基于边缘对比学习的持续关系抽取

打开文本图片集

中图分类号：TP391.1 文献标识码：A文章编号：1006-8228（2025）11-13-06

Continual Relation Extraction Through Margin Contrastive Learning

Zuo Yang，Wang Congcong，Ge Baoquan

（SchoolofArtificialIntelligence，XinjiangVocationalUniversityofTechnologyKashgar，Xinjiang844ia）

Abstract：TheContinualRelationExtractionmodelisre-trainedwithdatasetscontainingnewrelations，iteasilyleadstothe catastrophicforgetingproblemcausedbyre-weightingthemodelparametersofrelevantpriorrelationtypes，especiallynearthe decisionborderlineofanalogousrelationpairs.Therefore，weproposetheMarginalContrastiveLearing-basedContinualelation Extractionmodel（MCL-CRE）.Ourmodeladoptsmarginalcontrastivelearningtoenhancetheabilitytodistinguishexisting relationsfromnewrelationsandenlargesthedistancebetweeanalogousrelationrepresentationsnearthedecisionborderline. Experimentalresultsshowthatourmodelcanstablydistinguishanalogousrelations，andachievesignificantresultsonthe unbalanced TACRED dataset.

Keywords：ContinualRelation Extraction; CatastrophicForgeting;Analogous Relations;Marginal ContrastiveLearning

0引言

在自然语言处理（Natural Language Processing，NLP）领域，信息抽取（Information Extraction，IE）的核心作用是从非结构化的数据中提取结构化的知识。（剩余9613字）