知识图谱论文资源1911.07132.pdfVIP

Interstellar:SearchingRecurrentArchitecturefor

KnowledgeGraphEmbedding

YongqiZhangQuanmingYaoLeiChen

4ParadigmInc.

DepartmentofElectronicEngineering,TsinghuaUniversity

1DepartmentofComputerScienceandEngineering,HKUST

2

0{zhangyongqi,yaoquanming}@4,leichen@cse.ust.hk

2

r

pAbstract

A

8Knowledgegraph(KG)embeddingiswell-knowninlearningrepresentationsof

2KGs.Manymodelshavebeenproposedtolearntheinteractionsbetweenenti-

tiesandrelationsofthetriplets.However,long-terminformationamongmultiple

]tripletsisalsoimportanttoKG.Inthiswork,basedontherelationalpaths,which

Garecomposedofasequenceoftriplets,wedefinetheInterstellarasarecurrent

Lneuralarchitecturesearchproblemfortheshort-termandlong-terminformation

s.alongthepaths.First,weanalyzethedifficultyofusingaunifiedmodeltowork

castheInterstellar.Then,weproposetosearchforrecurrentarchitectureasthe

[InterstellarfordifferentKGtasks.Acasestudyonsyntheticdataillustratesthe

importanceofthedefinedsearchproblem.Experimentsonrealdatasetsdemon-

3

vstratetheeffectivenessofthesearchedmodelsandtheefficiencyoftheproposed

2hybrid-s