Machine Translation at the University of Michigan
Andreas Koutsoudas · 1961
This paper is intended to give a general idea of the methodology and aims of the research in machine translation (MT) being carried on at the University of Michigan. Before I begin, however, I think it worthwhile, in view of the nature of this conference, to make a few remarks on the question of hardware for machine translation. The main fact to be borne in mind, it seems to me, is that, because of the infancy of machine translation, it is very hard to say how well existing machines measure up to all its eventual needs, or what modifications would have to be incorporated in a machine designed specifically for translation. It is true that considerable weight is sometimes given to the fact that machine translation will require a computer with very large immediate access storage, but concern with this sort of problem seems premature. After all, storage media of indefinite capacity, such as magnetic tape, do exist, and it is an open question whether or not the slow access time of such media would create difficulties for translation; e.g. a bottleneck at the stage of dictionary lookup. The main task of machine translation research, at present and in the foreseeable future, is the linguistic task. Until we know in detail how to solve the problems of lexical ambiguity, inflection, and syntax (discussed below), for a given pair of languages, we shall be unable to construct a complete computer program for translation. And until such a program can be written, we can only guess at the problems which would be involved in adapting a computer to handle it. At present we are working at Michigan towards a program capable of translating Russian texts in the field of physics into English. We are pursuing two separate approaches, which may or may not turn