面向商品评论的语义模式挖掘

打开文本图片集

中图分类号：TP391.1 文献标识码：A 文章编号：2096-4706（2025）22-0044-06

SemanticPatternMiningforProductReviews

YAO Ya， WANG Ru

（1.SchoolofComputer Science，Nanjing UniversityofPostsand Telecommunications，Nanjing 21oo23，China; 2.School ofSoftware，Nanjing UniversityofPosts and Telecommunications，Nanjing21oo23，China; 3.SchoolofCyberspace Security，Nanjing UniversityofPostsandTelecommunications，Nanjing21oo3，China; 4.GuangxiKeyLaboratoryofHybridComputationandICDesignAnalysis，GuangxiMinzuUniversityNanning53o6，China）

Abstract：Userreviews onE-Commerce platforms contain rich product informationand serve as valuable resources for mining textual semantic patterns.However，existing neural topic modelsoften suffer from insuffcient interpretabilityandpor diversity ngeneratedtopics.To address theaforementioned issues，this paper proposes a BERT-based NeuralTopic Model （BNTM）.Firstly，aninter-topiccontrastivelearning mechanismisdesigned toenhancetopicdistinctivenessbymaximizing the divergence between differenttopics.Secondly，a prior distribution matching framework is introduced to ensure topic interpretabilitybyaligning thelearnedtopicdistributions withDirichletpriors.Experimentsonthreebenchmark datasets demonstrate that the proposed model can generate semantically coherent and high-quality topics，significantly outperforming existing topic modeling methods.

Keywords： pattern mining; pre-trained language model; Contrastive Learning;E-Commerce revie

0 引言

随着电子商务平台的普及，海量用户评论数据已成为分析消费者需求的重要载体。（剩余11129字）