计算机与现代化JISUANJIYUXIANDAIHUA2023年第1期总第329期摘要：为避免实体与关系独立抽取产生的误差累计及实体重叠问题，提出一种基于BERT和非自回归的联合抽取模型来进行医疗知识抽取。首先，通过BERT预训练语言模型进行句子编码；然后，采用非自回归（NAR，Non-autoregressive）的方法实现并行解码，抽取关系类型，并根据头尾实体的位置索引抽取实体，得到医疗实体的关系三元组；最后，将抽取出的实体和关系导入Neo4j图数据库中实现知识可视化。通过对电子病历中的数据进行人工标注得到数据集，实验结果表明，基于BERT和非自回归联合学习模型的F1值为0.92，precision值为0.93，recall值为0.92，与现有模型相比3项评价指标均有提升，表明本文方法能够有效抽取电子病历中的医疗知识。关键词：联合学习；非自回归；BERT；实体重叠；电子病历中图分类号：TP391.1文献标志码：ADOI：10.3969/j.issn.1006-2475.2023.020MedicalKnowledgeExtractionBasedonBERTandNon-autoregressiveYUQing，MAZhi-long，XUChun（SchoolofInformationManagement,XinjiangUniversityofFinanceandEconomics,Urumqi830012,China）Abstract：Inordertoavoidtheproblemsoferroraccumulationandentityoverlapcausedbythepipelineentityrelationextractionmodel，ajointextractionmodelbasedonBERTandNon-autoregressiveisestablishedformedicalknowledgeextraction.Firstly，withthehelpoftheBERTpre-trainedlanguagemodel，thesentencecodeisobtained.Secondly，theNon-autoregressivemethodisproposedtoachieveparalleldecoding，extracttherelationshiptype，extractentitiesaccordingtotheindexofthesubjectandobjectentities，andobtainthemedicaltriplet.Finally，weimporttheextractedtriplesintotheNeo4jgraphdatabaseandrealizeknowledgevisualization.Thedatasetisderivedfrommanuallabelingofdatainelectronicmedicalrecords.TheexperimentalresultsshowthattheF1value，precisionandrecallbasedonBERTandnon-autoregressivejointlearningmodelare0.92，0.93and0.92，respectively.Comparedwiththeexistingmodel，thethreeevaluationindicatorshavebeenimproved，indicatingthattheproposedmethodcaneffectivelyextractmedicalknowledgefromelectronicmedicalrecords.Keywords：jointlearning；non-autoregressive；BERT；entityoverlap；electronicmedicalrecord...
