易混淆样本驱动的簇间分布优化短文本聚类

打开文本图片集

中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2025）10-013-2996-09

doi：10.19734/j. issn.1001-3695.2025.03.0075

Confusing sample-driven inter-cluster optimization method for short text clustering

Enkaer Nuertai 1，2，3 ，Ma Bo 1，2，3† ，Wang Zhen 1，2，3 ， Aizimaiti Ainiwaer 1，2，3 ， Tuerhong Wusiman 1，3 ， Yang Yating1，2，3 （1.Xinjangltefic&strecef，Uin;Uitfd myofSciencesngbfiti

Abstract：Short textclusteringaims topartitionunlabeledshorttextinstances intodiferentsemanticclusters.This task faces challnges fromindistinguishableconfusing samplesandoverlappng feature distributions between semanticallysimilarclusters. Totacklethesechalenges，thispaperproposedaconfusingsample-driven inter-clusterdistributionoptimizationmethod for shorttextclustering.Thismethodfirstlysampledhigh-uncertaintyinstancesbasedoninformationentropyasconfusingsamples，adselectedtheir neighboring cluster samples ascandidates.Then，it leveragedlarge language models to semantic discriminationandformed“confusing-positive-negative”triplets.Meanwhile，this methodadoptedadataaugmentationmethod basedonparameterrandomperturbation to generate instance-levelpositivesforeachsample.Finall，iperformedjointoptimizationoftheinter-clusterdistributionwithinacontrastivelearning framework．Experimentalresultsonfourpublicshorttext datasets demonstratethattheproposedmethodoutperforms existingstate-of-the-artmodels，withanaverageacuracyimprovement of 5.14% and an average normalized mutual information increase of 2.51% . Further analysis confirms that the method significantlyenhancessemanticdiscriminationofconfusing samples between clustersand efectivelyalleviates featureoverlap between semantically similar clusters.

Key Words：confusing samples； short text clustering; large language models；contrastive learning

0 引言

随着数字化进程的加速，短文本作为信息传播和交互的主要形式之一（如社交媒体帖子、即时消息、新闻标题、商品评论等），其数量和重要性显著增加。（剩余23892字）