Page 112 - 水利学报2025年第56卷第4期
P. 112
[33] LIUJ,LIUC,ZHOUP,etal.Ischatgptagoodrecommender?apreliminarystudy[EB?OL].(2023 - 04 - 20)
[ 2024 - 10 - 16].https:??arxiv.org?abs?2304.10149.
[34] JIZ,LEEN,FRIESKER,etal.Surveyofhallucinationinnaturallanguagegeneration[J].ACM Computing
Surveys,2023,55(12):1 - 38.
[35] DAIS,SHAON,ZHAOH,etal.Uncoveringchatgpt’scapabilitiesinrecommendersystems[C]??Proceedings
ofthe17thACM ConferenceonRecommenderSystems.2023.
[36] KANGW C,NIJ,MEHTAN,etal.Dollmsunderstanduserpreferences?evaluatingllmsonuserratingpredic
tion [EB?OL].(2023 - 05 - 10)[2024 - 10 - 15].https:??arxiv.org?abs?2305.06474.
[37] LINJ,DAIX,XIY,etal.Howcanrecommendersystemsbenefitfrom largelanguagemodels:Asurvey[EB?
OL ].(2023 - 06 - 09)[2024 - 10 - 15].https:??arxiv.org?abs?2306.05817.
[38] SANMARTIND.KG - RAG: BridgingtheGapBetweenKnowledgeandCreativity[EB?OL].(2024 - 05 - 20)
[2024 - 08 - 16].https:??arxiv.org?abs?2405.12035.
[39] SHAOZ,GONGY,SHENY,etal.Enhancingretrieval - augmentedlargelanguagemodelswithiterativeretrieval -
generationsynergy [EB?OL].(2023 - 05 - 24)[2024 - 08 - 16].https:??arxiv.org?abs?2305.15294.
[40] YAOS,YUD,ZHAOJ,etal.Treeofthoughts:Deliberateproblem solvingwithlargelanguagemodels[C]??
AdvancesinNeuralInformationProcessingSystems ,2024.
[41] ZHANGY,WANGX,LIANGJ,etal.Chain - of - Knowledge:Integratingknowledgereasoningintolargelanguage
modelsbylearningfrom knowledgegraphs [EB?OL].(2024 - 06 - 30)[2024 - 08 - 13].https:??arxiv.org?abs?
2407.00653.
[42] 张殿元,余传明.基于知识增强与多模态融合的医疗视觉问答模型 [J].数据分析 与 知 识 发 现,2024,8
(S1):226 - 239.
[43] EVTIKHIEVM,BOGOMOLOVE,SOKOLOVY,etal.Outofthebleu:howshouldweassessqualityofthe
codegenerationmodels ?[J].JournalofSystemsandSoftware,2023,203:111741.
Multimodalknowledgegraphcollaboratedwithlargemodelfor
decisionrecommendationofwaterprojectsriskresponse
1
1
1
1
1,2
YANGYangrui,PANShifeng,LIUXuemei ,MAWei,FENGLing
(1.SchoolofInformationEngineering,NorthChinaUniversityofWaterResourcesandElectricPower,Zhengzhou 450000,China;
2.CollaborativeInnovationCentreforEfficientUtilizationofWaterResources,Zhengzhou 450000,China)
Abstract:Withthedevelopmentofwaterprojectsandtheincreasingdemandforinformatization,itbecomesin
creasinglycrucialtoeffectivelyprocessanduseinspectiondatafromwaterprojecttomakeriskresponsedecisions.
TheriskinspectiontextsoftheSouth - to - NorthWaterDiversionProjectinvolvenumerousinfluencingfactors,com
plexinterrelationships ,andmulti - sourceheterogeneousdata.Toeffectivelyutilizetheinspectiontextandensure
thetimelinessandeffectivenessoftheengineeringriskresponse ,amethodthatintegratesmultimodalknowledge
graphwithmultimodallargemodel (MLLM)basedoninspectiontextdataisproposed,aimingtoprovidetargeted
decision - makingrecommendationsforrisksintheoperationoftheSouth - to - NorthWaterDiversionProject.
Firstly ,theopen - sourcemultimodallargelanguagemodelVIP - LLaVA,whichfocusesmoreonregionallocations,
ischosenasthedialogmodel.Secondly,adomain - specificmultimodalknowledgegraphisconstructedbasedon
waterprojectinspectiondata.Lastly ,themultimodalknowledgegraph(MMKG)ofthewaterconservancydomain
isusedastheaugmentedknowledgeofthemultimodallargemodel,andtheretrievalandgenerationofcollaborative
anditerativeapproachisadoptedtoidentifythemostpertinentsolutionsandtoformulateriskresponsedecisions.
TheresultsindicatethatthemethodworkswellonBLEU,ROUGE - LandMETEORevaluationindexeswhengen
eratingriskresponsedecisions,andcanformulatecorrespondingresponsedecisionrecommendationstotheopera
tionalrisksintheprojectintime.Therelatedresearchcaneffectivelyimprovethereliabilityofengineeringsafety
maintenancemanagementandreducetheimpactofengineeringrisksonwatersupplysafety.
Keywords:multimodalknowledgegraph;multimodallargemodel;riskresponse;South - to - NorthWaterDiver
sionProject
(责任编辑:耿庆斋)
— 5 3 —
0