Experiences on the Improvement of Logic-Based Anaphora Resolution in English Texts

Abstract

Anaphora resolution is a crucial task for information extraction. Syntax-based approaches are based on the syntactic structure of sentences. Knowledge-poor approaches aim at avoiding the need for further external resources or knowledge to carry out their task. This paper proposes a knowledge-poor, syntax-based approach to anaphora resolution in English texts. Our approach improves the traditional algorithm that is considered the standard baseline for comparison in the literature. Its most relevant contributions are in its ability to handle differently different kinds of anaphoras, and to disambiguate alternate associations using gender recognition of proper nouns. The former is obtained by refining the rules in the baseline algorithm, while the latter is obtained using a machine learning approach. Experimental results on a standard benchmark dataset used in the literature show that our approach can significantly improve the performance over the standard baseline algorithm used in the literature, and compares well also to the state-of-the-art algorithm that thoroughly exploits external knowledge. It is also efficient. Thus, we propose to use our algorithm as the new baseline in the literature.