive to stimulate large-scale funding inthe United States and to inspire the initiation of MTprojects elsewhere, notably in the USSR.

High Expectations, 1954–1966

When MT research began, there was little help to behad from current linguistics. As a consequence, in the1950s and 1960s, researchers tended to polarize be-tween empirical trial-and-error approaches, oftenusing statistical methods to discover grammaticaland lexical regularities that could be applied com-putationally, and theoretical approaches involvingfundamental linguistic research and indeed the begin-nings of what was later called computational linguis-tics (see Computational Linguistics: History). Thisdecade saw the beginnings of three basic approachesto MT (see Machine Translation: Overview): (1) thedirect translation model, in which programming ruleswere developed for the translation specifically fromone source language (SL) into one particular targetlanguage (TL) with a minimal amount of analysisand syntactic reorganization and in which problemsof homonymy and ambiguity were simplified by

providing single TL equivalents for SL words tocover most senses; (2) the interlingua model (seeMachine Translation: Interlingual Methods), inwhich translation was into and from some kind oflanguage-neutral representation and that involvedcomplex syntactic and semantic analysis; and (3) thetransfer model, in which SL texts were analyzedinto disambiguated SL representations and then con-verted into equivalent TL representations, fromwhich output text was generated.

The direct translation approach was epitomized byresearch by Erwin Reifler (at the University ofWashington, Seattle) and by Gilbert King (at IBM).Large bilingual dictionaries were compiled for aphotoscopic store (a purpose-built memory device),in which lexicographic information was used not onlyfor selecting lexical equivalents but also for solvinggrammatical problems without the use of syntacticanalysis. A Russian-English system was installed forthe U.S. Air Force, producing translations until theearly 1970s – the output was crude and barely intelli-gible, but appeared to satisfy basic information needsof users.

Dictionary development was also the focus ofresearch at a number of other centers. AnthonyOettinger’s group (Harvard University) compiled amassive Russian-English dictionary, both to serveas an aid for translators (a forerunner of the now-common computer-based dictionary aids) and to pro-duce crude word-for-word translations. Sydney Lamb(at the University of California, Berkeley) concen-trated on developing maximally efficient dictionaryroutines and a linguistic theory appropriate for MT –his theory of stratificational grammar.

The system developed at Georgetown University,for many years the largest research group in theUnited States, was also essentially direct translationin approach, although incorporating various levels ofstructure analysis: morphological, syntagmatic(noun-adjective agreement, verb government, etc.),and syntactic (subjects and predicates, clause rela-tionships, etc.). Systems installed by Euratom in1963 and by the U.S. Atomic Energy Commissionin 1964 continued in regular use until the late1970s. A direct translation approach similar to theGeorgetown system was made at the Institute ofPrecision Mechanics (under D.Y. Panov), but withless practical success, primarily because of lack ofadequate computer facilities.

Research on interlinguas was almost wholly theo-retical. The Cambridge Language Research Unit(Margaret Masterman) investigated a prototype inter-lingua to produce crude (almost word-for-word)translations, which would be refined by means of thesemantic networks of a thesaurus (using mathematical

Machine Translation: History 377

lattice theory.) Silvio Ceccato (Milan University) de-veloped an interlingua based on cognitive processes,involving the conceptual analysis of words and theirpossible correlations in texts – a forerunner of theneural networks of later years. Igor A. Mel’cuk (Insti-tute of Linguistics, Moscow) worked on the linguisticfoundations of an interlingua, his stratificational de-pendency (meaning-text) model of language (seeMel’cuk, Igor Aleksandrovic (b.1932)). NikolajAndreev (Leningrad State University) conceived aninterlingua as composed of those features statisticallymost common in a large number of languages.

The transfer approach was epitomized by the re-search at MIT led by Victor Yngve on syntactic analy-sis and production, mainly for German and English.Despite Chomsky’s association with the group for ashort time, transformational grammar (see GenerativeGrammar) had little influence – indeed Chomskyanlinguistics had little impact on any MT group in thisperiod.

Apart from MIT, the most explicit concentrationon syntactic issues was at Harvard (after 1959),where research focused on the predictive syntacticanalyzer (originally developed at the National Bureauof Standards under Ida Rhodes), a system forthe identification of permissible sequences of grammat-ical categories (nouns, verbs, adjectives, etc.) and theprobabilistic prediction of following categories. How-ever, the results were often unsatisfactory, causedprimarily by the enforced selection at every stage ofthe ‘most probable’ prediction. (Nevertheless, animproved version, the Multiple-Path Predictive Analyz-er, led later to William Woods’s familiar AugmentedTransition Network parser.)

