基于BERT的中文电子病历命名实体识别  

Chinese electronic medical record named entity recognition based on BERT methods

作　　者：封红旗[1] 孙杨[1] 杨森[1] 李文杰[2,3] FENG Hong-qi;SUN Yang;YANG Sen;LI Wen-jie(School of Computer Science and Artificial Intelligence,Changzhou University,Changzhou 213164,China;School of Microelectronics and Control Engineering,Changzhou University,Changzhou 213164,China;Key Laboratory of Biomedical Information Technology,Changzhou University,Changzhou 213164,China)

机构地区：[1]常州大学计算机与人工智能学院,江苏常州213164 [2]常州大学微电子与控制工程学院,江苏常州213164 [3]常州大学生物医学信息技术重点实验室,江苏常州213164

出　　处：《计算机工程与设计》2023年第4期1220-1227,共8页Computer Engineering and Design

基　　金：江苏省科技厅社会发展基金项目(BE2018638);常州市社会发展基金项目(CE20195025)。

摘　　要：针对中文电子病历命名实体识别过程中实体特征利用率低,语义表示不充分等问题,提出一种基于BERT语言模型的命名实体识别方法。运用Char-CNN学习字符的多种特征,将特征加入BERT预训练生成的词向量中,获得融合领域信息和汉字特征的词向量表示,将词向量输入迭代扩张卷积神经网络中进行特征抽取,引入注意力机制加强实体特征的关注度,通过CRF解码标注命名实体。实验结果表明,该方法在CCKS17中取得91.64%的F1值,识别性能优于现有方法。Aiming at the problems of low entity features utilization and insufficient semantic representation in the process of identifying named entities in Chinese electronic medical records,a method for identifying named entities based on BERT methods was proposed.Char-CNN was used to learn various features of character,and features were added to the word vector generated using BERT pre-trained model.A word vector representation with fusion domain information and character features was obtained.An iterative dilated convolutional neural network was used for feature extraction,the attention mechanism was introduced to enhance the attention of entity features.The named entity was labeled through CRF decoding.Experimental results show that the proposed method achieves 91.64%F 1 value in CCKS17,and its recognition performance is better than that of existing methods.

关 键 词：中文电子病历 命名实体识别 深度学习 语言模型 卷积神经网络 注意力机制 词向量 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]