基于多任务的中文电子病历中命名实体识别研究  

Multi-task based Chinese electronic medical record entity recognition

作　　者：杨晓辉 毕雪华[2] 张琳琳[3] 高颖 YANG Xiao-hui;BI Xue-hua;ZHANG Lin-lin;GAO Ying(School of Software,Xinjiang University,Urumqi 830046,China;College of Medical Engineering and Technology,Xinjiang Medical University,Urumqi 830011,China;College of Information Science and Engineering,Xinjiang University 830046)

机构地区：[1]新疆大学软件学院,新疆乌鲁木齐830046 [2]新疆医科大学医学工程技术学院,新疆乌鲁木齐830011 [3]新疆大学信息科学与工程学院,新疆乌鲁木齐830046

出　　处：《东北师大学报（自然科学版）》2020年第1期81-87,共7页Journal of Northeast Normal University(Natural Science Edition)

基　　金：新疆维吾尔自治区自然科学基金资助项目(2017D01C232).

摘　　要：针对电子病历结构化中的命名实体识别问题,提出一种基于分词任务和命名实体识别任务相结合的多任务双向长短期记忆网络模型(Bidirectional Long Short-Term Memory Conditional Random Feld,Bi-LSTM-CRF).该模型通过加入共享LSTM捕获分词任务中的词边界信息,丰富了命名实体识别任务的特征集,进而达到提高命名实体的效果.实验采集数据由新疆某三甲医院提供的500份冠心病患者的出院小结和2000份心血管疾病患者的出院小结组成.实验结果表明:F-measure值达到了0.927,满足临床的实际需求.Aiming at the problem of named entity recognition in electronic medical record structuring,a multi-task Bidirectional long short-term memory network model based on word segmentation task and named entity recognition task is proposed.By adding the word boundary information in the word segmentation task by adding the shared LSTM,the model enriches the feature set of the named entity recognition task,there by achieving the effect of improving the named entity.The experimental data consisted of 500 discharge symptoms of coronary heart disease patients and a discharge summary of 2000 patients with cardiovascular disease provided by a top three hospital in Xinjiang.The experimental results F-measure value reached 0.927,meeting the actual clinical needs.

关 键 词：电子病历 命名实体识别 多任务 Bi-LSTM-CRF 

分 类 号：R319[医药卫生—基础医学] TP391[自动化与计算机技术—计算机应用技术]