Computer facilities were frequently inadequate,and much effort was devoted to improving basichardware (paper tapes, magnetic media, accessspeeds, etc.) and to devising programming tools suit-able for language processing (e.g., COMIT developedat MIT). Some groups were inevitably forced to con-centrate on theoretical issues, particularly in Europeand the Soviet Union. For political and militaryreasons, nearly all U.S. research was for Russian-English translation, and most Soviet research fo-cused on English-Russian systems, although themultilingual policy of the Soviet Union inspired re-search there on a much wider range of languagesthan elsewhere.

By the mid-1960s MT research groups hadbeen established in many countries throughout theworld, including most European countries (Hungary,Czechoslovakia, Bulgaria, Belgium, Germany, France,etc.), China, Mexico, and Japan. Many of these wereshort-lived; an exception was the project that begun in1960 at Grenoble University.

Throughout this period, research on MT becamean umbrella for much contemporary work in struc-tural and formal linguistics (particularly in the SovietUnion), semiotics, logical semantics, mathematicallinguistics, quantitative linguistics, and nearly all ofwhat is now called computational linguistics and lan-guage engineering – and in the Soviet Union withclose ties with cybernetics and information theory(Leon, 1997).

The ALPAC Report and Its Consequences

In the 1950s, optimism was high; developments incomputing and in formal linguistics, particularlyin the area of syntax, seemed to promise greatimprovements in quality; there were many predic-tions of fully automatic systems operating within afew years. However, disillusion grew as the com-plexity of the linguistic problems became more appar-ent and research seemed to face an apparentlyinsuperable semantic barrier. In an influential survey,Bar-Hillel (1960) criticized the prevailing assumptionthat the goal of MT research should be the creation offully automatic high-quality translation (FAHQT)systems producing results indistinguishable fromthose of human translators. He argued that itwas not merely unrealistic, given the current stateof linguistic knowledge and computer systems, butimpossible in principle.

Although the first working systems (from IBM andGeorgetown) were showing that poor-quality transla-tions could be useful in many circumstances, therewas growing disappointment at the slow develop-ment of good-quality MT. In 1964, the governmentsponsors of MT in the United States (mainly militaryand intelligence agencies) asked the National ScienceFoundation to set up the Automatic Language Proces-sing Advisory Committee (ALPAC) to examine thesituation. It concluded that MT was slower, less accu-rate, and twice as expensive as human translation andthat ‘‘there is no immediate or predictable prospect ofuseful machine translation’’ (ALPAC, 1966: 32). Itsaw no need for further investment in MT research;instead, it recommended the development of machineaids for translators, such as automatic dictionaries,and the continued support of basic research incomputational linguistics. Paradoxically, ALPAC re-jected MT because it fell well short of human quality(i.e., it required post-editing) even though humantranslations are invariably revised before publicationand even though the sponsoring bodies were primari-ly interested in information gathering and analysis, inwhich lower quality would be acceptable. The influ-ence of ALPAC was profound, bringing virtually toan end MT research in the United States for over a

378 Machine Translation: History

decade and indirectly bringing to an end much MTresearch elsewhere. As a consequence, MT was to beno longer the leading area of research in computersand natural language.

The Quiet Decade, 1967–1976

Research did not stop completely, however. Even inthe United States, groups continued for a few moreyears at the University of Texas and at Wayne StateUniversity. But there was a change of direction. In theUnited States, activity had concentrated on Englishtranslations of Russian scientific and technical mate-rials. In Canada and Europe, the needs were quitedifferent. The Canadian government’s bicultural pol-icy created a demand for English-French (and to alesser extent French-English) translation beyond thecapacity of the translation profession. The demandwas even greater within the offices of the EuropeanCommunity for translations of documents from andinto all the EC languages.

At Montreal, research began in 1970 on a syntactictransfer system for English-French translation. TheTAUM project (Traduction Automatique de l’Univer-site de Montreal) had two major achievements: first,the Q-system formalism for manipulating linguisticstrings and trees (later developed as the Prologprogramming language) and, second, the Meteo sys-tem designed specifically for translating the restrictedvocabulary and limited syntax (sublanguage) of me-teorological reports, which were publicly broadcastfrom 1976.

Other groups continued with essentially interlinguaapproaches. In the Soviet Union, Mel’cuk continuedhis research on a meaning-text model for applicationin MT. At Grenoble University, Bernard Vauquoisdeveloped a pivot language for translating Russianmathematics and physics texts into French – not,however, a full interlingua because it representedonly the logical properties of syntactic relationships;there were no interlingual representations for lexicalitems, which were translated by a bilingual transfermechanism. A similar model was adopted at the Uni-versity of Texas (under Winfred Lehmann) in itsMETAL system for German and English: Sentenceswere analyzed into ‘normal forms’, that is, interlin-gual semantic propositional dependency structureswith no interlingual lexical elements.

By the mid-1970s, however, the Grenoble andTexas groups recognized major problems with theirinterlingua approaches: the rigidity of the levels ofanalysis (failure at any one stage meant failure toproduce any output at all), the inefficiency of parsers(too many partial analyses that had to be filtered out),and in particular loss of information about surface

forms of SL input that might be used to guide selec-tion of TL forms and construction of acceptable TLsentence structures. To many at the time it seemedthat the less ambitious transfer approach offeredbetter prospects.

The Revival, 1976–1989

In the decade after ALPAC, more MT systems cameinto operational use and attracted public attention.Most significant were the first Systran installations.Developed by Peter Toma, its oldest version is theRussian-English system at the USAF Foreign Technol-ogy Division (Dayton, Ohio) installed in 1970. TheCommission of the European Communities installedits first Systran system in 1976 and now has versionsfor most languages of the European Community(now European Union). Over the years, the original(basically, direct translation) design has been greatlymodified, with increased modularity and greatercompatibility of the analysis and synthesis compo-nents of different versions, permitting cost reductionswhen developing new language pairs. During the1980s and subsequently, Systran was also installedat many major companies (e.g., General Motors ofCanada, Dornier, and Aerospatiale).

From the early 1980s until the mid-1990s, Systran’smain commercial rival was the Logos Corporation.After experience with an English-Vietnamese systemfor translating aircraft manuals during the 1970s,Logos developed a successful German-English systemthat first appeared on the market in 1982; during the1980s, other language pairs were also developed.Also in the 1980s a commercial METAL German-English system appeared, the research at the Univer-sity of Texas University now funded by Siemens ofMunich (Germany). The system was no longer inter-lingua-based but was now essentially a transfer-basedapproach. Other language pairs were later marketedfor Dutch, French, and Spanish, as well as for Englishand German.

Research also revived in the late 1970s and early1980s. In contrast to the focus in the late 1960sand early 1970s on interlingua approaches, this wascharacterized by the widespread adoption of thethree-stage transfer-based approach, predominantlysyntax-oriented and founded on the formalizationof lexical and grammatical rules influenced by thelinguistic theories of the time.

One major exemplar was the system developed atGrenoble (GETA, Groupe d’Etudes pour la Traduc-tion Automatique), the influential Ariane system.Regarded as the paradigm of the second-generationlinguistics-based transfer systems, Ariane influ-enced projects throughout the world in the 1980s.

Machine Translation: History 379

Of particular note were its flexibility and modularityand its algorithms for manipulating tree representa-tions that incorporated many different levels andtypes of representation (dependency, phrase struc-ture, and logical). Similar in conception were themultilingual transfer system at Saarbrucken (calledSUSY), incorporating a heterogeneity of techniques(phrase structure rules, transformational rules, casegrammar and valency frames, dependency grammar,the use of statistical data, etc.), and the Mu systemdeveloped at the University of Kyoto under MakotoNagao, where the most prominent features were theuse of case grammar analysis and dependency treerepresentations. (In 1986, the research prototypewas converted to an operational system for use bythe Japanese Information Center for Science andTechnology for the translation of abstracts.)

One of the best-known projects of the 1980s wasthe Eurotra project of the European Community,intended as an advanced multilingual transfer systemfor translation among all the EC languages. LikeGETA-Ariane and SUSY, the design combined lexical,logico-syntactic, and semantic information in multi-level interfaces at a high degree of abstractness. Nofacilities for human assistance during translation pro-cesses were envisaged, and problems of the lexiconwere also neglected; at the end of the 1980s with nooperational system in prospect, the project ended,having, however, achieved its secondary aim of sti-mulating cross-national research in computationallinguistics.

From the mid-1980s, there was a revival of inter-est in interlingua systems, motivated in part bycontemporary research in artificial intelligence andcognitive linguistics. Two major projects were basedin the Netherlands. The DLT (Distributed Lan-guage Translation) system in Utrecht (under ToonWitkam) was intended as a multilingual interactivesystem operating over computer networks, in whicheach terminal was to be a translating machine fromand into one language only, using a modified formof Esperanto as an interlingua. The project made asignificant effort in the construction of largelexical databases and proposed the use of a BilingualKnowledge Bank from a corpus of (human) trans-lated texts (anticipating later example-basedapproaches). The Rosetta project at Eindhoven(under Jan Landsbergen) explored the use of Monta-gue grammar (see Montague Semantics) – semanticinterlingual representations were derived from theinterpretation of the derivation trees of syntacticrepresentations – and the feasibility of reversiblegrammars for analysis and generation. Reversibilitysubsequently became a feature of many MT projects.

In the latter half of the 1980s, Japan witnessed asubstantial increase in activity. Most of the computercompanies (Fujitsu, Toshiba, Hitachi, etc.) began toinvest large sums into areas that government andindustry saw as fundamental to the coming fifth gen-eration of the information society – this included MT.The research, initially greatly influenced by the Muproject at Kyoto University, showed a wide variety ofapproaches. Although transfer systems predomi-nated, there were also interlingua systems, such asthe PIVOT system at NEC and the Japanese-fundedmultilingual multinational project, with participantsfrom China, Indonesia, Malaysia, and Thailand.There was also considerable commercial activity andmost of the computer companies marketed transla-tion software, mainly for the Japanese-English andEnglish-Japanese markets. Many of these systemswere low-level direct translation or transfer systemslimited to morphological and syntactic analysis andoften with little attempt made to automaticallyresolve lexical ambiguities. Often restricted to specif-ic subject fields (computer science and informa-tion technology were popular choices), they reliedon substantial human assistance at both the prepara-tory (pre-editing) and the revision (post-editing)stages.

During the 1980s, many research projects werealso established in Korea, in Taiwan, in mainlandChina, and in Southeast Asia, particularly in Malay-sia. There was also an increase in activity in the SovietUnion, where from 1976 research was concentratedat the All-Union Center for Translation in Moscow.Systems for English-Russian and German-Russiantranslation were developed based on the direct trans-lation approach, but there was also work under thedirection of Yurij Apres’jan, based on Mel’cuk’smeaning-text model, leading to systems for French-Russian and English-Russian. Apart from this group,however, most activity in the Soviet Union focused onthe production of relatively low-level operational sys-tems, often involving the use of statistical analyses(influenced by the Speech Statistics group underRajmund Piotrovskij at Leningrad State University).

During this period, many researchers believed thatquality improvements would come from languageresearch in the context of artificial intelligence (AI;see Natural Language Understanding, Automatic),such as the investigations by Yorick Wilks on prefer-ence semantics and semantic templates and by RogerSchank on expert systems and knowledge-based textunderstanding. The Japanese investment in artificialintelligence projects also had a major impact and maywell have stimulated the revival of government fund-ing in the United States. A number of projects applied

380 Machine Translation: History

knowledge-based approaches in Japan, in Europe,and particularly in North America. In the UnitedStates, the most important was at Carnegie-MellonUniversity (Pittsburgh) under Jaime Carbonell andSergei Nirenburg, which experimented with a num-ber of knowledge-based MTsystems (Nirenburg et al.,1992).

For syntactic processing, there was a trend towardthe adoption of unification and constraint-basedformalisms (e.g., Lexical-Functional Grammar, Head-Driven Phrase Structure Grammar; Categorial Gram-mar, etc.). Complex multilevel representationsand large sets of transformation and mapping ruleswere replaced by lexicalist approaches: monostratalrepresentations, a restricted set of abstract rules, andconstraints incorporated into specific lexical entries.

