Login
Computer Science @ Rochester
Thursday, July 11, 2002
11:00 AM
CSB 209
Donna K. Byron
University of Rochester
Resolving Pronominal Reference to Abstract Entities
In natural languages, the same word can have different meanings in different contexts, when spoken by a different person, or said with different intonation. Humans deal quite well with this ambiguity, yet it remains the primary obstacle to building software that approaches human competence in language understanding.

Nowhere is the problem more apparent than in the interpretation of pronouns. Pronouns are tiny empty vessels whose meaning must be filled in by their surroundings. Automated techniques to understand pronouns have traditionally calculated a pronoun's meaning by finding the noun phrase that is its antecedent. Such techniques cannot find a meaning for pronouns with more complex meanings, such as:

I am defending soon. _That_ makes me very happy.

In spoken discourse, around 50% of the pronouns are of this type---being used to re-mention a concept that was not introduced as a noun phrase. This talk will describe a technique to interpret a broad range of these more complex pronouns. The technique has been implemented within the TRIPS spoken dialog system, and results in an approximate doubling of the number of pronouns interpreted correctly by the system.