ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2019, Vol. 56 ›› Issue (9): 1851-1858.doi: 10.7544/issn1000-1239.2019.20180733

• 人工智能 • 上一篇    下一篇

结合GAN与BiLSTM-Attention-CRF的领域命名实体识别

张晗1,2, 郭渊博1, 李涛1   

  1. 1(战略支援部队信息工程大学密码工程学院 郑州 450001); 2(郑州大学软件学院 郑州 450001) (zhang_han@zzu.edu.cn)
  • 出版日期: 2019-09-10
  • 基金资助: 
    国家自然科学基金项目(61501515);河南省重点科技攻关项目(172102210002);郑州大学青年骨干教师项目(2017ZDGGJS048)

Domain Named Entity Recognition Combining GAN and BiLSTM-Attention-CRF

Zhang Han1,2, Guo Yuanbo1, Li Tao1   

  1. 1(Department of Cryptogram Engineering, Strategic Support Force Information Engineering University, Zhengzhou 450001); 2(Software College, Zhengzhou University, Zhengzhou 450001)
  • Online: 2019-09-10
  • Supported by: 
    This work was supported by the National Natural Science Foundation of China (61501515), the Key Scientific and Technological Research Project of Henan Province (172102210002), and the Young Scholar Teachers Project of Zhengzhou University (2017ZDGGJS048).

摘要: 领域内命名实体识别通常面临领域内标注数据缺乏以及由于实体名称多样性导致的同一文档中实体标注不一致等问题.针对以上问题,利用生成式对抗网络(generative adversarial network, GAN)可以生成数据的特点,将生成式对抗网络与BiLSTM-Attention-CRF模型相结合.首先以BiLSTM-Attention作为生成式对抗网络的生成器模型,以CNN作为判别器模型,从众包标注数据集中整合出与专家标注数据分布一致的正样本标注数据来解决领域内标注数据缺乏的问题;然后通过在BiLSTM-Attention-CRF模型中引入文档层面的全局向量,计算每个单词与该全局向量的关系得出其新的特征表示以解决由于实体名称多样化造成的同一文档中实体标注不一致问题;最后,在基于信息安全领域众包标注数据集上的实验结果表明,该模型在各项指标上显著优于同类其他模型方法.

关键词: 领域命名实体识别, 生成式对抗网络, 众包标注数据, 实体标注一致, BiLSTM-Attention-CRF模型

Abstract: Domain named entity recognition usually faces the lack of domain annotation data and the inconsistency of entity annotation in the same document due to the diversity of entity names in the domain. To issue the above problems, our work draws on the combination of the generative adversarial network (GAN) which can generate data and the BiLSTM-Attention-CRF model. Firstly, BiLSTM-Attention is used as the generator model of GAN, and CNN is used as the discriminant model. The two models use the crowd annotations and the expert annotations to train respectively, and integrate the positive annotation data consistent with the expert annotation data distribution from the crowd annotations to solve the problem of lack of annotation data in the domain; then we also introduce a new method to obtain the new feature representation of each word in the document through introducing a document-level global feature in the BiLSTM-Attention-CRF model in order to solve the problem of inconsistency of the entity in the same document caused by the diversification of the entity name. Finally, taking the crowd annotations in the information security field as a sample, a comprehensive horizontal evaluation experiment is carried out by learning the common features and applying them to the training BiLSTM-Attention-CRF model for the identification of named entities in the information security field. The results show that compared with the existing models and methods, the model we proposed has made great progress on various indicators, reflecting its superiority.

Key words: domain named entity recognition, generative adversarial network (GAN), crowd annotations, entity annotations consistent, BiLSTM-Attention-CRF model

中图分类号: