基于多模板提示调优和知识增强的事件因果关系识别方法

Event Causality Identification via Multi-template Prompt Tuningand Knowledge Enhancement

事件因果关系识别(Event Causality Identification,ECI)是自然语言处理领域的一项重要研究任务,旨在识别文本中事件之间的因果关系。现有方法大都基于微调范式,不能较好发挥预训练语言模型的作用,难以有效捕获隐式因果关系识别的线索。为此,该文提出了一种基于多模板提示调优和知识增强的事件因果关系识别方法。针对ECI任务设计独特的总提示模板,对显式和隐式事件因果关系分别设计不同的种子提示模板,集成训练所有提示模板,形成适应于ECI任务的提示调优方式。通过引入ConceptNet、Oxford Dictionaries等外部知识库,丰富事件的解释性知识和事件之间的关系性知识,将不同的知识融入提示模板,强化隐式因果关系线索。在EventStoryLine和Causal-TimeBank两个广泛使用的数据集上的实验结果表明,该文方法性能优于现有方法。

Event causality identification(ECI)is to aimed at identifying causal relationships between events in text.Most of the existing methods are based on fine-tuning paradigm,which cannot fully exploit the pre-trained language to capture the cues of implicit causality.This paper proposes an event causality identification method based on multi-template cue tuning and knowledge enhancement.A unique total cue template is designed for the ECI task,with different seed cue templates for explicit and implicit event causality.All cue templates are integrated and trained a cue tuning approach adapted to the ECI task.By introducing external knowledge bases such as ConceptNet and Oxford Dictionaries,different knowledge of events and event relations is integrated into the cue templates to strengthen implicit causality cues.Experimental results on EventStoryLine and Causal-TimeBank show that our approach outperforms existing methods.