Author Login Editor-in-Chief Peer Review Editor Work Office Work

Computer Engineering ›› 2021, Vol. 47 ›› Issue (12): 103-111. doi: 10.19678/j.issn.1000-3428.0059574

• Artificial Intelligence and Pattern Recognition • Previous Articles     Next Articles

Joint Extraction of Chinese Entity Relations Based on Graph Convolutional Neural Network

ZHANG Junlian1,2,3, ZHANG Yifan1,2,3, WANG Mingquan1,3, HUANG Yongjian1,2,3   

  1. 1. Shanghai Carbon Data Research Center, Shanghai Advanced Research Institute, Chinese Academy of Sciences, Shanghai 201210, China;
    2. University of Chinese Academy of Sciences, Beijing 100049, China;
    3. Key Laboratory of Low-Carbon Conversion Science and Engineering, Chinese Academy of Sciences, Shanghai 201210, China
  • Received:2020-09-25 Revised:2020-12-10 Published:2020-12-21

基于图卷积神经网络的中文实体关系联合抽取

张军莲1,2,3, 张一帆1,2,3, 汪鸣泉1,3, 黄永健1,2,3   

  1. 1. 中国科学院上海高等研究院 碳数据与碳评估研究中心, 上海 201210;
    2. 中国科学院大学, 北京 100049;
    3. 中国科学院低碳转化科学与工程重点实验室, 上海 201210
  • 作者简介:张军莲(1996-),女,硕士研究生,主研方向为自然语言处理;张一帆,硕士研究生;汪鸣泉,副研究员、博士;黄永健(通信作者),工程师、硕士。
  • 基金资助:
    国家自然科学基金面上项目(51778601);中国科学院青年创新促进会基金(2018327)。

Abstract: The existing methods for extracting entity relations usually ignore the complex structural features of Chinese sentences.To address the problem, a Graph Convolutional neural Network(GCN)-based method is proposed for joint extraction of Chinese entity relations.Based on the sequence features extracted by the bidirectional long short term memory network, this method uses GCN to encode the grammatical structure information in dependency analysis results, and employs the idea of an improved entity tagging strategy to build an end-to-end model for the joint extraction of Chinese entity relations.Experimental results show that this method displays an F score of 61.4%, which is 4.1% higher than the LSTM-LSTM model.GCN can effectively encode the prior relations between words contained in the text, and effectively improve the performance of entity relation extraction.

Key words: information extraction, relation extraction, joint extraction, Graph Convolutional neural Network(GCN), dependency analysis

摘要: 现有实体关系联合抽取方法未充分考虑中文句子中实体关系的复杂结构特征,为此,提出一种基于图卷积神经网络(GCN)的中文实体关系联合抽取方法。在双向长短时记忆网络抽取序列特征的基础上,利用GCN编码依存分析结果中的语法结构信息,借鉴改进的实体标注策略构建端到端的中文实体关系联合抽取模型。实验结果表明,该方法的F值可达61.4%,相比LSTM-LSTM模型提高了4.1%,GCN能有效编码文本的先验词间关系并提升实体关系抽取性能。

关键词: 信息抽取, 关系抽取, 联合抽取, 图卷积神经网络, 依存分析

CLC Number: