machine translation

machine translation,

in computational linguistics, publishing, and other fields, the use of computers to conduct large-scale translation operations. The electronic translation of one language into another or the electronic syntactic analysis of a text has been attempted since the mid 20th cent. However, the complexities of this type of operation, both practical and theoretical, have resulted in only a limited measure of success.

Machine Translation

(also automatic translation), translation of texts from one language into another using automatic devices. There are two research trends in machine translation: the applied trend (industrial machine translation of scientific and technical texts, automation of information services, and so on) and the theoretical trend (simulation of human speech as a method of studying speech; development of formal mathematical structures for linguistic descriptions; the search for algorithms for processing linguistic objects; and study of the relations between human thought and machines).

A machine translation system usually consists of linguistic descriptions of the source and target languages (automatic vocabularies and formal grammars at all levels) and an algorithm (instructions for using the vocabularies and grammars, oriented only to their form), on the basis of which the translation itself is performed. The complete process of machine translation consists of the following principal stages: (1) analysis of the texts in the source language (search for words in the vocabulary, as well as morphological and syntactical analysis—that is, simulation of comprehension of the text), (2) conversion (transfer from the structure of the text in the source language to the structure of the text in the target language), and (3) synthesis of the text in the target language (syntactical and morphological presentation of the text—that is, simulation of construction of the text). In actual machine translation systems all these stages may be closely interrelated, and some may be absent.

The machine translation algorithm is usually performed by a general-purpose digital computer. The text produced as a result of the machine translation may be edited by a human, the “post-editor,” who corrects mistakes and ambiguities in the translation. Here is an example, in general terms, of translation of the sentence “He was seen at 6 o’clock” from English into Russian according to the stages described above. In the analysis stage it is determined that “he” is the subject, “was seen” is the predicate (the verb “see” in the past indefinite, indicative mood, passive voice), and “at 6 o’clock” is an adverbial modifier of time. In the conversion stage, Russian translations are placed incorrespondence to the English words and word combinations: “he” -» on, “see” -» videt’ “6 o’clock” -> shest’ chasov. Since the verb videt’ is not used in the passive voice, the English passive construction is converted into a Russian impersonal construction: on becomes the direct object (ego) of the predicate (the verb videt’, imperfective aspect, indicative mood, active voice, past tense, plural). In the synthesis state the case and prepositional markers of syntactical connections among words are worked out; in particular, the preposition “at” is translated into “v + accusative case” as an indicator of a time modifier, and on as a direct object is given the marker “accusative case.” The word order is determined, and then the necessary forms of the words are shaped so that the result is Ego videli v 6 chasov. If the initial sentence had contained the English pronoun “it” instead of “he,” ambiguity could have arisen during translation (without considering preceding sentences); Ego videli . . . (if “it” is an airplane [samolet, masculine]), Ee videli . . . (if “it” is a rocket [raketa, feminine]), or Eto videli .. . (if “it” is an event or phenomenon, neuter). In this case the human posteditor can select the correct version.

The problem of machine translation is at the junction of theoretical and applied linguistics (including structural and statistical linguistics), mathematical linguistics, the theory and practice of computer programming and automatic programming, and information science. The automation of linguistic research has developed parallel to machine translation. The methods of machine translation developed for natural languages are used in problems connected with artificial languages (automatic programming languages and information languages).

The idea of machine translation was first expressed by the French inventor J. Arzruni and independently by the Soviet inventor P. P. Smirnov-Troianskii (1933). With the appearance of electronic computers in the 1940’s and 1950’s, work on machine translation began in the USA and the USSR. The first experiment with machine translation of Russian into English was conducted at Georgetown University (Washington, D.C.) in 1954. In 1955-56 the first tests of machine translation (English-Russian and French-Russian) were made in the USSR. Research on machine translation developed subsequently in many other countries. The main source languages are English, Russian, and French; the main target languages are those three and German, Japanese, Czech, and Vietnamese.

The initial period of work on the problems of machine translation (until about 1961) was characterized by increased attention to technical and programming questions, orientation toward specific pairs of languages (“binary translation”), development of only the morphological and syntactical rules of translation, direct formulation of the rules of translation in the form of algorithmic rules, and output of usually only one version of the translation for each sentence. As the development of machine translation continued, the results of modern structural and mathematical linguistics came into extensive use. Primary attention was directed toward the development and refinement of general schemes of machine translation suitable for the most diverse languages. The rules of text processing for specific languages were formulated more as conditions imposed on the correct result of processing than as algorithmic rules. The process of machine translation is carried out by a highly universal algorithm, which identifies and performs all possible procedures for text processing in a given stage that lead to permissible results according to given rules (multiversion processing); in subsequent stages the superfluous and incorrect versions are rejected (the filter method).

The morphological problems of machine translation, and also many syntactical problems, within isolated sentences have been basically solved. The main difficulties in creating a fully automated system for high-quality translation are related to the inadequate level of development of the semantic theory of languages. Such a theory would make possible precise formulation of the rules for processing the sense and meanings of sentences of a language.

In practice, simplified and specialized machine translation systems are used to automate the processing of scientific and technical data (word-for-word translation with partial grammatical processing and automatic abstracting for purposes of express information services and patent documentation, and also for information retrieval systems).

Publications devoted in part or full to machine translations include the journals Mechanical Translation (Cambridge, since 1954), T. A. Informations (Paris, since 1965; from 1960 to 1964 published as Traduction automatique), Communications of the Association for Computing Machinery (Philadelphia, since 1958), and Nauchno-tekhnicheskaia informatsiia: Seriia 2 (Scientific and Technical Information: Series 2; Moscow, since 1961) and the collections of articles Mashinnyiperevod iprikladnaia lingvistika (Machine Translation and Applied Linguistics; Moscow, since 1959) and Problemy kibernetiki (Problems of Cybernetics; Moscow, since 1958).

REFERENCES

Perevodnaia mashina P. P. Troianskogo: Sbornik materialov o perevodnoi mashine dlia perevoda s odnogo iazyka na drugie, predlozhennoi P. P. Troianskim v 1933 g. Moscow, 1959.
Mashinnyi perevod: Sb. st. Moscow, 1957. (Translated from English.)
Panov, D. Iu. Avtomaticheskii perevod. Moscow, 1958.
Revzin, I. I., and V. Iu. Rozentsveig. Osnovy obshchego i mashinnogo perevoda. Moscow, 1964.
Mel’chuk, I. A., and R. D. Ravich. Avtomaticheskii perevod, 1949-1963:
Kritiko-bibliograficheskii spravochnik. Moscow, 1967.
Avtomaticheskiiperevod: Sb. st. Moscow, 1971. (Translated from English, Italian, German, and French.)

I. A. MEL’CHUK and S. IA. FITIALOV