BSQA: Bidirectional Stacked Question Answering Architecture for End-to-end Event Extraction

Event extraction (EE) is a fundamental task of natural language processing, which aims to recognize the occurrence of events and their arguments in the texts. Most previous works on EE employ the pipeline framework to firstly extract event triggers, then to extract arguments for the given event triggers. This framework is very simple, but error cascades are inevitable. Besides, in some cases, the extraction of arguments is easier than event triggers. Furthermore, arguments can help in event trigger extraction. In order to reduce error cascades and take advantage of the complementarity between event triggers and arguments, we propose a bidirectional stacked question answering (BSQA) framework for event extraction. Specifically, we first unify event trigger extraction and argument extraction into an independent span detection and classification component. Secondly, for each component, we devise a non-restrictive extraction query, a restrictive extraction query and a set of restrictive classification queries to accomplish the corresponding span detection and classification task. Finally, we stack the two components for event trigger and argument extraction bidirectionally. One direction sequentially extract event triggers and arguments, while the other direction recognizes arguments first, then event triggers. Experimental results on the FewFC corpus show the effectiveness of our proposed approach.