Incorporating Generation Method and Discourse Structure to Event Coreference Resolution

Congcheng Huang, Sheng Xu, Longwang He, Peifeng Li, Qiaoming Zhu

Abstract: Event coreference resolution is an important task of information extraction. Previous work often focused on modeling the sentence structure, ignoring the structure between paragraphs which is also important to event coreference resolution. Moreover, almost all previous work modeled event coreference resolution as a classification task. In this paper, we introduce macro discourse structure to help event coreference resolution through a Relational Graph Convolutional Network (R-GCN), which can take advantage of structure and relations between paragraphs. Moreover, we are the first to introduce an encoder-decoder style generation model to further boost event coreference resolution task. The experimental results on the English KBP2016 and KBP2017 datasets show that our model CGECR (Classification and Generation models for Event Coreference Resolution) outperforms the SOTA baselines.