Abstract: Most existing Chinese anaphora resolution models determine whether two mentions are coreferent by a binary classifier. This approach can lead to incorrect decisions as lower precision features often overwhelm the precision features. We propose a modified multi-pass sieve model for Chinese anaphora resolution to adapt to Chinese. We add a new semantic-based sieve to the original model for incorporating word sense information. The Web word sense information is imported to solve resource constraints. Furthermore, we modify the mention detection sieve based on the Chinese characters. The proposed model is evaluated on five different testing methods on the ACE2005 corpus. Results show that the proposed model outperforms two other baseline models by 4% and 9% respectively.

Key words: artificial intelligence, multi-pass sieve model, semantic information, anaphora resolution, natural language processing