计算机应用

• 人工智能与仿真 •    下一篇

结合注意力机制的 Bi-LSTM-CRF中文电子病历命名实体识别

张华丽,康晓东,李博,王亚鸽,刘汉卿,白放   

  1. 天津医科大学 医学影像学院
  • 收稿日期:2019-08-07 修回日期:2019-10-11 发布日期:2019-10-11 出版日期:2020-05-12
  • 通讯作者: 康晓东
  • 作者简介:张华丽(1995—),女,河南信阳人,硕士研究生,主要研究方向:医学图像处理; 康晓东(1964—),男,天津人,教授,博士,CCF高 级会员,主要研究方向:医学图像处理、医疗信息系统集成; 李博(1987—),男,天津人,硕士研究生,主要研究方向:医学图像处理; 王亚鸽 (1992—),女,河南平顶山人,硕士研究生,主要研究方向:医学图像处理; 刘汉卿(1997—),男,湖南衡阳人,硕士研究生,主要研究方向:医学 图像处理; 白放(1995—),男,陕西榆林人,本科生,主要研究方向为医学图像处理。
  • 基金资助:
    京津冀协同创新项目(17YFXTZC0020)

Medical name entity recognition based on Bi-LSTM-CRF and attention mechanism

ZHANG Huali,KANG Xiaodong,LI Bo,WANG Yage,LIU Hanqing,BAI Fang   

  • Received:2019-08-07 Revised:2019-10-11 Online:2019-10-11 Published:2020-05-12

摘要: 在中文电子病历命名实体识别任务中,为了消除传统命名实体识别方法高度依赖人工提取特征这一不足,设计了双向长短时记忆(Bi-LSTM)网络与条件随机场(CRF)结合的网络模型,并在联合网络的基础上添加注意力机制,从而优化实体识别准确率。首先,将中文电子病历数据集进行脱敏处理及序列标注等预处理;其次,结合词嵌入技术将电子病历文本序列进行词向量化表示,并利用 Bi-LSTM 网络模型构造包含前向和后向文本的语义特征;然后,将双向特征序列输入到注意力层,利用注意力机制对文本特征向量的语义编码分配不同的注意力权重,进一步强化当前信息与上下文信息之间潜在的语义关联性;最后,输入到 CRF层中,由此提取出实体。实验结果表明,该注意力机制与 Bi-LSTM-CRF模型融合的新方法能有效提高中文电子病历命名实体识别的准确率。

关键词: 电子病历, 双向长短时记忆网络, 条件随机场, 注意力机制, 实体识别

Abstract: In the Chinese electronic medical record named entity recognition task,in order to eliminate the limitation that the traditional named entity recognition method is highly dependent on manually extracting features,a network model combining the Bidirectional Long Short-Term Memory (Bi-LSTM) network and Conditional Random Field (CRF) was designed;and attention mechanism was added to the joint network to optimize the accuracy and effectiveness of entity recognition. Firstly,the Chinese electronic medical record dataset was subjected to desensitization processing and sequence labeling,etc. Secondly,the electronic medical record text sequence was vectorized and represented by word embedding technique,and the Bi-LSTM model was used to construct the forward and backward text semantic features. Then,the bidirectional feature sequences were input into the attention layer,and the attention mechanism was used to assign different attention weights to the semantic codes of the text feature vectors,therefore further strengthening the potential semantic correlation between the current information and the context information. Finally,the weighted semantic representation was input to the CRF layer,thereby extracting the entity. The results show that the new method of combining the attention mechanism and Bi-LSTM-CRF model is effective to improve the accuracy of Chinese electronic medical record named entity recognition.

Key words: electronic medical record, Bidirectional Long Short-Term Memory (Bi-LSTM) network, Conditional Random Field (CRF), attention mechanism, entity recognition

中图分类号: