基于对抗式提示样本增强与原型学习的少样本持续关系抽取

打开文本图片集

中图分类号：TP391 文献标识码：A 文章编号：2096-4706（2025）18-0045-08

Abstract： Recent advances in Large Language Models have demonstrated impressive performance in terms of relation extraction.However，mtigatingtheisuesofcatastrophicforgetingandsimilarelationclassconfusionemainschallngingin dynamicenvironments withlimitedsamples.Toaddressthese isses，this paper proposesaMemory-basedContinualAdversarial Prototype Leaming （CAPL）method.CAPLemploysamemory-basedarchitecture thatidentifiesrepresentativesamples through extractedrelaonprotosformemoyplayostlvatetessuefatastroicgetig.Idion，Csigs anadversarialsamplegeneration strategy tailored for few-shotscenarios toaugmenttraining data and efectivelyreduce class confusion probability.Experimental results ontwopublic few-shot datasets show that CAPLachieves significant improvements over mainstream baseline methods， outperforming the previous best-performing methods by 4.82% and 5.91% in terms of averageaccuracyrespectively.

KeyWords： Continual Learning;Few-shot Learning; Catastrophic Forgeting; Prototype Learning;Adversarial Sample Augmentation

0 引言

持续关系抽取（ContinualRelationExtraction，CRE）是机器学习的子领域，核心是在动态环境下保持对旧有关系的持续记忆能力，在任务间转移知识，并通过学习新关系应对新任务。（剩余13983字）