融合胶囊网络与细粒度语义匹配的多标签文本分类

打开文本图片集

关键词：多标签文本分类；标签注意力；文本自注意力；协同注意力；胶囊网络中图分类号：TP391 文献标志码：A 文章编号:1001-3695(2025)12-022-3698-09doi:10.19734/j. issn.1001-3695.2025.04.0115

Integrating capsule networks with fine-grained semantic matching for multi-label text classification

SongKe，Li Xiang†，LinJianchu (FacultyofComputer&Software，HuaiyinInstituteofTechnology，HuaianJiangsu223OO3，China)

Abstract：Deeplearning-basedmulti-labeltextclassificationfaces twomainchallnges:theneglectoflabelsemanticnformationand inter-labeldependencies，andthelackoffine-grainedsemanticmatchingbetweentextsandlabels.Toaddressthese isues，thispaperproposedamulti-labeltextclasficationmethodintegratingcapsulenetworksandfine-grainedsemantic matching.Itfrstlyusedlabelembeddings to incorporate label information intothe graph structure,andemployeda graphconvolutionalnetorktopropagateinfomationacross nodes，ectivelyapturinglabelcorelations.Subsequently,italculated labelattentionandtextself-tention,andconstucedlabelco-attentionandtextcoatentionnetworks.Fiall,itfedtheextractedfeatures intoacapsulenetwork classfier tocapturemulti-levelsemantic informationandmatch thecorresponding labels.Experimentsconductedonthe AAPD,RCV1-V2,andEUR-Lexdatasetsdemonstratethat theproposedmethodoutperforms existing approaches in terms of P@k and nDCG@ k metrics. The results confirm the effectiveness of the capsule network and the fine-grained semantic matching strategy.

Keywords：multi-label text clasification；labelattention；textself-attention；co-attention；capsulenetwork:

0 引言

多标签文本分类（multi-labeltextclassification，MLTC）的任务是为输入的每个样本分配与其相关的所有标签[1]。（剩余32237字）