Improving cross-document event coreference resolution by discourse coherence and structure

Abstract

Cross-Document Event Coreference Resolution (CD-ECR) is to identify and cluster together event mentions that occur across multiple documents. Existing methods exhibit two limitations: (1) In contrast to within-document event mentions, which are linked by rich, coherent contexts, cross-document event mentions lack such contexts, posing a challenging for the model to understand the relation between two event mentions in different documents. (2) The lack of coherent textual information between cross-document event mentions lead to the inability to capture their global information, which is important to mine long-distance interactions between them. To tackle these issues, we propose a novel discourse coherence enhancement mechanism and introduce discourse structure to improve cross-document event coreference resolution. Specifically, we first introduce a new task: Event-oriented cross-document coherence enhancement (ECD-CoE), which selects coherent sentences that form a coherent text for two cross-document event mentions. Second, we represent the coherent text as a tree structure with rhetorical relation information between textual units. We then obtain the global interaction information of event mentions from the tree structures and finally resolve coreferent events. Experimental results on both the ECB+ and GVC datasets indicate that our proposed method outperforms several state-of-the-art baselines.