Combining rationalist and empiricist approaches to machine translation
Ralph Grishman, Michiko Kosaka · 1992
Better methods are needed for acquiring the knowledge which must go into machine translation systems. The call for papers for this conference contrast two approaches: the rationalist (based on linguistic theory) and the empiricist (based on analysis of large corpora). We suggest in this paper an intermediate approach which draws on the strengths of both. In this approach, parallel corpora in the source and target languages would be analyzed to produce parses and syntactically regularized tree structures. The individual source and target language trees would then be aligned, yielding a set of correspondences between source and target structures involving specific words. Classes of closely related words would be identified from a distributional analysis of the parsed corpora, and these classes would be used in turn to generalize the correspondences. These generalized correspondences would then serve as the transfer rules of a machine translation system.