基于RoBERTa-Prompt-R-Drop新闻主题分类

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）12-44-06

Abstract：Toaddressthechallengesofmissngcontextanddatasparsityinnewstopicclasification，weproposeajoint optimizationframeworkthatintegratesRoBERTaPromptLearning，andR-Drop.Theframeworkreformulatesclasificationasa masked-language-modeltaskviaprompts，therebyactivatingRoBERTa'spre-trainedsemanticknowledgetocompensateforthelack ofcontext.ConurentlyR-DropapliesaKL-divergenceconsrinttotwostochastic-forwardasssofthesameinputyieldinga negative-samplefreecontrastiveregularizationthatdrivesthemodeltolearnnoise-robustrepresentationsandavoidsthepitfallof low-quality negative construction.Experiments on THUCNews show that our method achieves 96.61% accuracy，significantly outperformingallbaselinesandfullyconfirmingitseffectivenessinimprovingbothclasificationaccuracyandmodelrobustes.

Keywords：Text Classification；News Topic;RoBERTa;Prompt Learning;R-Drop

0引言

随着移动互联网的普及，新闻数据呈爆炸式增长。（剩余11228字）