Login
Computer Science @ Rochester
Tuesday, May 07, 2002
11:00 AM
CSB 209
Ph.D. Thesis Proposal
Scott Stoness
University of Rochester
Incrementality and Feedback Integration in Stochastic Parsing
Humans process language in an incremental fashion which allows the almost immediate incorporation of context and reference at all levels, even in such seemingly low-level tasks as lexical disambiguation and parsing. Despite psycholinguistic evidence from eye-tracking experiments and research into parsing preferences, the bulk of computational approaches to the parsing task have continued to follow a whole-sentence model of information flow. Several researchers have indicated that the incremental incorporation of feedback from higher-level modules should be a productive research direction. Unfortunately, the basic dynamic programming techniques which underly both hidden Markov models and probabilistic context-free grammars are not suitable for incrementally incorporating information. A new architecture is introduced which features a mediation algorithm capable of integrating stochastic judgments from a standard dynamic programming algorithm and a higher level process. The mediation algorithm then surreptitiously changes the chart of the dynamic programming algorithm, violating the independence assumptions which render it unable to interact incrementally, while keeping the efficiency of the initial model relatively intact. A specific instantiation of this architecture in the part-of-speech tagging domain is developed, and the complications and benefits of moving to a parser are discussed.