A minimal transfer conception for Verbmobil

Abstract

In this paper we introduce the transfer conception MinT that is currently being developed for the prototype of the face-to-face translation system verbmobil. The acronym MinT stands for Minimal Transfer. MinT is a semantic-oriented transfer model that is based on some central ideas of the MRS-based approach outlined in [Copestake et al., 1995], and the Shake-and-Bake approach to machine translation sketched in [Whitelock, 1992]. The central idea of minimal transfer is to relate the source and target language semantic descriptions on a maximal abstract level, without falling back into the well-known problems of the Interlingua approach. Minimal transfer results in simultaneously decreasing the number of transfer rules and leaving a maximal set of options for lexicalization and grammaticalization up to the generator. In sum, MinT can be characterized as a semantic-oriented, unification-based and lexicalist transfer model. Its main knowledge base are transfer statements which provide the correspondences between underspecied semantic predicates of the source and target language. Transfer statements comprise both bilingual and monolingual correspondences. Bilingual correspondences, on the one hand, establish the equivalence between sets of semantic predicates of the source and target languages. They are formulated in a strictly declarative way and can be applied bidirectionally. In order to solve translational ambiguities, the roles and instances of a predicate are typed with fine-grained sorts that are supplied by an elaborated sort hierarchy. Monolingual correspondences, on the other hand, provide a solution to divergences in the logical structure of the languages involved. The idea is to allow the transfer component to initiate further compositional processes if this is motivated by the contrastive situation. Thus, the input structure is transformed into a logically equivalent semantic representation that is shared by the target language. This way, all contrastive knowledge is contained in the transfer component, which allows strict modularity of analysis and generation.