基于深度学习的中文实体关系抽取方法  

Chinese Entity Relation Extraction Method Based on Deep Learning

作　　者：孙紫阳 顾君忠[1] 杨静[1] SUN Ziyang;GU Junzhong;YANG Jing(Department of Computer Science and Technology,East China Normal University,Shanghai 200062,China)

机构地区：[1]华东师范大学计算机科学技术系,上海200062

出　　处：《计算机工程》2018年第9期164-170,共7页Computer Engineering

基　　金：国家科技支撑计划项目(2015BAH01F02);上海市自然科学基金(17ZR1444900)

摘　　要：实体关系抽取技术通过文本内容确定句子中实体对之间的关系类别,但由于中文的语法结构复杂、词义理解多样等因素,其对中文实体关系的分类效果不佳。为此,提出一种基于最短依存路径表示文本的深度学习方法。利用依存分析对语句良好的表示性,配合词性特征,利用长短期记忆(LSTM)网络单元双向结构学习最短依存路径的表示信息,并对LSTM的输出使用卷积神经网络(CNN)训练分类模型。实验结果表明,该方法能够准确地抽取实体关系,其F1值较CNN和Bi-LSTM方法有所提高。Entity relation extraction technology aims to determine the relationship among entity pairs of statements via the text content.Due to the complexity of the grammatical structure,the variety of word meanings and other factors in Chinese,the relationship classification of Chinese entities is very limited.Aiming at this problem,this paper presents a deep learning method expressing texts based on the Shortest Dependency Path(SDP),which makes full use of the dependency analysis to express the statement and word class.It uses the bidirectional structure with Long Short-Term Memory(LSTM)units while learning the information expressed by the SDP,and then uses the Convolutional Neural Network(CNN)training classification model for LSTM output.Experimental result shows that this method can extract the entity relation accurately,its F1 value is higher than that of CNN method and BI-LSTM method

关 键 词：关系抽取 依存分析 最短依存路径 长短期记忆网络 卷积神经网络 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]