Exploiting Data Characteristics for Document-level Event Extraction

Document-level event extraction (DEE) extracts structured information of events from a document. Previous studies focus on improving the model architecture. We propose to exploit data characteristics: 1) we utilize more coreference information to obtain better document-level entity representations; 2) we manually identify core roles of each event type and propose the hybrid extraction to shallow the memory and alleviate error propagation. Experiments on a large dataset demonstrate that our methods significantly improve model performance on both the role-level and record-level metrics. Our code is available at https://github.com/coszeros/CAB.