网络与信息安全学报 ›› 2018, Vol. 4 ›› Issue (9): 44-51.doi: 10.11959/j.issn.2096-109x.2018074

• 论文 • 上一篇    下一篇

基于CNN和双向LSTM融合的实体关系抽取

张晓斌(),陈福才,黄瑞阳   

  1. 国家数字交换系统工程技术研究中心,河南 郑州 450002
  • 修回日期:2018-08-20 出版日期:2018-09-15 发布日期:2018-10-15
  • 作者简介:张晓斌(1993-),男,福建漳州人,国家数字交换系统工程技术研究中心硕士生,主要研究方向为文本挖掘、实体关系抽取。|陈福才(1974-),男,江西高安人,硕士,国家数字交换系统工程技术研究中心研究员、硕士生导师,主要研究方向为大数据分析与处理。|黄瑞阳(1986-),男,福建漳州人,国家数字交换系统工程技术研究中心助理研究员,主要研究方向为文本挖掘、信息抽取。
  • 基金资助:
    国家自然科学基金创新群体基金资助项目(61521003);国家重点研发计划基金资助项目(2016YFB0800101)

Relation extraction based on CNN and Bi-LSTM

Xiaobin ZHANG(),Fucai CHEN,Ruiyang HUANG   

  1. China National Digital Switching System Engineering &Technological R&D Center,Zhengzhou 450002,China
  • Revised:2018-08-20 Online:2018-09-15 Published:2018-10-15
  • Supported by:
    The National Natural Science Fund for Creative Research Groups Project(61521003);The National Key Research and Development Program of China(2016YFB0800101)

摘要:

实体关系抽取旨在识别网络文本中的实体,并提取出文本中实体之间隐含的关系。研究表明,深度神经网络在实体关系抽取任务上具有可行性,并优于传统关系抽取方法。目前的关系抽取方法大都使用卷积神经网络(CNN)和长短期记忆神经网络(LSTM),然而CNN只考虑连续词之间的相关性而忽略了非连续词之间的相关性。另外,LSTM虽然考虑了长距离词的相关性,但提取特征不够充分。针对这些问题,提出了一种CNN和LSTM结合的实体关系抽取方法,采用3种结合方法进行了实验,验证了该方法的有效性,在F1值上有一定的提升。

关键词: 实体关系抽取, 卷积神经网络, 长短期记忆网络, 注意力机制

Abstract:

Relation extraction aims to identify the entities in the Web text and extract the implicit relationships between entities in the text.Studies have shown that deep neural networks are feasible for relation extraction tasks and are superior to traditional methods.Most of the current relation extraction methods apply convolutional neural network (CNN) and long short-term memory neural network (LSTM) methods.However,CNN just considers the correlation between consecutive words and ignores the correlation between discontinuous words.On the other side,although LSTM takes correlation between long-distance words into account,the extraction features are not sufficiently extracted.In order to solve these problems,a relation extraction method that combining CNN and LSTM was proposed.three methods were used to carry out the experiments,and confirmed the effectiveness of these methods,which had some improvement in F1 score.

Key words: relation extraction, convolution neural networks, long short-term memory, attention mechanism

中图分类号: 

No Suggested Reading articles found!