As far as practical use was concerned, one methodof improving quality was seen to be the ‘control’ ofvocabulary and syntax. Systems such as Systran,Logos, and METAL were in principle designed forgeneral application, although in practice their dic-tionaries are adapted for particular subject domainsby corporation purchasers. The Xerox Corporationwent further by restricting, or controlling, inputtexts, so that transfer and generation was simplerand there was less need for post-editing. They havebeen followed by many others in subsequent years –indeed, the design of controlled languages is itself anactive area of research (see Controlled Languages).There have been several companies that design con-trolled language systems for specific clients (e.g.,LANT (later Xplanation) and ESTeam). The majorexample since the early 1980s has been the SmartCorporation (New York), with customers suchas Citicorp, Ford, and the Canadian Departmentof Employment and Immigration. Restriction toparticular subject areas does not necessarily meancontrolled input. Special-purpose systems can bedesigned for particular environments, allowing forless complication in lexicons. The main exampleshere are the successful Spanish-English (SPANAM)and English-Spanish (ENGSPAN) systems developedduring the 1970s and 1980s at the Pan AmericanHealth Organization in Washington (under MurielVasconcellos).

Developments since 1989

The dominant framework for MT research until theend of the 1980s was essentially based on linguisticrules for syntactic analysis, for lexical transfer, formorphology, and so forth. Since 1989, however, thisdominance has been broken by the emergence of newmethods and strategies loosely called corpus-basedmethods.

Most dramatic has been the revival of statistics-based approaches – seen as a return to the empiricismof the first decade as opposed to the rationalism of thelater rule-based methods. The two sides, empiricistsand rationalists, were first directly confronted at aconference in Montreal in June 1992 (TMI, 1992).

With the success of their stochastic techniques inspeech recognition, a group at IBM (YorktownHeights, New York) developed a statistical MT(SMT) model, Candide (Brown et al., 1990). Its dis-tinctive feature was that no linguistic rules were ap-plied; statistical methods were virtually the solemeans of analysis and generation. Candide was ap-plied to the corpus of French and English texts in thereports of Canadian parliamentary debates. The firststep was the alignment of SL and TL sentences andwords that are potentially translation equivalents.Translation itself was achieved through a translationmodel (frequency statistics derived from previouslyaligned SL and TL words and phrases) and a languagemodel of TL word sequences. During the 1990s,statistical methods (see Language Processing: Statisti-cal Methods) have been the main focus of many re-search groups, for example, the improvement ofbilingual text and word alignment techniques andstatistical methods for extracting lexical and termi-nological data from bilingual corpora. Some haveconcentrated on the development of purely statis-tics-based systems, statistical MT (e.g., at the univer-sities of Southern California, Aachen, and HongKong), and others have investigated the integrationof statistical and traditional rule-based approaches(e.g., at Carnegie-Mellon University; the NationalTsing-Hua University, Taiwan; and MicrosoftCorporation).

The second major corpus-based approach, benefit-ing likewise from improved rapid access to largedatabanks of text corpora, has been what is knownas the example-based (or memory-based) approach.Although first proposed in 1981 by Makoto Nagao, itwas only toward the end of the 1980s that experi-ments began, initially in some Japanese groups andduring the DLT project. The underlying hypothesis isthat translation often involves the finding (or recal-ling) of analogous examples, that is, how a particularexpression or similar phrase has been translatedbefore. Example-based MT (EBMT) involves thematching of input phrases against equivalent or simi-lar SL phrases in a corpus of parallel bilingual texts(aligned either statistically or by traditional rule-based methods), the extraction of TL equivalentor closely matching phrases (using, e.g., semanticnetworks or lexical frequencies), and – the most diffi-cult process – the ‘re-combination’ of selected TLphrases to produce fluent and grammatical output.

Machine Translation: History 381

Example-based MT research is pursued actively ata number of Japanese and U.S. centers (Carl andWay, 2003).

Although the main innovations since 1990 havebeen corpus-based approaches, rule-based researchhas continued with projects on both transfer andinterlingua systems. Examples of transfer systemsare the PaTrans system for Danish/English transla-tion of patents (based on Eurotra research) andthe LMT system of Michael McCord at IBM. Inter-lingua-based projects are the CATALYST knowledge-based system at Carnegie-Mellon University for theCaterpillar company and the special-purpose systems(DIPLOMAT) developed for military operations,also at Carnegie-Mellon; the ULTRA system atNew Mexico State University (Sergei Nirenburg);the UNITRAN system based on the linguistictheory of Principles and Parameters (see Principlesand Parameters Framework of Generative Grammar)(Bonnie J. Dorr, University of Maryland); the Pan-gloss project, a collaborative project involving theuniversities of Southern California, New MexicoState, and Carnegie-Mellon; and the Universal Net-working Language project at the Institute of Ad-vanced Studies of the United Nations University(Tokyo), involving groups in some 15 countries.

