基于检索增强生成的主题推理模型

打开文本图片集

关键词：主题模型；大语言模型；检索增强生成；思维链；最优传输理论中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2025）10-016-3019-08doi：10.19734/j. issn. 1001-3695.2025.03.0059

Topic inference model based on retrieval augmented generation

PanLihu，Li Jie，Zhao Hongyan+（ColegeofComputerScience&Technology，TaiyuanUniversityofSience&Technology，TaiyuanO3oO24，China）

Abstract：Toaddressthechallngesof insuficienttopicdiversityandlimitedmodel interpretabilityfacedbylargelanguage models （LLMs）in topic modeling，this paper proposeda topic inference model based on retrieval augmented generation（TIM_ RAG），whichachieved topicgenerationanddistribution inferencethrough thre-stagearchitecture.Firstly，intheRAGretrieval phase，itdesignedamulti-dimensionaldocumentsimilarityretrievalmethodthatfilteredsimilardocumentsehibiting bothterm-frequencyrelevanceanddeepsemanticassociations，thus enrichingthetopic informationof individualdocumentand enhancingtopicdiversitySecondly，intheRAGgenerationphase，itimplementedamulti-perspectivetopicgenerationstrategy，whichusedchain-of-thought prompting toguidetheLLMin extracting multi-angletopic termsand generated interediate reasoningsteps toincreaseprocesstransparency.Finally，intheindependent inferencephase，itintroducedanembedding transport plan basedonoptimal transport theoryto modelsemanticrelationships between document-topic andtopic-word，significantlyimproving model interpretabilityExperimentsontheWikiText-103，BBCNews，and2ONewsgroupsdatasetsshow thatTIM_RAG effectivelymitigates topic diversity limitations while enhancing topic modeling performance.

Key words：topicmodel;large language model;retrievalaugmented generation;chain-of-thought；optimaltransporttheory

0 引言

随着信息技术的高速发展，文本数据呈爆炸式增长，如何有效地从文档集合中自动推断潜在的主题结构（即主题推理）已成为文本分析领域的核心挑战。（剩余19939字）