基于BERT和改进对抗训练的痛风病历分类方法  

Text classification of gout medical records based on BERT and adversarial training

作　　者：李胜煜 王磊[2] 徐文畅 贺玉伟 李鑫德[3] LI Sheng-yu;WANG Lei;XU Wen-chang;HE Yu-wei;LI Xin-de(School of Biomedical Engineering,University of Science and Technology of China,Hefei 260026,China;CAS Key Lab of Bio-Medical Diagnostics,Suzhou Institute of Biomedical Engineering and Technology,Chinese Academy of Sciences,Suzhou 215163,China;Shandong Clinical Medical Research Center for Immune Diseases and Gout,Shandong Province Key Laboratory of Metabolic Disease,The Affiliated Hospital of Qingdao University,Qingdao 266000,China)

机构地区：[1]中国科学技术大学生物医学工程学院,安徽合肥260026 [2]中国科学院苏州生物医学工程技术研究所生物医学检验技术重点实验室,江苏苏州215163 [3]青岛大学附属医院山东省代谢性疾病重点实验室山东省免疫疾病与痛风临床医学研究中心,山东青岛266000

出　　处：《计算机工程与设计》2024年第6期1668-1673,共6页Computer Engineering and Design

基　　金：国家重点研发计划基金项目(2022YFC2503305);山东省重点研发计划(重大科技创新工程)基金项目(2021CXGC011103、2021SFGC0104)。

摘　　要：为解决痛风电子病历文本稀疏性高、数据集小导致的分类任务准确率低问题,提出一种基于BERT预训练模型和改进对抗训练的痛风病历文本分类算法。使用中文生物医学语言预训练模型MC-BERT初始化病历文本,下接Text-CNN网络捕捉文本中不同长度的关键词信息,在模型训练过程中采用改进的对抗训练策略,在词嵌入中添加对抗性扰动提高模型的泛化性。实验结果表明,该算法可以提高中文痛风病历文本分类任务的精度,增强模型的鲁棒性。To solve the problem of low accuracy of classification task due to high sparsity of text and small data set,a text classification algorithm of gout medical records based on BERT and improved adversarial training was proposed.The Chinese biome-dical language pre-training model MC-BERT was used to initialize the medical record text,followed by the Text-CNN network to capture the key words of different lengths in the text.In the process of model training,the improved adversarial training strategy was adopted,and the adversarial disturbance was added into the word embedding to improve the generalization of the model.Experimental results show that the algorithm can improve the accuracy of Chinese gout text classification task and enhance the robustness of the model.

关 键 词：痛风 电子病历 文本分类 卷积神经网络 对抗训练 预训练模型 词嵌入 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]