摘要
目前大多数端到端的关系抽取方法解决了流水线方法的误差累计问题,但是仍然存在两个问题:结构特征表示不足,缺乏句法结构等信息;句子中存在着大量实体重叠关系,未考虑这些实体重叠关系的抽取导致关系三元组召回率低的问题。针对这些问题,提出一种基于图卷积神经网络的实体关系抽取模型。使用依存句法分析对文本构图,然后通过双向GraphSage提取其结构特征,融入句法结构的特征向量在预测关系时有着更好的表现。而对于关系重叠问题,一次性将所有关系抽取较为困难,因此将该任务分为两步:第一步只抽取非重叠关系与单实体重叠关系;第二步中语言模型抽取关系触发词,并与实体词一起构建实体关系图,这一步能够提高实体对重叠关系的抽取能力。
At present,most of the end⁃to⁃end relation extraction methods can get rid of the error accumulation of the pipeline method.However,there are still two problems.One is that the structural features are not well expressed and are lack of syntactic information,the other is that there are a large number of entity overlap relations in sentences,and without considering the extraction of entity overlap relations leads to low recall of the relational triad.To solve the problems,an entity relation extraction model based on graph convolutional neural network is proposed.Dependency parsing analysis is used to compose the text,and then bidirectional GraphSage is used to extract its structural features.The feature vector integrated into the syntactic structure has a better performance in predicting the relation.As for the relationship overlap,it is difficult to extract all relations at one time,so it is divided into two steps to extract the relations.In the first step,only non⁃overlapping relation and single entity overlap(SEO)relation are extracted.In the second step,the language model is used to extract the relation trigger words and construct the entity relation graph together with the entity words.The later step can improve the extraction ability of entity pair overlap(EPO)relation.
作者
刘源
刘胜全
常超义
孙伟智
LIU Yuan;LIU Shengquan;CHANG Chaoyi;SUN Weizhi(College of Information Science and Engineering,Xinjiang University,Urumqi 830000,China;College of Computer Science and Cyber Security(Oxford Brookes College),Chengdu University of Technology,Chengdu 610000,China)
出处
《现代电子技术》
2022年第13期111-117,共7页
Modern Electronics Technique
基金
国家自然科学基金资助项目(61966034)。
关键词
关系抽取
图神经网络
关系触发词
深度学习
实体重叠
依存句法树
多任务学习
预训练语言模型
relation extraction
graph neural network
relational trigger word
deep learning
entity overlap
dependency parsing tree
multi⁃task learning
pre⁃trained language model
