SymCoNLL: A Symmetry-Based Approach for Document Coreference Resolution

Ying Mao, Xinran Xie, Lishun Wang, Zheyu Shi, Yong Zhong

Abstract: Coreference resolution aims to identify the referential relationships between all entities in a text. However, traditional rule-based or hand-crafted feature-based methods encounter difficulties with long sentences and complex structures. Although many neural end-to-end coreference resolution models have been proposed, there are still challenges in resolving coreferences among multiple sentences due to weak contextual connections and ambiguity caused by distant mentions. To address these problems, we propose a symmetry-based method called SymCoNLL that studies coreference resolution from two aspects: mention identification and coreference prediction. Specifically, at the local level, our method focuses on mention identification based on semantic, syntactic, and mention type features, and improves coreference prediction by enhancing the similarity between mentions using semantic and mention type features. At the global level, it strengthens the internal connections of mentions in the same cluster using symmetry. We validate the effectiveness of our method in mention identification and coreference prediction through experiments on OntoNotes 5.0 dataset. The results show that our method significantly outperforms the baseline methods.