一种注意力引导知识增强的事件因果关系识别方法  

An Attention-guided Knowledge-driven Event Causality Identification Method

作　　者：徐博[1] 孙晋辰 林鸿飞[1] 宗林林 XU Bo;SUN Jinchen;LIN Hongfei;ZONG Linlin(School of Computer Science and Technology,Dalian University of Technology,Dalian,Liaoning 116023,China;School of Computer Science and Technology,University of Science and Technology of China,Hefei,Anhui 230027,China;School of Software,Dalian University of Technology,Dalian,Liaoning 116620,China)

机构地区：[1]大连理工大学计算机科学与技术学院,辽宁大连116023 [2]中国科学技术大学计算机科学与技术学院,安徽合肥230027 [3]大连理工大学软件学院,辽宁大连116620

出　　处：《中文信息学报》2025年第1期89-100,共12页Journal of Chinese Information Processing

基　　金：辽宁省社会科学规划基金(L22CTQ002)。

摘　　要：事件因果关系识别是自然语言处理领域的重要任务,由于因果关系表达方式多样且以隐式表达为主,现有方法难以准确识别。该文将外部结构化知识融入事件因果关系识别任务,提出一种注意力引导知识增强的事件因果关系识别方法。首先,通过BERT模型对事件对及其上下文进行编码;然后,提出零跳混合匹配方案挖掘事件相关的描述型知识和关系型知识,通过注意力机制对事件的描述型知识序列进行编码,通过稠密图神经网络对事件对的关系型知识进行编码。最后,融合前三个编码模块识别事件因果关系。基于EventStoryLine和Causal-TimeBank数据集的实验结果表明,该文所构建模型的识别效果优于现有模型,在零跳概念匹配、描述性和关系型知识编码等层面均获得了识别性能的提升。Event causality identification is an important task in the field of natural language processing.To address the diversity and implicitness of the event causality’s expressions,this paper integrates external structured knowledge into the task of event causality identification,and proposes an attention-guided knowledge-driven event causality identification model.First,our model encodes event pairs and their context through the BERT model.Second,the corresponding zero-hop concepts are matched for events through a mixed matching scheme and related external knowledge is introduced.This model encodes the descriptive knowledge sequence of events through the attention mechanism,and induces the relational knowledge graph of events through the attention mechanism and then uses the attention guided graph neural network to encode the graph.Finally,our model fuses the encoding results of the above three modules to identify the event causality.The experimental results on two dataset EventStoryLine and Causal-TimeBank show that the proposed model is better than other existing models.

关 键 词：事件抽取 因果识别 知识图谱 注意力机制 自然语言处理 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]