《计算机应用》唯一官方网站 ›› 2023, Vol. 43 ›› Issue (S1): 61-66.DOI: 10.11772/j.issn.1001-9081.2022060921

• 人工智能 • 上一篇    下一篇

基于相互学习和SoftLexicon的中文命名实体识别模型

陈田1, 黄泓毓2(), 杨东升2, 董淑婷2   

  1. 1.中国航空工业集团公司 成都飞机设计研究所,成都 610091
    2.电子科技大学 信息与软件工程学院,成都 610054
  • 收稿日期:2022-06-24 修回日期:2022-10-11 接受日期:2022-10-17 发布日期:2023-07-04 出版日期:2023-06-30
  • 通讯作者: 黄泓毓
  • 作者简介:陈田(1984—),男,湖北武汉人,高级工程师,硕士,主要研究方向:航电系统、人机交互
    黄泓毓(1998—),女,四川自贡人,硕士,主要研究方向:知识图谱、自然语言处理。hyhuang@std.uestc.edu.cn
    杨东升(1996—),男,重庆人,硕士研究生,主要研究方向:知识图谱、自然语言处理
    董淑婷(1998—),女,安徽芜湖人,硕士研究生,主要研究方向:数字人、动画驱动。
  • 基金资助:
    四川省科技服务业示范项目(2020GFW068);四川省科技成果转移转化示范项目(2021ZHCG0007)

Chinese named entity recognition model based on mutual learning and SoftLexicon

Tian CHEN1, Hongyu HUANG2(), Dongsheng YANG2, Shuting DONG2   

  1. 1.Chengdu Aircraft Design Research Institute,Aviation Industry Corporation of China,Chengdu Sichuan 610091,China
    2.School of Information and Software Engineering,University of Electronic Science and Technology of China,Chengdu Sichuan 610054,China
  • Received:2022-06-24 Revised:2022-10-11 Accepted:2022-10-17 Online:2023-07-04 Published:2023-06-30
  • Contact: Hongyu HUANG

摘要:

中文自然语言文本中实体边界区分难、语法复杂度大,中文命名实体识别(NER)难度往往比英文命名实体识别大。针对中文NER中分词误差传播的问题,提出一种基于相互学习和SoftLexicon的中文命名实体识别模型MM-SLLattice。首先,向字级别表示的句子中加入词信息的模型;然后,在词信息的引入过程中通过结合开放词典与领域词典信息来提高模型的精度;最后,在训练过程中,引入了深度相互学习减小泛化误差提高模型的性能。实验结果表明,该模型在不同类型的中文数据集的实体识别能力有提升,MM-SLLattice在MSRA数据集上F1值为94.09%,比独立网络提高了0.41个百分点,对比实验中F1值也优于其他主流模型协同图形网络(CGN)、卷积注意力网络(CAN)、LR-CNN。所提模型可以更精确地提取中文实体。

关键词: 知识图谱, 命名实体识别, SoftLexicon, 双向长短期记忆, 自注意力

Abstract:

The difficulty of entity boundary differentiation and syntactic complexity in Chinese natural language text are often greater than those of English named entity recognition. Aiming at the problem of word segmentation error propagation in Chinese Named Entity Recognition (NER), a Chinese named entity recognition model based on mutual learning and SoftLexicon, MM-SLLattice (SoftLexicon Lattice with Deep Multi-Mutual Learning Network), was proposed. Firstly, word information was added into sentences represented at character level. Then, the accuracy of the model was improved by combining open dictionary and domain dictionary information during the introduction of word information. Finally, deep mutual learning was introduced to reduce the generalization error during the training process to improve the performance of the model. The experimental results show that the model has improved entity recognition ability in different types of Chinese datasets, and the F1 value of MM-SLLattice on MSRA dataset is 94.09%, 0.41 percentage points better than that of the isolated network. The F1 value in comparison experiments is also better than those of other mainstream models Collaborative Graph Network(CGN), Convolutional Attention Network(CAN), and LR-CNN(Lexicon Rethinking Convolutional Neural Network). The proposed model can more accurately extract Chinese entities.

Key words: knowledge graph, Named Entity Recognition (NER), SoftLexicon, Bi-directional Long Short-Term Memory(BiLSTM), self-attention

中图分类号: