基于Biaffine机制和词汇增强的中文命名实体识别方法

打开文本图片集

关键词：中文命名实体识别；多级词汇增强；Biaffine机制；特征融合

中图分类号：TP391.1 文献标志码：A 文章编号：1001-3695（2026）01-014-0120-09

doi：10.19734/j.issn.1001-3695.2026.05.0152

Chinese named entity recognition method based on Biafine mechanism and lexical enhancement

Zhang Runmeia，b，Wang Mingxia†，Chen Zhongh （a.Schoolofecocdfoatning，olfcaldlecrcalEneiJUitei 230601，China）

Abstract：To address the challenges of fuzzy entityboundaries，complex structures，and scarce domain-specificdata in Chinese named entityrecognition（NER），this paper proposed WLASC，a Chinese NER model based ona Biaffine mechanism andlexicalenhancement.Intheencodinglayer，themodelincorporatedadynamic biafine moduleandamult-levellexical enhancement module.Byintroducing relativepositionencodingand Biafinetransformation，themodelstrengthenedcontextual modelingcapabilities，fectivelyresolvingfuyntityboundarie.Simultaneously，itleveragedmulti-levellexicalifoation andamulti-eadatentionmechanismtoweightandfusefeaturesofdierentgranularities，improvingnestedentityrecognition accuracywhilereducingrelianceonannotateddata.Additionall，themodelemployedabidirectional gatedrecurrentneural networktofuseextractedfeatures，furterenhancing itsexpressivepower.Experimentalresultsontheaviationflightsafetydomaindataset CANERand publicdatasets（Weibo，Resume）demonstrate that the improved algorithmachieves maximum F1 1 score improvements of 9. 77% ， 5.97% and 1.38% ，respectively，with minimum improvements of 2.72% ， 1.27% and （20 0.31% .Testson the CANER dataset confirm that the model efectively handles structurallycomplexentitiesand domainspecificterminology inspecialized Chinesedomains.Experimentsonpublicdatasetsfurtherindicatethatthemodelexhibits strong generalization capabilities.

Keywords：Chinese named entityrecognition；multilevel word enhancement；Biaffne mechanism；feature fusion

0 引言

命名实体识别（namedentityrecognition，NER）作为自然语言处理（natural language processing，NLP）技术中的一项关键任务[1]，其主要目标是确定实体的类型和边界。（剩余21620字）