Improving Chinese pronominal anaphora resolution by extensive feature representation and confidence estimation

Abstract

Pronominal anaphora resolution denotes antecedent identification for anaphoric pronouns expressed in discourses. Effective resolution relies on the kinds of features to be concerned and how they are appropriately weighted at antecedent identification. In this paper, a rich feature set including the innovative discourse features are employed so as to resolve those commonly-used Chinese pronouns in modem Chinese written texts. Moreover, a maximum-entropy based model is presented to estimate the confidence for each antecedent candidate. Experimental results show that our method achieves 83.5% success rate which is better than those obtained by rule-based and SVM-based methods.