图书情报工作 ›› 2020, Vol. 64 ›› Issue (17): 131-144.DOI: 10.13266/j.issn.0252-3116.2020.17.014

• 知识组织 • 上一篇    下一篇

跨语言情境下基于对抗的实体关系抽取模型研究

余传明1, 王曼怡2, 安璐3   

  1. 1 中南财经政法大学信息与安全工程学院 武汉 430073;
    2 中南财经政法大学统计与数学学院 武汉 430073;
    3 武汉大学信息管理学院 武汉 430072
  • 收稿日期:2020-02-08 修回日期:2020-04-22 出版日期:2020-09-05 发布日期:2020-09-05
  • 作者简介:余传明(ORCID:0000-0001-7099-0853),教授,E-mail:yucm@zuel.edu.cn;王曼怡(ORCID:0000-0002-0633-0073),硕士研究生;安璐(ORCID:0000-0002-5408-7135),教授,博士生导师。
  • 基金资助:
    本文系国家自然科学基金面上项目"面向跨语言观点摘要的领域知识表示与融合模型研究"(项目编号:71974202)和国家自然科学基金重大课题"国家安全大数据综合信息集成与分析方法"(项目编号:71790612)研究成果之一。

Research on the Model of Adversarial Entity Relation Extraction in Cross-Lingual Context

Yu Chuanming1, Wang Manyi2, An Lu3   

  1. 1 School of Information and Safety Engineering, Zhongnan University of Economics and Law, Wuhan 430073;
    2 School of Statistics and Mathematics, Zhongnan University of Economics and Law, Wuhan 430073;
    3 School of Information Management, Wuhan University, Wuhan 430072
  • Received:2020-02-08 Revised:2020-04-22 Online:2020-09-05 Published:2020-09-05

摘要: [目的/意义] 从实体关系抽取视角出发,将单一语言情境下的知识获取任务扩展到跨语言情境,提升低资源语言的关系抽取效果。[方法/过程] 提出一种跨语言对抗关系抽取(Cross-Lingual Adversarial Relation Extraction,CLARE)框架,将跨语言关系抽取分解为平行语料获取和对抗适应关系抽取两个子模块。通过词典扩展或自学习方法将源语言关系抽取数据集转换为目标语言数据集,在此基础上利用对抗特征适应将源语言的特征表示迁移给目标语言,再利用训练得到的目标语言关系抽取网络对目标语言进行关系分类。[结果/结论] 将本文方法应用到以ACE2005多语言数据集为基础的英语-中文、中文-英文两种跨语言关系抽取任务上,最优模型的Macro-F1值分别为0.880 1和0.842 2。实验结果表明本文提出的跨语言对抗关系抽取CLARE框架能显著提升低资源语言实体关系抽取的效果。研究结果对于改进跨语言情境下的关系抽取模型以及促进实体关系抽取研究在情报学领域的应用具有重要意义。

关键词: 跨语言信息抽取, 实体关系抽取, 深度学习, 生成对抗网络

Abstract: [Purpose/significance] From the perspective of entity relation extraction, the knowledge acquisition task in a single language context is extended to a cross-language context, and the relation extraction effect of low-resource languages is improved.[Method/process] This paper proposed a Cross-Lingual Adversarial Relation Extraction (CLARE) framework, which decomposed cross-lingual relation extraction into parallel corpus acquisition and adversarial adaptation relation extraction. Through dictionary expansion or self-learning methods, the source language relation extraction data set was converted into the target language data set. On this basis, the feature representation of the source language was transferred to the target language using adversarial feature adaptation, and then the target language relation extraction network obtained by training was used to classify the target language.[Result/conclusion] The method in this paper is applied to the English-Chinese and Chinese-English cross-lingual relation extraction task based on the ACE2005 multilingual dataset. The Macro-F1 values of the optimal models on the two tasks are 0.880 1 and 0.842 2 respectively, indicating that the proposed CLARE framework for cross-language adversarial relation extraction can significantly improve the effect of low-resource language entity relation extraction. The research results are of great significance for improving the relation extraction model in the cross-lingual context and promoting the application of entity relation extraction research in the field of information science.

Key words: cross-lingual information extraction, entity relation extraction, deep learning, generative adversarial network

中图分类号: