GraphGen4Code | A Toolkit for Generating Code Knowledge ...

Knowledge graphs have been proven extremely useful in powering diverse ... applying it to 1.3 million Python files drawn from GitHub, 2,300 Python modules, ... Knowledgegraphshavebeenprovenextremelyusefulinpoweringdiverseapplicationsinsemanticsearchandnaturallanguageunderstanding. Inthiswork,wepresentGraphGen4Code,atoolkittobuildcodeknowledgegraphsthatcansimilarlypowervariousapplicationssuchasprogramsearch,codeunderstanding,bugdetection,andcodeautomation.GraphGen4Codeusesgenerictechniquestocapturecodesemanticswiththekeynodesinthegraphrepresentingclasses,functionsandmethods.Edgesindicatefunctionusage(e.g.,howdataflowsthroughfunctioncalls,asderivedfromprogramanalysisofrealcode),anddocumentationaboutfunctions(e.g.,codedocumentation,usagedocumentation,orforumdiscussionssuchasStackOverflow).OurtoolkitusesnamedgraphsinRDFtomodelgraphsperprogram,orcanoutputgraphsasJSON.Weshowthescalabilityofthetoolkitbyapplyingitto1.3millionPythonfilesdrawnfromGitHub,2,300Pythonmodules,and47millionforumposts.Thisresultsinanintegratedcodegraphwithover2billiontriples.Wemakethetoolkittobuildsuchgraphsaswellasthesampleextractionofthe2billiontriplesgraphpubliclyavailabletothecommunityforuse. Paper:https://arxiv.org/abs/2002.09440 DownloadGraphGen4Codedatasetasnquadsfromhere. TableofContents SampleGraphsGeneratedbyGraphGen4Code Applications AutomatedMachineLearning(AutoML) Recommendationenginefordevelopers BuildingLanguageModelsforCodeUnderstanding Enforcingbestpractices Learningfrombigcode SampleSchema Createyourowngraph ExampleQueries Publications SampleGraphsGeneratedbyGraphGen4Code 1.3MillionPythonProgramsfromGithub TodemonstrateGraphGen4Code’sscalability,webuildgraphsfor1.3millionPythonprograms(whereprogramreferstoasinglePythonscript)onGitHub,eachanalyzedintoitsownseparategraph.Wealsousethetoolkittolinklibrarycallstodocumentationandforumdiscussions,byidentifyingthemostcommonlyusedmodulesincode,andtryingtoconnecttheirclasses,methodsorfunctionstorelevantdocumentationorposts.Forforumposts,weusedinformationretrievaltechniquestoconnectittoitsrelevantmethodsorclasses.Weperformedthislinkingfor257Kclasses,5.8Mmethods,and278Kfunctions,andprocessed47MpostsfromStackOverflowandStackExchange.ThisshowsthefeasibilityofusingtheGraph4CodeGentoolkitforbuildinglarge-scaleknowledgegraphsforcodethatcapturescodesemanticsaswellasnaturallanguageartifcactsaboutcode. Allgraphfilesareavailablehere. Toloadandquerythisdata,pleasefollowtheinstructionshere.Wealsoprovidescriptsforcreatingadockerimagewiththegraphdatabasereadytouse. ETH150kPythonDataset WealsousedGraphGen4CodetoproducegraphsforETH150kPythonDatasetcollectedfromGithub.ETH-150Kdatasethasbeenusedtotrainmodelsforcoderecommendation,typeinferencing,programrepairs,…etc.WeprovidegraphdataforthisdatasetinbothJSONandRDFN-Quadsformats. Schema ThefollowingshowsacodesnippetexampleaswellasahighleveloverviewoftheinformationgeneratedbyGraphGen4Codefromcodeanalysis,StackOverflow,anddocstrings.WeprovidearandomsampleofeachdatasourceinRDFformathere. CodeSnippetExample Dataflowgraphfortherunningexample StackOverflowGraphExample DocstringsGraphExample Publications IfyouuseGraphGen4Codeinyourresearch,pleaseciteourwork: @article{abdelaziz2020codebreaker, title={ADemonstrationofCodeBreaker:AMachineInterpretableKnowledgeGraphforCode}, author={Abdelaziz,IbrahimandSrinivas,KavithaandDolby,JulianandMcCusker,JamesP}, journal={InternationalSemanticWebConference(ISWC)(DemonstrationTrack)}, year={2020} } @article{abdelaziz2021graph4code, title={AToolkitforGeneratingCodeKnowledgeGraphs}, author={Abdelaziz,IbrahimandDolby,JulianandMcCusker,JamesPandSrinivas,Kavitha}, journal={TheEleventhInternationalConferenceonKnowledgeCapture(K-CAP)}, year={2021} } @inproceedings{abdelaziz2022blanca, title={CanMachinesReadCodingManualsYet?--ABenchmarkforBuildingBetterLanguageModelsforCodeUnderstanding}, author={IbrahimAbdelazizandJulianDolbyandJamieMcCuskerandKavithaSrinivas}, booktitle={ProceedingsoftheAAAIConferenceonArtificialIntelligence(AAAI2022)}, year={2022} }