Exploiting event-aware and role-aware with tree pruning for document-level event extraction

Document-level event extraction aims to extract event information from a passage, which is more challenging than sentence-level event extraction. Despite some success, the existing document-level event extraction methods still face the following two shortcomings: (a) Insufficient Dependencies: the dependencies between event type and role type are not fully exploited in a document. (b) Redundancy: The redundant role features in document-level event extraction hinder the performance improvement. In this paper, we propose dual-channel Conditional Interaction with Tree Pruning (CITP) to solve the above two challenges simultaneously. For the insufficient dependencies issue, CITP constructs a dual-channel condition interaction module to simultaneously extract both event types and role types in a document, which improves the feature interaction between event types and role types. Also, it can effectively enhance candidate argument features and sentence features with event-aware and role-aware. For the redundancy issue, CITP expands the ordered tree and adopts the classified role types as guidance to prune unnecessary branches, which reduces the impact of redundant role features and improves event extraction performance. Experimental results on the widely used ChFinAnn dataset have proved that our model achieves state-of-the-art compared to other models, with higher effectiveness.