Another new departure for MT research in the1990s has been the growing interest in spoken lan-guage translation, with the challenge of combiningspeech recognition and the interpretation andproduction of conversation and dialog. The firstlong-standing group was established in 1986 at ATRInterpreting Telecommunications Research Labora-tories (Nara, Japan), which has been developing asystem for telephone registrations at internationalconferences and for telephone booking of hotelaccommodation (Kurematsu and Morimoto, 1996).Slightly later came the JANUS project (under AlexWaibel) in a consortium of Carnegie-Mellon Univer-sity, the University of Karlsruhe, and ATR. TheJANUS project has also focused on travel planning,but the system is designed to be readily expandableto other domains (Levin et al., 2000). Both projectscontinue. A third shorter-lived group, by SRI(Cambridge, UK) as part of its Core Language proj-ect, investigated Swedish-English translation viaquasi-logical forms (Rayner et al., 2000); and on alarger scale there was a fourth project, Verbmobil(directed by Wolfgang Wahlster), funded from 1993until 2000 by the German government. Its aim was todevelop a transportable aid for face-to-face English-language commercial negotiations by Germans andJapanese not speaking fluent English. Although thebasic goal was not achieved, efficient methodologies

for dialog and speech translation were developedand the project established top-class research groupsin German universities as a valuable by-product(Wahlster, 2000).

Evaluation of MT systems emerged as a majorfocus of research activity, now recognized as crucialfor progress in MT research itself. Since the early1990s, there have been numerous workshops dedicat-ed specifically to problems of evaluation (oftenattached to MT conferences), and in particular therehas been the series of Language Resources and Evalu-ation Conferences (LREC). Methodologies developedby the Japan Electronic Industry DevelopmentAssociation (JEIDA) and the Expert Advisory Groupon Language Engineering Standards (EAGLES) andfor the evaluation of ARPA supported projects havebeen particularly influential.

The use of MT systems expanded greatly in the1990s, particularly in commercial agencies, govern-ment services, and multinational companies, wheretranslations are produced on a large scale, primarilyof technical documentation. This was and remainsthe major market for the mainframe systems (Systran,Logos, METAL, and ATLAS), now usually on client-server configurations. Already in 1995, it was esti-mated that over 300 million words a year were beingtranslated by such companies.

The most significant practical development forhuman translators has been the appearance in theearly 1990s of the first translation workstations,which combine various machine aids (see Machine-aided Translation, Methods): multilingual word pro-cessing, OCR facilities, terminology managementsoftware, facilities for concordancing (facilities thattranslators had become familiar with in the 1980s ), –and in particular translation memories. The historicalorigins are described in Hutchins (1998).

Although MT systems for personal computersbegan to be marketed in the 1980s (e.g., the systemsfrom Weidner, Globalink, and Toshiba), there hasbeen a great expansion since 1990. The increasingcomputational power and storage capacities of per-sonal computers makes these commercial systems theequal of previous mainframe systems of the 1980sand earlier – and, in many cases, more powerful.However, there has not been a matching improve-ment in translation quality. Nearly all are basedon older transfer-based (or even direct translation)models; few have substantial and well-founded dic-tionaries; and most attempt to function as general-purpose systems, although most vendors do offerspecialist dictionaries. In nearly all cases, systemsare sold in three basic versions: systems for largecorporations (enterprise systems), usually running

382 Machine Translation: History

on client-server configurations; systems intended forindependent translators (professional systems); andsystems for nontranslators (home use).

The Internet has had a major impact since the mid-1990s. First, there has been the appearance of MTsoftware products specifically for translating Webpages and electronic mail messages. Second, begin-ning in the mid-1990s, many MT vendors haveprovided Internet-based online translation servicesfor translation on demand – pioneered by Systranon the French Minitel network during the 1980s,by CompuServe in 1995 based on the Transcend sys-tem, and then shortly afterward by AltaVista(the Babelfish service using Systran). There arenow numerous other online services, some offeringpost-editing by human translators (revisers), at extracost, but in most cases presenting unrevised results.Hence, translation quality is often poor, inevitablygiven the colloquial nature of many source texts, butthese services are undoubtedly filling a significant(and apparently widely acceptable) demand for im-mediate rough translations for information purposes.MT is now reaching a mass market.

