Other-Anaphora Resolution in Biomedical Texts with Automatically Mined Patterns

This paper proposes an other-anaphora resolution approach in bio-medical texts. It utilizes automatically mined patterns to discover the semantic relation between an anaphor and a candidate antecedent. The knowledge from lexical patterns is incorporated in a machine learning framework to perform anaphora resolution. The experiments show that machine learning approach combined with the auto-mined knowledge is effective for other-anaphora resolution in the biomedical domain. Our system with auto-mined patterns gives an accuracy of 56.5%., yielding 16.2% improvement against the baseline system without pattern features, and 9% improvement against the system using manually designed patterns.