Underspecified Japanese Semantics in a Machine Translation System

Language, Information and Computation(PACLIC 11)1996, 53-62 Underspecified Japanese Semantics in a Machine Translation System* Bjorn Gambick Christian Lieske Yoshiki Mori Department of Computational Linguistics University of the Saarland Postfach 151150 D-66041 Saarbriicken, Germany e-mail: {gam,lieske,mori}ecoli.uni-sb.de Abstract Semantic representations which are underspecified with respect to, for example, scope, have recently attracted much attention. Most research in this area has focused on treating English language phenomena in a theoretical fashion. Our paper deviates from this twofold: We take Japanese as the language of our in- vestigations and describe how our ideas about underspecified Japanese semantics (e.g., on modality adverbs) have been implemented in a spoken-language machine translation system. 1 Introduction Natural language expressions are inherently ambiguous. Ambiguities can be caused, for example, by the fact that one of the words used does not have a unique meaning, that more than one syntactic structure may be assigned to the expression, that the scopal relations are not clear, etc. Regardless of the cause, ambiguities are problematic for Natural Language Processing, one of the problems being that they decrease processing efficiency: usually all of the possible interpretations have to be assumed to be right until hard facts prove the contrary. Unfortunately, this can oftentimes not be decided on until after a lot of processing already has been done. A way around this dilemma is to have a common representation for all of the possible interpretations of an ambiguous expression, as in (Alshawi et al., 1992; Kameyama, 1995). Recent research (Reyle, 1993; Bos, 1995; Pinkal, 1995; Cooper et al., 1996) has used the term underspecification to describe this idea: One does not use representations (we will assume formulae of a certain logic in what follows) that encode one single concrete interpretation but a set of interpretations. One prominent area where underspecification can be used, namely for leaving possible scopal domains undecided on, can be explicated with the Japanese sentence in (1): Here, the scope of the focus particle `dake' (`only') cannot be determined. Thus, all of the *This work was funded by the German Federal Ministry of Education, Science, Research, and Tech- nology (BMBF) under grant 01 IV 101 R. We are grateful to Johan Bos, Karsten Worm, and Manfred Pinkal for comments and to Julia Heine, Daniela Kurz, and Feiyu Xu for their work on the lexicon. 53 translations given in (2) are viable, the expression is ambiguous and an underspecified representation is called for. (1) Yamada kara kari� to� dake desu Yamada from borrow past only cop.+pres. (2) I only borrowed it from Y amada. I only borrowed it from Y amada. I only borrowed it from Yamada. The main reason for introducing the underspecified representations has been pro- cessing efficiency. For us, however, using underspecified representations has another motivation. In this paper we will discuss underspecification within the semantic formal- ism of the machine translation (MT) system Verbmobil, a system where the transfer from source language expressions into target language expressions takes place at the (compositional) semantic level. Since ambiguities on the side …

Content truncated to 500 words