Verbnet: a Broad-Coverage, Comprehensive Verb Lexicon

Despite the proliferation of approaches to lexicon development, the field of natural language processing has yet to develop a clear consensus on guidelines for computational verb lexicons, which has severely limited their utility in information processing applications. James Pustejovsky's Generative Lexicon has concentrated on nouns rather than verbs. WordNet does not provide a comprehensive account of possible syntactic frames and predicate argument structures associated with individual verb senses and ComLex provides syntactic frames but ignores sense distinctions. Dorr's LCS lexicon attempts to address these limitations, but does not provide broad coverage of syntactic frames or different senses or links to actual instances in corpora. In order to address this gap, we created VerbNet, a verb lexicon compatible with Word-Net but with explicitly stated syntactic and semantic information, using Levin verb classes to systematically construct lexical entries. Classes are hierarchically organized to ensure that all their members have common semantic and syntactic properties. Each class in the hierarchy is characterized extensionally by its set of verbs, and intensionally by syntactic frames and semantic predicates and a list of typical verb arguments. 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 FameNet 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 were investigated. Additional verbs derived from Kingsbury's clustering experiments and from Loper's VerbNet-WordNet correlation experiment were integrated into the lexicon. These experiments show that it is possible to semi-automatically supplement and tune VerbNet with novel information from corpus data. These approaches reduce the manual classification and enable easy adaptation of the lexicon to specific tasks and applications.