Research On Chinese Event Extraction Based On Question Answering

Event extraction is a fundamental task in the field of natural language processing,which aims to extract the triggers of events and their arguments.The extracted event information can give powerful help to downstream tasks such as knowledge graph,dialogue and question answering.In recent years,with the rapid development of deep learning,pre-training techniques have emerged as an alternative to traditional static word embedding.With the help of pretraining models,the study of event extraction has started to take a new direction,and using question answering to perform event extraction is a new idea.While new methods have been applied to the event extraction task with success,some problems need to be solved.For example,sentence-level event extraction models have limited prior knowledge,and traditional pipelined event extraction has error cascades.In addition,since question answering models using pre-training techniques are more suitable for low-resource scenarios and the questions are customizable,it is also a worthwhile research topic to use the models to feedback the event extraction task and accelerate the annotation of the corpus.In this regard,this paper provides an in-depth study from the following three perspectives.(1)We propose a trigger extraction framework incorporating task-related discourse information.The framework retrieves discourse information through the questions related to trigger extraction task and uses these information to broaden the vision of the model,so that the acquisition range of model knowledge is not limited to the sentence itself.(2)We propose a bidirectional stacked question answering framework.The framework firstly unify trigger extraction and argument extraction into a consistent processing pattern,i.e.,span detection and classification.secondly,the framework cast the unified processing pattern as a two-round question answering component.Lastly,The framework stacks two components bidirectionally by devising additional restrictive queries,which can make full use of the complementarity between event triggers and arguments,thus reduce the impact of cascading errors.(3)We propose a corpus annotation method based on question answering.The event extraction model based on question answering is suitable for low-resource scenarios and questions can be customized.By using these characteristics,the annotation method greatly simplifies corpus annotation work and speeds up corpus annotation,thus providing positive feedback on event extraction.