基于多粒度阅读器和图注意力网络的文档级事件抽取

Document level event extraction based on multi granularityreaders and graph attention networks

文档级事件抽取面临论元分散和多事件两大挑战,已有工作大多采用逐句抽取候选论元的方式,难以建模跨句的上下文信息。为此,提出了一种基于多粒度阅读器和图注意网络的文档级事件抽取模型,采用多粒度阅读器实现多层次语义编码,通过图注意力网络捕获实体对之间的局部和全局关系,构建基于实体对相似度的剪枝完全图作为伪触发器,全面捕捉文档中的事件和论元。在公共数据集ChFinAnn和DuEE-Fin上进行了实验,结果表明提出的方法改善了论元分散问题,提升了模型事件抽取性能。

Document level event extraction faces two major challenges:argument dispersion and multiple events.Most exis-ting work adopts the method of extracting candidate arguments sentence by sentence,which makes it difficult to model contextual information across sentences.Therefore,this paper proposed a document level event extraction model based on multi granularity readers and graph attention networks.It used multi-granularity readers to achieve multi-level semantic encoding,and used the graph attention network to capture local and global relations between entity pairs.It constructed a pruned complete graph based on entity pair similarity as a pseudo trigger to comprehensively capture events and arguments in the document.Experiments conducted on the public datasets of ChFinAnn and DuEE-Fin show that the proposed method improves the problem of argument dispersion and enhances model’s event extraction performance.