基于多头注意力的中文电子病历命名实体识别  

NAMED ENTITY RECOGNITION BASED ON MULTI-HEAD ATTENTIONIN CHINESE ELECTRONIC MEDICAL RECORDS

作　　者：肖丹 杨春明[1,2] 张晖 赵旭剑[1,3] 李波 Xiao Dan;Yang Chunming;Zhang Hui;Zhao Xujian;Li Bo(School of Computer Science and Technology,Southwest University of Science and Technology,Mianyang 621010,Sichuan,China;Sichuan Big Data and Intelligent Systems Engineering Technology Research Center,Mianyang 621010,Sichuan,China;Chengdu Tianfu New Area Innovation Research Institute,Southwest University of Science and Technology,Chengdu 610299,Sichuan,China)

机构地区：[1]西南科技大学计算机科学与技术学院,四川绵阳621010 [2]四川省大数据与智能系统工程技术研究中心,四川绵阳621010 [3]西南科大成都天府新区创新研究院,四川成都610299

出　　处：《计算机应用与软件》2024年第1期133-138,160,共7页Computer Applications and Software

基　　金：教育部人文社科基金资助项目(17YJCZH260);赛尔网络下一代互联网技术创新项目(NGII20180403)。

摘　　要：针对中文电子病历中复杂医疗实体的识别问题,提出一种联合特征与多头注意力相结合的实体识别方法。该方法使用字符、词性和词典组成的联合特征,利用BiLSTM和多头注意力分别提取句子的全局特征和局部特征,利用CRF结合所有特征完成实体标签的预测。实验结果表明,该方法F1值达89.16%,其中治疗和疾病两类实体分别达到94.76%和95.56%。Aimed at the recognition problem of complex medical entities in Chinese electronic medical records(EMRs),an entity recognition method combining joint features and multi-head attention is proposed.This method used the joint feature composed of characters,parts of speech and dictionary,and used BiLSTM and multi-head attention to extract separately the global feature and local feature of the sentence.CRF was used to combine all the features to complete the prediction of the entity labels.Experimental results show that the F1-score of this method reaches 89.16%,among which the two types of entities,treatment and disease,reach 94.76%and 95.56%respectively.

关 键 词：命名实体识别 中文电子病历 多头注意力 长短期记忆网络 条件随机场 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]