Event Argument Extraction via a Distance-Sensitive Graph Convolutional Network

Lu Dai, Bang Wang, Wei Xiang, Yijun Mo

Abstract: Event argument extraction (EAE) is a challenging task due to that some event arguments are far away from event triggers. Recently there have been many works using dependency trees to capture long-range dependencies between event triggers and event arguments. However, not all information in the dependency trees are useful for the EAE task. In this paper, we propose a Distance-Sensitive Graph Convolutional Network (DSGCN) for the EAE task. Our model can not only capture long-range dependencies via the graph convolution over the dependency trees, but also keep the information relevant to EAE via our designed distance-sensitive attention. Furthermore, dependency relation type information is utilized to enrich the representation of each word to further distinguish the roles of event arguments playing in an event. Experiment results on the ACE 2005 English dataset show that the proposed model achieves superior performance than the peer state-of-the-art methods.