Abstraction and underspecification in semantic transfer

Abstraction and Underspecification in Semantic Transfer Bernd Abb, Bianka Buschbeck-Wolf, Christel Tschernitschek Institute for Logic and Linguistics IBM Deutschland Informationssysteme GmbH Vangerowstr.18, 69115 Heidelberg, Germany e-mail: {abb,bianka,christel}@heidelbg.ibm.com Abstract This paper introduces the semantic transfer approach MinT (Minimal Transfer) that has been developed in the speech-to-speech MT system VERBMOBIL. As a unification-based and lexicalist semantic transfer model, it relies on some central ideas of the MRS-based transfer approach outlined in [Copestake et al., 1995]. It differs, however, from the latter in certain aspects: in MinT, the idea of abstraction and underspecification is worked out in much more detail and has been applied to a variety of translation phenomena. MinT relates SL and TL semantic descriptions on a maximally abstract level, which results in simultaneously decreasing the number of transfer rules and leaving a considerable amount of options for lexicalization and grammaticalization up to the generator. To preserve ambiguities that hold across the involved languages MinT processes underspecified semantic representations. 1 Introduction In this paper, we present the semantic transfer approach MinT (Minimal Transfer) that has been developed for the face-to-face MT system Verbmobil and is applied to the German-English transfer part.1 Verbmobil is designed to produce English output for spoken German and Japanese input in the domain of appointment scheduling dialogs. The input for the MinT module are UMRS (Underspecified Minimal Recursion Semantics) struc- tures (see section 3) which are the result of the syntactic and semantic analysis. On the basis of these semantic representations (SR), an evaluation component provides information about the dialog context and the speech acts by integrating domain-specific world knowledge. It allows the transfer and other components to access additional information. The transfer provides the genera- tor with target UMRS representations, which are mapped to TL expressions and transformed into speech by the synthesis component. The central idea of MinT ([Abb and Buschbeck-Wolf, 1995]) is to relate underspecified SL and TL semantic descriptions on a level with a maximal degree of abstractness so that analysis and transfer efforts can be reduced significantly. MinT is a semantic, unification-based and lexical- ist 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 MT sketched in [Whitelock, 1992]. 1 This work was partially funded by the German Ministry for Research and Technology (BMBF) in the framework of the Verbmobil Project under grant no. 01IV101V. The responsibility for the contents of this study lies with the authors. 56 But it differs from them in certain aspects: in MinT, the idea of abstraction and underspecification is worked out in much more detail and has been applied to a variety of translational phenomena. MinT involves techniques for the resolution of translational ambiguities and copes with structural- semantic divergences in a systematic way. This paper is organized as follows: in section 2, we briefly discuss the utility of abstraction and underspecification in MT. Section 3 introduces the framework of UMRS with focus on underspecified representations. In section 4, we describe …

Content truncated to 500 words