结合注意力机制的BERT-BiGRU-CRF中文电子病历命名实体识别  

Named Entity Recognition for Chinese Electronic Medical Record Based on BERT-BiGRU-CRF and Attention Mechanism

作　　者：陈娜[1] 孙艳秋[1] 燕燕[1] CHEN Na;SUN Yan-qiu;YAN Yan(School of Information Engineering,Liaoning University of Traditional Chinese Medicine,Shenyang 110847,China)

机构地区：[1]辽宁中医药大学信息工程学院,沈阳110847

出　　处：《小型微型计算机系统》2023年第8期1680-1685,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(82274580)资助;辽宁省教育厅科学研究项目(L2020059)资助;辽宁省教育厅高等学校基本科研项目(LJKZ0894)资助;辽宁中医药大学人文社科类项目(2021LNZYQN014)资助。

摘　　要：为了改善中文电子病历命名实体识别模型的性能,本文提出了基于BERT、双向门控循环单元(Bidirectional Gate Recurrent Unit,BiGRU)与条件随机场CRF相结合的中文电子病历命名实体识别模型,并在此基础上引入了注意力机制.利用BERT(Bidirectional Encoder Representation from Transformers)预训练模型得到结合语境信息的动态字向量,通过双向门控循环单元(Bidirectional Gate Recurrent Unit,BiGRU)提取全局语义特征,利用注意力机制获得增强语义特征,最后通过CRF(Conditional Random Field)解码输出概率最大的全局最优标签序列.利用含有解剖部位、手术、疾病和诊断、药物、实验室检验、影像检查6类实体的CCKS19中文电子病历数据集训练模型.对比实验表明了本文提出的命名实体识别模型的有效性,本文模型在CCKS19数据集上获得了84.11%的F1值。In order to improve the performance of Chinese electronic medical record named entity recognition model,this paper proposes a named entity recognition model based on BERT,Bidirectional Gate Recurrent Unit and conditional random field;and the attention mechanism is added to this model.A dynamic word vector combining contextual information is obtain by BERT pre-training model.Global semantic features are extracted by Bidirectional Gate Recurrent Unit.Enhanced semantic features are obtained by attention mechanism.Finally,the global optimal label sequence with the highest probability is output by CRF.The model is trained with the CCKS19 Chinese e-case datasets which contains six types of entities:body,operation,diagnosis,medicine,check and examination.The effectiveness of the named entity recognition model proposed in this paper is demonstrated by comparison experiments,and the model in this paper obtains 84.11%entity recognition F_(1) score on the CCKS19 datasets.

关 键 词：中文电子病历 命名实体识别 BERT BiGRU 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]