Matha Palmer
U. Pennsylvania
As our national interests become increasingly global, timely access to information in other languages becomes more and more necessary. This can only be provided efficiently through the use of automated or semi-automated information processing technology. Computational lexical semantics plays a critical role in multilingual information processing, especially machine translation where it is essential for accurate lexical choice. This talk will give examples of different ranges of translation choices, and demonstrate how the use of cross-linguistic semantic components based on lexical generalizations can lead to accurate predictions. It will introduce Levin classes and a refinement of them, Intersective Levin classes, and show how the closely coupled syntactic frames and semantic components they supply provide a methodology for defining regular sense extensions. The implementation of these classes in Tree-Adjoining Grammars will be briefly discussed, with examples of how the adjunction of prepositional phrases or adverbs can extend a verb's event type representation. The regular sense extensions exemplified by the adjunctions also supply concrete criteria for sense distinctions, which provide a basis for VerbNet, and extension of WordNet, a public domain lexical resource. The VerbNet sense distinctions are currently being used for semantic annotation of on-line corpora which provides training material for word sense disambiguation systems and information extraction systems. The recent word sense disambiguation system evaluation framework, SENSEVAL, will be described, as well as the plans for SENSEVAL2.