Some Linguistic Problems in Machine Translation

T HE GEORGETOWN-IBM EXPERIment in machine translation, in which I participated, raised a number of problems of linguistic method and threw light on some important phases of general linguistic theory. The purpose of machine translation (MT) is to have a logical machine perform a task which so far has been performed by skilled human beings only that of translation, that is, “the transference of meaning from one patterned set of symbols occurring in a given culture . . . into another set of patterned symbols occurring in another culture .. . .” Two questions have to be answered before MT can seriously be attempted: (1) what are the discrete steps involved in the process of translation; (2) how can these steps be stated in terms of the modus operandi of logical machines. The steps in translation can be discovered by a detailed reconstruction of the translation process from a comparison of the original text with its translation, that is, by translation analysis. The results of the translation analysis, in order to be compatible with the modus operandi of logical machines, must be stated explicitly and unequivocally, that is, with each logical step including the obvious ones spelled out in detail, and in terms of the yes-no decisions required by the binary operation of electronic circuits. The above two aspects of MT research the translation analysis, and the verbal statement of its results in acceptable terms – are the proper concern of the linguist, since they must be based upon