双路注意力引导图卷积网络的关系抽取  

Dual Attention Guided Graph Convolutional Networks for Relation Extraction

作　　者：李志欣[1] 孙亚茹 唐素勤[1] 张灿龙[1] 马慧芳 LI Zhi-xin;SUN Ya-ru;TANG Su-qin;ZHANG Can-long;MA Hui-fang(Guangxi Key Laboratory of Multi-source Information Mining and Security,Guangxi Normal University,Guilin,Guangxi 541004,China;College of Computer Science and Engineering,Northwest Normal University,Lanzhou,Gansu 730070,China)

机构地区：[1]广西师范大学广西多源信息挖掘与安全重点实验室,广西桂林541004 [2]西北师范大学计算机科学与工程学院,甘肃兰州730070

出　　处：《电子学报》2021年第2期315-323,共9页Acta Electronica Sinica

基　　金：国家自然科学基金(No.61966004,No.61663004,No.61967002,No.61866004,No.61762078);广西自然科学基金(No.2019GXNSFDA245018,No.2018GXNSFDA281009,No.2017GXNSFAA198365)。

摘　　要：为了更好地学习节点依赖并利用结构信息,本文提出一种以完全依赖树作为直接输入的新方法,利用图卷积网络并结合两个并行的注意力模块,自主学习如何有选择地关注对关系抽取任务有用的信息.该方法将样本表示成图上的各节点,一个模块用于计算节点特征位置之间的影响,使特征向量可以包含更广范围的语义信息,另一个用于计算节点依赖的关系特征,以增强节点间的全局依赖.两个模块并行相互提升,可以得到完整的特征表示.在TACRED和SemEval数据集上的实验结果表明,该方法能够更有效地获取对关系抽取任务有益的信息,在各评价指标上取得了更好的性能.To better learn node dependence and make use of structural information,this paper proposes a new method that takes the tree of complete dependence as the direct input.The method uses the graph convolutional network and combines two parallel attention modules to learn how to select the useful information.The method represents the samples as nodes on the graph.One module is used to compute the influence between positions of node features,which allows the feature vector to contain a wider range of semantic information.The other one is used to compute the relational features of node dependence,so as to enhance the global dependence between nodes.The two modules promote each other in parallel to obtain complete feature representation.The experimental results on the TACRED and SemEval datasets show that the method can obtain more useful information for relation extraction,thus achieve better performances on various evaluation metrics.

关 键 词：关系抽取 图卷积网络 注意力机制 多跳关系推理 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]