基于语言学知识增强的自监督式图卷积网络的事件关系抽取方法  

Linguistic Knowledge-Enhanced Self-Supervised Graph Convolutional Network for Event Relation Extraction

作　　者：徐康 余胜男 陈蕾 王传栋 Xu Kang;Yu Shengnan;Chen Lei;Wang Chuandong(School of Computer Science,Software and Network Security,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学计算机学院、软件学院、网络空间安全学院,南京210003

出　　处：《数据分析与知识发现》2023年第5期92-104,共13页Data Analysis and Knowledge Discovery

基　　金：国家自然科学基金项目(项目编号:62202240,61872190);南京邮电大学人才引进项目(项目编号:NY218118,NY219104)的研究成果之一。

摘　　要：【目的】解决事件关系抽取中因缺少大规模高质量的标注数据以及事件关系复杂的语言表达模式而导致的现有方法难以捕获结构化事件知识的问题。【方法】提出一种基于语言学知识增强的自监督式图卷积网络模型,利用预训练BERT模型编码文本特征,将其输入图卷积网络中学习词之间的句法依存关系以增强文本表示,引入多头注意力机制对不同依赖特征加以区分,再利用分段最大池化操作提取结构信息,然后组合多个段的池化结果作为事件对的关系特征,使用关系特征进行自适应聚类生成伪类别标签,并将其作为自监督信息,通过迭代的自监督训练模式优化事件关系特征。【结果】在TACRED和FewRel数据集上进行实验,B3-F1相较于最好的基线模型分别提高了2.1和1.2个百分点。【局限】该模型将句法依存树当作无向图处理,未考虑边的方向和依赖边的标签信息。【结论】本文所提基于语言学知识增强的自监督式图卷积网络模型能有效增强文本的表征效果,为缺少标注数据的事件关系抽取提供了一种自监督学习框架。[Objective]This paper proposes a Linguistic Knowledge-enhanced Self-Supervised Graph Convolutional Network(LKS-GCN)model,aiming to improve the existing method for event relation extraction.[Methods]First,we used the BERT model to encode the input texts,and learned the syntactic relationships between words with graph convolutional network to enhance text representations.Then,we introduced a multihead attention mechanism to distinguish different dependency features and utilized segment-level max pooling operation to extract structural information.Next,the pooled results of multiple segments were combined as the relation features of event pairs.We conducted adaptive clustering based on the relation representation features and generated pseudo-labels as the self-supervision information.Finally,we optimized event relation features through iterative self-supervised training.[Results]We evaluated the new model on TACRED and FewRel datasets,which made the B3-F12.1% and 1.2% higher than the best baseline methods.[Limitations]The model treated the syntactic dependency tree as an undirected graph and did not consider the edges’direction and dependency edges’label information.[Conclusions]The LKS-GCN model could effectively enhance text representation and provide a self-supervised learning framework for event relation extraction with limited labeled data.

关 键 词：事件关系抽取 BERT 自监督模型 图卷积网络 多头注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]