VERB SEMANTICS AND LEXICAL SELECTION

Abstract

A Predicate with Selection Restrictions (PSR) is the main verb representation scheme in today's rule-based MT systems. They arc the consequences of linguistic theories such as Fillmore's case theory, and Jackendoff's lexical conceptual structure. As Levin has addressed (Levin 1985), the decomposition of verbs is proposed for the purposes of accounting for systematic semantic-syntactic correspondences. This results in a series of problems for MT systems: inflexible verb sense definitions; difficulty in handling metaphor and new usages; imprecise lexical selection and insufficient system coverage. It seems one approach is to apply probability methods and statistical models for some of these problems. However, the question remains has PSR exhausted the potential of the knowledge-based approach? If not, are there any alternatives that can improve the handling of these problems? We suggest an alternative that represents verb semantic knowledge and accounts for not only fine-tuned systematic semantic-syntactic correspondences, but also semantic-interpretation correspondences. A verb is not represented by a predicate or simple primitives, but by a set of semantic components that are sensitive to the syntactic alternations the semantic interpretations. This system TranStar to see how PSR affects the lexical selection. We first ran UNICON with selection restrictions. Experiments show that the accuracy and coverage of the lexical selection are hard to be improved. Because PSR scheme docs not care much about the relation among meanings and the way to guess the meaning with world knowledge, the only way to have an improvement under PSR is to exhaustively list every possible translation pairs. We then built a prototype system UNICON to demonstrate our suggested method. English and Chinese verb senses are defined on conceptual lattices. Experiments show that the accurate rate for lexical selection has been improved 13.8% after the system makes use of the extended selection process. This proves that once the verb senses are defined with the consideration about the relation among meanings and ways of guessing the new meaning, the system performance can be improved.