融合多特征和迭代扩张卷积的中文电子病历命名实体识别  

Chinese electronic medical record named entity recognition based on multi-features and IDCNN

作　　者：封红旗[1] 孙杨[1] 吴涛 王少聪 李文杰[2,3] FENG Hongqi;SUN Yang;WU Tao;WANG Shaocong;LI Wenjie(School of Computer Science and Artificial Intelligence,Changzhou University,Changzhou 213164,China;School of Microelectronics and Control Engineering,Changzhou University,Changzhou 213164,China;Changzhou Key Laboratory of Biomedical Information Technology,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]常州市生物医学信息技术重点实验室,江苏常州213164

出　　处：《常州大学学报（自然科学版）》2023年第1期59-67,共9页Journal of Changzhou University：Natural Science Edition

基　　金：江苏省科技厅社会发展资助项目(BE2018638);常州市社会发展资助项目(CE20195025)。

摘　　要：针对中文电子病历命名实体识别过程中文本语义表示不充分、特征抽取效率低等缺陷,提出一种融合多特征和迭代扩张卷积的命名实体识别方法。该方法首先构建基于卷积神经网络(CNN)的字嵌入算法,将生成的字向量与词向量等外部特征信息融合后送入迭代扩张卷积神经网络(IDCNN)中进行特征抽取,引入注意力机制加强序列间依赖关系,最后通过CRF解码最优标签序列。该方法在CCKS2017中文电子病历数据集中取得了91.36%的F1值,识别性能优于现有方法,同时验证了融合多特征的语义表示对中文实体识别有一定性能提升。Aiming at the problems of word segmentation errors,fuzzy word boundaries and low model calculation efficiency in the process of entity recognition task,a Chinese electronic medical record named entity recognition method that combines multiple features and IDCNN is proposed.This method first constructs a CNN-based character embedding algorithm to train the char vector,then splices it with the word vector and other additional features,then sends it to the iterative expanded convolutional neural network for feature learning,and finally decodes the optimal label sequence through CRF.Experimental results show that the F1 value of this method in the CCKS 2017 Chinese electronic medical record dataset reaches 91.36%,and the training efficiency is better than the existing model,which verifies the effectiveness of the method.

关 键 词：中文电子病历 命名实体识别 卷积神经网络 自注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]