Further Reading

The general history of MT is covered by Hutchins(1986), updated by Hutchins (1988, 1993). Basicsources for the early period are Locke andBooth (1955), Edmundson (1961), Booth (1967),Rozencvejg (1974), Henisz-Dostert et al. (1979),Bruderer (1982), and Hutchins (2000). For the1970s and 1980s, there are good descriptions ofthe main systems in Nirenburg (1987), King (1987),and Slocum (1988). For systems developed during the1990s, sources include Dorr et al. (1999); Somers(2003), the journal Machine Translation, the biennialMachine Translation Summit conferences, work-shops and other conferences for MT, conferences forlanguage resources and evaluation, computationallinguistics, and the Machine Translation Archive.

See also: Chomsky, Noam (b. 1928); Computational Lin-

guistics: History; Controlled Languages; Generative

Grammar; Language Processing: Statistical Methods;

Machine Translation: Interlingual Methods; Machine

Translation: Overview; Machine-Aided Translation: Meth-

ods; Mel’cuk, Igor Aleksandrovic (b.1932); Montague Se-

mantics; Natural Language Understanding, Automatic;

Parsing and Grammar Description, Corpus-Based; Prin-

ciples and Parameters Framework of Generative Gram-

mar; Speech Acts and Artificial Intelligence Planning

Theory.

Bibliography

ALPAC (1966). Language and machines: computers intranslation and linguistics. Report by the AutomaticLanguage Processing Advisory Committee, Division ofBehavioral Sciences, National Academy of Sciences,National Research Council. Washington, DC: NationalAcademy of Sciences, National Research Council.

Bar-Hillel Y (1960). ‘The present status of automatictranslation of languages.’ Advances in Computers 1,91–163.

Booth A D (ed.) (1967). Machine translation. Amsterdam:North-Holland.

Brown P F, Cocke J, Della Pietra S A et al. (1990). ‘Astatistical approach to machine translation.’ Computa-tional Linguistics 16(2), 79–85.

Bruderer H E (ed.) (1982). Automatische Sprachuberset-zung. Darmstadt: Wissenschaftliche Buch-Gesellschaft.

Carl M & Way A (eds.) (2003). Recent advances inexample-based machine translation. Dordrecht: KluwerAcademic Publishers.

Dorr B J, Jordan P W & Benoit J W (1999). ‘A survey ofcurrent paradigms in machine translation.’ Advances inComputers 49, 1–68.

Edmundson H P (ed.) (1961). Proceedings of the NationalSymposium on Machine Translation. London: Prentice-Hall.

Henisz-Dostert B, Macdonald R R & Zarechnak M (1979).Machine translation. The Hague: Mouton.

Hutchins W J (1986). Machine translation: past, present,future. Chichester, UK: Ellis Horwood/New York: JohnWiley.

Hutchins W J (1988). ‘Recent developments in machinetranslation: a review of the last five years.’ In MaxwellD et al. (eds.) New directions in machine translation.Dordrecht: Foris. 9–63.

Hutchins W J (1993). ‘Latest developments in machinetranslation technology: beginning a new era in MT re-search.’ In MT Summit IV: international cooperation forglobal communication. Tokyo: AAMT. 11–34.

Hutchins W J (1997). ‘From first conception to first dem-onstration: the nascent years of machine translation,1947–1954. A chronology.’ Machine Translation 12(3),195–252.

Hutchins W J (1998). ‘The origins of the translator’s work-station.’ Machine Translation 13(4), 287–307.

Hutchins W J (ed.) (2000). Early years in machine transla-tion: memoirs and biographies of pioneers. Amsterdam/Philadelphia: John Benjamins.

King M (ed.) (1987). Machine translation today: the state ofthe art. Edinburgh, UK: Edinburgh University Press.

Kurematsu A & Morimoto T (1996). Automatic speechtranslation: fundamental technology for future cross-language communications. Amsterdam: Gordon andBreach.

Leon J (1997). ‘Les premieres machines a traduire(1948–1960) et la filiation cybernetique.’ Bulag 22, 9–33.

