融合知识图谱的大语言模型研究综述

打开文本图片集

键词：大语言模型；知识图谱；可解释性；幻觉问题；知识语言协同

中图分类号：TP391.1 文献标志码：A 文章编号：1001-3695（2025）08-002-2255-12

doi：10.19734/j.issn.1001-3695.2024.12.0532

Review of large language models integrating knowledge graph

Cao Rongrong'，Liu Lin1†，Yu Yandong²，Wang Hailong1 （1.CollegeofpuerSiece&TchoogrogolaNalUiestHoo2ina;2.UanqabKLbatofI telligent InformationProcessing&Security，JiningNormal University，UlanqabNei MongolO12ooo，China）

Abstract：LLMshavedemonstrated exceptional performanceacrossmultipleverticaldomains，yettheir practicaldeployment remainsconstrainedbylimitedexplainabilityandhallcinationisues ingeneratedcontent.KGs，whichstorefactualknowledgein structured semantic networks，providea novel pathway toenhance thecontrolabilityand knowledgeconstraintsof LLMs.To addressthesechallenges，this paper systematicallyreviewed technical approaches for integrating KGs withLLMs.It analyzedrepresentative methodsacross threekeystages—pretrainingadaptation，architectural modification，andfine-tuning optimization，andsummarized their mechanisms for improvingmodel explainabilityandsuppressinghallucinations.Furthermore，itidentifiedcorechalengessuchasmultimodalknowledgerepresentationalignmentandlatencyindynamicknowledge integration.Theanalysisrevealsthatdeep integrationofKGssignificantlyenhancesthefactualconsistencyofLLM-generated content.However，futureresearchmustovercomecriticaltechnicalbotlenecksinmultimodal knowledgealignment，lightweight incrementalfusion，andcomplexreasoning verification toshiftLLMsfromlanguage-centrictoknowledge-language-augmented paradigms，therebyestablishing theoreticalandtechnicalfoundationsforbuilding trustworthyandinterpretableAIsystems.

Key words：large languagemodels（LLM）；knowledgegraph（KG）；explainability；halucinationproblem；knowledgelanguage synergy

0 引言

大语言模型（LLM）[1]作为人工智能领域的前沿技术，特指基于超大规模语料库训练，参数量超过百亿级别的深度神经网络模型。（剩余37680字）