Martha Stone Palmer

My primary research is in the representation of semantic information and its use in natural language processing applications. The meaning of a sentence is a central aspect of natural language understanding, yet an elusive one, since there is no accepted methodology for determining meaning. There is not even a consensus on criteria for distinguishing word senses, as can easily be seen by comparing entries for the same word in any two dictionaries. Linguistics offers potentially useful insights into semantic representations, and initially I used Jackendoff’s Lexical Conceptual Structures as a basis for computational lexical semantics [1]. This was implemented in the Pundit/Kernel text processing system at Unisys [2]. These representations proved to be effective for driving reference resolution, temporal analysis and recovery of implicit information, and led to this system being internationally recognized as providing path breaking coverage of semantics and pragmatics [3]. However, this experience also revealed the domain-specific limitations of the approach, and the difficulty of extending hand-crafted lexical entries to new vocabulary. The desire to develop a more robust technology for semantic processing focused my attention on more data driven techniques, with new insights from linguistics. Beth Levin has correlated syntactic alternations with a verb’s semantic content to create classes of verbs with similar syntactic and semantic behavior [4]. The accessibility of syntactic structure gives rise to the potential for using a distributional analysis of text as a methodology for determining semantic components. My students and I have been developing VerbNet, a class based computational English verb lexicon that contains explicit syntactic frames and semantic components for individual verbs [5, 6]. As a validity check on the semantic components we have used them to drive animations of the actions the verbs describe [7]. VerbNet has recently been incorporated in the PARC parser used by PowerSet [8], as the basis for lexical acquisition at Rochester [9], and as a lexical component for discourse analysis at the University of Illinois/Chicago [10].