针对篇章级事件抽取任务主要存在的以下3个挑战:模型复杂、事件论元分散以及多事件抽取,提出一种基于异质交互图和有序树的篇章级事件抽取方法DTHIGN.支持以解耦实体抽取部分的方式精简模型的参数;构建面向触发词的篇章级异质交互图,更全面地对语义进行建模;支持使用基于统一事件模板的树形事件论元抽取方式获得事件论元结果.实验结果表明,该方法能够获得较好的效果并显著降低模型的训练难度.
To solve three main challenges in the task of document-level event extraction(DEE),namely model complexity,scat-tered event arguments,and multi-events extraction,a DEE method DTHIGN based on heterogeneous interaction graphs and or-der tree was proposed.This method supported reducing parameters of itself by decoupling entity extraction part.Document-level trigger-oriented heterogeneous interaction graph was constructed to model semantics more comprehensively.The tree-based event argument extraction method with unified event template was used to obtain event arguments results.Experimental results show that the proposed method achieves better performance and significantly reduces the training complexity.