Reporter: Jun Lang
Date: 2005-12-26(Monday) 16:00
Location: Room 618, New Tech Building
Paper Information:
Author: Shane Bergsma
Title: Automatic Acquisition of Gender Information for Anaphora Resolution
Conference: Canadian AI 2005, May 9-11
-- Winner, AI'2005 Best Paper Award
Abstract:
We present a novel approach to learning gender and number information for anaphora resolution. Noun-pronoun pair counts are collected from gender-indicating lexico-syntactic patterns in parsed corpora, and occurrences of noun-pronoun pairs are mined online from the web. Gender probabilities gathered from these templates provide features for machine learning. Both parsed corpus and web-based features allow for accurate prediction of the gender of a given noun phrase. Together they constructively combine for 96% accuracy when estimating gender on a list of noun tokens, better than any of our human participants achieved. We show that using this gender information in simple or knowledge-rich pronoun resolution systems significantly improves performance over traditional gender constraints. Our novel gender strategy would benefit any of the current top-performing coreference resolution systems.
没有评论:
发表评论