Mining Information from Event Structure Relation Graph for Event Argument Extraction

Abstract

Event Argument Extraction is a vital subtask of Event Extraction. Despite the achievements in existing methods, they can not fully use the event structure information and the rich semantics of the labels, which can provide richer external knowledge for extracting event arguments. To this end, we propose an efficient and end-to-end event argument extraction model based on the Event Structure and Question Answering (ESQA-EAE): (1) we model a multi-relational graph of event ontologies to get the structure-aware node representations; (2) we encode the questions and event mentions separately to avoid premature fusion of the two features. Experiments on the ACE2005 show that ESQA-EAE surpasses the baseline models, which further show that ESQA-EAE can use the structural information to improve the accuracy of event argument extraction.

Tasks

Reproductions