Lexical Semantic Resources for Information Processing Karin Kipper Schuler
Monday May 1, 2006, EE/CSci 5-212
This talk will focus on VerbNet, a verb lexicon compatible with WordNet but with explicitly stated syntactic and semantic information, created originally using Levin verb classes. One of VerbNet's primary applications has been as a basis for Parameterized Action Representations (PARs), which are used to animate the actions of virtual human agents in a simulated 3D environment. In order to support the animation of the actions, PARs have to make explicit many details that are often underspecified in the language. This detailed level of representation also provides a suitable pivot representation for generation in other natural languages, i.e., a form of interlingua. To evaluate VerbNet's syntactic coverage it has been mapped to the Proposition Bank. VerbNet syntactic frames account for over 84\% exact matches to the frames found in PropBank. VerbNet provides mappings between its verbs and WordNet senses and between its verbs and FrameNet II frames, and mappings between the syntactic frames and Xtag tree families. All these resources are complementary and can be used as extensions of each other. The original set of classes described by Levin has been refined and extended in many ways through systematic efforts: the coverage experiment against PropBank corpus instances proposed a large set of new syntactic frames and a better treatment of prepositions; new classes from Korhonen and Briscoe's resource were integrated into the lexicon; and new members from the LCS database were added. Taking advantage of VerbNet's class-based approach automatic acquisition methods are being investigated. These approaches help reduce the manual classification and enable easy adaptation of the lexicon to specific tasks and applications.