Abstract: Event coreference resolution is an asignificant subtask of information extraction and plays an import role in information fusion,QA system and reading comprehension.This paper introduced a multi-attention-based CNN neural network,called CorefNet,to resolve document-level event coreference.CorefNet uses a deep CNN to extract event features and a multi-attention mechanism to capture important features.Compared with most previous studies with probability-based or graph-based models,the proposed model only uses a few features.Compared with the current main stream nueral network model,this menthod can extract deep event features,and significantly improve the performance of event coreference resolution.The experimental results on the ACE2005 corpus show that this model achieves the state-of-the-art results

Key words: Event coreference, Document-level, Attention mechanism, Deep CNN