Machine Translation: Interlingual Methods 383

Levin L, Lavie A, Woszczina M, Gates D et al. (2000). ‘TheJANUS-III translation system: speech-to-speech translationin multiple domains.’ Machine Translation 15(1–2), 3–25.

Locke W N & Booth A D (eds.) (1955). Machine transla-tion of languages: fourteen essays. Cambridge, MA: MITTechnology Press.

Nirenburg S (ed.) (1987). Machine translation: theoreticaland methodological issues. Cambridge, UK: CambridgeUniversity Press.

Nirenburg S, Carbonell J, Tomita M & Goodman K (1992).Machine translation: a knowledge-based approach. SanMateo, CA: Morgan Kaufmann.

Rayner M, Carter D, Bouillon P et al. (2000). The spokenlanguage translator. Cambridge, UK: Cambridge Univer-sity Press.

Rozencvejg V J (ed.) (1974). Machine translation andapplied linguistics (2 vols). Frankfurt: Athenaion Velag[Also published as: Essays on lexical semantics (2 vols).Stockholm: Skriptor.].

Slocum J (ed.) (1988). Machine translation systems.Cambridge, UK: Cambridge University Press.

Somers H L (2003). ‘Machine translation: latest develop-ments.’ In Mitkov R (ed.) The Oxford handbook ofcomputational linguistics. Oxford: Oxford UniversityPress. 512–528.

Machine Translation: Interlingual MB Dorr, UMIACS, College Park, MD, USAE Hovy, University of Southern California,Los Angelos, CA, USA

L Levin, Carnegie Mellon University,Pittsburgh, PA, USA

� 2006 Elsevier Ltd. All rights reserved.

Introduction

As described in the article on Machine Translation:Overview, machine translation (MT) methodologiesare commonly categorized as direct, transfer, and in-terlingual. The methodologies differ in the depth ofanalysis of the source language and the extent towhich they attempt to reach a language-independentrepresentation of meaning or intent between thesource and target languages. Interlingual MT typicallyinvolves the deepest analysis of the source language.

Figure 1 – the Vauquois triangle (Vauquois, 1968) –illustrates these levels of analysis. Starting with theshallowest level at the bottom, direct transfer is madeat the word level. Moving upward through syntacticand semantic transfer approaches, the translationoccurs on representations of the source sentencestructure and meaning, respectively. Finally, at the

TMI (1992). Quatrieme Colloque international sur lesaspects theoriques et methodologiques de la traductionautomatique. Proceedings of the 4th InternationalConference on Theoretical and Methodological Issues inMachine Translation: Empiricist vs. rationalist methodsin MT. Actes du colloque. Montreal, Canada: CCRIT-CWARC.

Wahlster W (ed.) (2000). Verbmobil: foundations of speech-to-speech translation. Berlin: Springer.

Relevant Websites

http://www.aamt.info – Asia-Pacific Association for Ma-chine Translation, and its conferences.

http://www.aclweb.org – Association for ComputationalLinguistics and its conference and publication archive.

http://www.amtaweb.org – Association for Machine Trans-lation, in the Americas, and its conferences.

http://www.eamt.org – European Association for MachineTranslation, and its conferences.

http://www.elra.info – Conferences on Language Resourcesand Evaluation.

http://www.mt-archive.info – Machine Translation Archive.

ethodsinterlingual level, the notion of transfer is replacedwith a single underlying representation – the interlin-gua – that represents both the source and target textssimultaneously. Moving up the triangle reduces theamount of work required to traverse the gap betweenlanguages, at the cost of increasing the requiredamount of analysis (to convert the source input intoa suitable pretransfer representation) and synthesis(to convert the posttransfer representation into thefinal target surface form). For example, at the baseof the triangle, languages can differ significantly inword order, requiring many permutations to achievea good translation. However, a syntactic dependencystructure expressing the source text may be convertedmore easily into a dependency structure for the targetequivalent because the grammatical relations (sub-ject, object, modifier) may be shared despite wordorder differences. Going further, a semantic represen-tation (interlingua) for the source language may to-tally abstract away from the syntax of the language,so that it can be used as the basis for the targetlanguage sentence without change.

Comparing the effort required to move up anddown the sides of the triangle to the effort to performtransfer, interlingual MT may be more desirable in