Automatic translation

machine translation

Translation of texts from one natural language to another using automatic devices.

Automatic translation is a problem in the area of modelling and automation of human mental activities — in the present case, of language. It is performed by an algorithm, the rules of which do not interact with the knowledge and intuition of man, i.e. are strictly formal. An algorithm for automatic translation can be implemented for any suitable automatic device, e.g. by a digital computer.

Automatic translation is closely connected with the development of modern structural and mathematical linguistics, including difficult and fundamental linguistic problems, many of which were neglected in the past or were not even explicitly stated.

The theoretical basis of the development of automatic translation is the theory of formal grammars (cf. Grammar, formal). An automatic translation algorithm realizes a certain correspondence between two languages which are defined by their grammars — a so-called translation correspondence. It is formulated in terms of structural descriptions of the properties of these languages within the framework of their grammars. The most frequently used structural descriptions are constituent structures and dependency trees (cf. Syntactic structure). An automatic translation algorithm accordingly consists of three main parts: 1) a parsing of the text in the source language, i.e. an analysis of the structure of the input text, based on the given grammar of the input language; 2) a transfer, i.e. a transformation from the structure of the text in the source language to the structure of the text in the target language, based on the given translation correspondence; and 3) a synthesis of the text in the target language, i.e. a transition from the structure of the output text to the specific word sequence.

During the early period of studying automatic translation the transition rules were usually formulated directly, in the form of algorithmic rules, without a preliminary construction of the formal grammars of the source and target languages, and without an explicit formulation of the translation correspondence. As a result, purely linguistic information became confused with the mathematical formulation of the algorithm. In a later period, automatic translation algorithms were constructed on the base of general schemes applicable to a large number of languages, with the maximum possible particularization of the translation stages; the synthetic and the analytical stages were made mutually independent, with an itemization of all possible varieties of interpretation of the text, permissible grammars and translation correspondences (so-called multi-variant translation). Information about the specific language was in most cases rigidly separated from the algorithmic part. In this way it became possible to disregard a large number of details specific for each language, and to concentrate on the development of general procedures for obtaining solutions valid under the specified conditions.

As regards the purely linguistic aspects of automatic translation, all morphological and almost all syntactic problems in a given sentence may be considered as solved. The principal difficulties in the creation of completely automatic systems of high-quality translation are due to the relatively backward state-of-the-art of the semantic theory of languages, which could be used for an exact formulation of the rules of processing the sense and the various meanings, and also for understanding the logical connection between the separate sentences of a coherent text. Unless due allowance is made for the semantics of the text, the translation often proves ambiguous or faulty. A faulty translation is usually due to the fact that the translation rules do not include all possible meanings of the text to be translated or do not invariably ensure a proper selection of the meaning, which may depend on a fairly wide context or even on the knowledge of the subject matter of the text being translated, on general knowledge, etc.

Operational industrial systems of high-quality automatic translations are as yet (1970s) non-existent. There exist experimental systems of varying degrees of complexity. Simplified and specialized systems of automatic translation are successfully used in practical work; these are based on automatic processing of scientific and technical information (word-for-word translation with partial grammatical processing, partial translation and automatic referencing for on-the-spot information, patent documentation, and for information search systems). Automatic translation can serve in the foreseeable future for translating scientific and technical publications only; automatic translation of literature and fiction is both unrealistic and unnecessary.

References

[1]	D.Yu. Panov, "Automatic translation" , Moscow (1958) (In Russian)

Comments

Machine translation started in the USA around 1955.

From a linguistic point of view the programs were naive (in principle they gave word-by-word translations), and they were designed in an unstructured way (e.g. linguistic information was not separated from the translation algorithm). The programs were not satisfactory: they were slower, less accurate and more costly than translations provided by human translators. Therefore it was decided in the USA in 1966 to stop all support by the government for machine translation projects. In the years 1966–1975 hardly any fundamental research concerning machine translation was done. The programs from the first period were developed further in commercial surroundings, and it are mainly these programs that are used practically nowadays. Their use is not in the quality of the translations, but in the saving of time they give to professional translators when used as auxiliary tool (e.g. as text editor). There is up till now only one fully automatic translation system: the Canadian system TAUM that translates weather forecasts from English into French. Since 1975 there is a revival of research in machine translation, especially in western Europe and Japan. Important factors in this research are developments in theoretical linguistics, hardware developments which change the factors of costs and speed, and the increased need for specialized translations (especially for technical manuals).

References

[a1]	W.J. Hutchins, "Machine translation: past, present, future" , E. Horwood (1986)
[a2]	J. Slocum, "A survey of machine translation: its history, current status, and future prospects" Computational Linguistics , 11 (1985) pp. 1–17