基于篇章图模型的中文事件时序关系识别

Chinese Event Temporal Relation Identification via Document-Level Graph

事件时序关系识别是信息抽取中一项十分重要且极具挑战性的任务。此前大部分工作都集中在句子级别上,忽略了篇章级别(同句、邻句和跨句)的事件时序关系识别。针对此问题,该文提出了一种基于篇章图模型的篇章级事件时序关系识别方法,该方法通过分别构造句法层次和篇章级别的图卷积神经网络将篇章中融合了句法特征的所有事件进行信息交互,以达到丰富特征的目的,进而使得模型可更好地识别篇章级时序关系。该文提出的方法在ACE2005-extended中文事件时序关系语料库上的微平均F1值达到71.81%,比最好的基准系统提升了1.76个百分点。

Event temporal relation identification is a challenging subtask of information extraction.Most previous works are focused on identifying sentence-level temporal relation,failing to address document-level relation(i.e.,intra-sentence,adjacent-sentence and nonadjacent-sentence relation).To address this issue,we propose a model of event temporal relation identification on document-level graph.It constructs two Graph Convolutional Networks to encode syntactic information and event interaction information,respectively.The experimental results on the Chinese ACE2005-extended dataset show that the proposed model achieves 71.81%in micro-F 1 measure,with 1.76%improvement compared with the best baseline.