程序自動生成技術(shù)之研究動態(tài)代碼大數(shù)據(jù)的形成引發(fā)了

“利用表示學習模型解決開發(fā)問題”的研究海量高質(zhì)量開源代碼解決方案編程任務表示模型學習重要國際會議相關(guān)論文的數(shù)量不斷攀升分布：54%來源于人工智能領(lǐng)域，46%來自于軟件工程領(lǐng)域相關(guān)領(lǐng)域的研究雖然，相關(guān)領(lǐng)域（特別是自然語言處理）的研究已經(jīng)取得了重要進展，但卻無法直接應用于代碼大數(shù)據(jù)的分析。？程序代碼數(shù)據(jù)的特性抽象性

——對比自然語言，代碼的淺層信息難以體現(xiàn)其操作語義。結(jié)構(gòu)性

——代碼的結(jié)構(gòu)（如AST的復雜度）遠比自然語言要復雜。演化性

——代碼的迭代比自然語言頻繁，演化速度比自然語言快。代碼的自動生成面臨諸多挑戰(zhàn)挑戰(zhàn)之一：

代碼詞匯表與自然語言詞匯表存在巨大差異代碼詞匯表中詞匯數(shù)量通常比自然語言多很多；代碼中詞匯的語義具有較強的區(qū)域性，且與代碼結(jié)構(gòu)密切相關(guān)；代碼中自定義標識符具有較強的動態(tài)性，容易發(fā)生變化；代碼的自動生成面臨諸多挑戰(zhàn)挑戰(zhàn)之二：

代碼之間的語義依賴關(guān)系不同于自然語言代碼語句之間的依賴關(guān)系與代碼結(jié)構(gòu)或體系結(jié)構(gòu)相關(guān)，復雜程度較高；與自然語言相比，代碼語句之間存在“較遠”的語義依賴關(guān)系；代碼的自動生成面臨諸多挑戰(zhàn)挑戰(zhàn)之三：

代碼與背景知識的綁定關(guān)系較強，且演化快代碼的語義與所使用的API庫、框架等可復用資源密切相關(guān)，存在非常強的綁定關(guān)系；API庫、可復用框架的演化對代碼語義的影響非常大；代碼的語義對其具體運行環(huán)境也存在較大的依賴；代碼的自動生成面臨諸多挑戰(zhàn)其他挑戰(zhàn)：需求是軟件開發(fā)的起點，代碼的語義必須滿足需求才具有實際意義；與自然語言相比，代碼具有無二義性，且需要通過編譯，并能夠執(zhí)行，因此，更加難以生成；開發(fā)者對代碼的要求不僅僅是能夠執(zhí)行，代碼的優(yōu)劣對最終解決方案的效果，存在較大的影響；相關(guān)研究團隊PublicationsAreDeepNeuralNetworkstheBestChoiceforModelingSourceCode,

FSE’17Onthe“Naturalness”ofBuggyCode,

ICSE’16CACHECA:ACacheLanguageModelBasedCodeSuggestionTool(Demo),

ICSE’15OntheLocalnessofSoftware,

FSE’14

Onthe“Naturalness”ofsoftware,

ICSE’12

PremDevanbuProfessorUniversityofCaliforniaatDavisThe

Naturalness

of

Source

Code相關(guān)研究團隊相關(guān)研究團隊PublicationsSmartPaste:LearningtoAdaptSourceCode.TechReport,2017AConvolutionalAttentionNetworkforExtremeSummarizationofSourceCode.ICML’16MiningSemanticLoopIdiomsfromBigCode.TechReport,2016BimodalModellingofSourceCodeandNaturalLanguage.ICML’15SuggestingAccurateMethodandClassNames.FSE’15LearningNaturalCodingConventions.FSE’14Code

ConventionsEdinburgh相關(guān)研究團隊PublicationsAPICodeRecommendationUsingStatisticalLearningfromFine-grainedChanges,FSE’16T2API:SynthesizingAPICodeUsageTemplatesfromEnglishTextswithStatisticalTranslation,FSE’16UsingTopicModeltoSuggestFine-grainedSourceCodeChanges,ICSME’16DoContextsHelpinPhrase-based,StatisticalSourceCodeMigration?,ICSME’16MappingAPIElementsforCodeMigrationwithVectorRepresentations,ICSE’16TienN.Nguyen,Ph.D.

AssociateProfessorIowaStateUniversityResearchInterests:ProgramAnalysisandUnderstanding,SoftwareEvolutionandMaintenance,MiningSoftwareRepositories,VersionandConfigurationManagement,andWebandConfigurableCodeAnalysis.API相關(guān)研究團隊PublicationsDeepAM:MigrateAPIswithMulti-modalSequencetoSequenceLearning,IJCAI’17DeepAPILearning,FSE’16SungKimAssociateProfessorHongKongUniversityofScienceandTechnologyZhiJinGeLiLiliMou*PekingUniversity*PostdoctoralfellowUniversityofWaterlooPublicationsLearningToInferAPIMappingsFromAPIDocuments,KSEM2017.ConvolutionalNeuralNetworksoverTreeStructuresforProgrammingLanguageProcessing,AAAI2016.LearningEmbeddingsOfAPITokensToFacilitateDeepLearningBasedProgramProcessing,KSEM2016.AComparativeStudyonRegularizationStrategiesforEmbedding-basedNeuralNetworks,EMNLP2015.Onend-to-endprogramgenerationfromuserintentionbydeepneuralnetworks,arxiv’15BuildingProgramVectorRepresentationsforDeepLearning,arxiv’14,KSEM2015.ScottReed&NandodeFreitas,NEURALPROGRAMMER-INTERPRETERS,CandidatePaperof5thInternationalConferenceonLearningRepresentations(ICLR2017),underreview.QuchengGongandYuandongTianandC.LawrenceZitnick,UNSUPERVISEDPROGRAMINDUCTIONWITHHIERARCHICALGENERATIVECONVOLUTIONALNEURALNETWORKS,CandidatePaperof5thInternationalConferenceonLearningRepresentations(ICLR2017),underreview.RecentlySearch-basedSynthesisIR

based

Search

&

SynthesisGeneration

based

Search

&

SynthesisLanguage-basedDeductionTranslate-basedgenerationProbability-baseddeductionLearning-basedSimulation I-ObasedNaturalStyleLearningOperationProcessorientedLearningCodeGenerationGoogleDeepMind，2016NeuralProgrammer-InterpretersMicrosoftResearch,2017DeepCoder:LearningtoWriteProgramsFacebookAIResearch,2017HGCNN:HierarchicalGenerativeConvolutionalNeuralNetworksNeural

ProgrammingbyExample(NPBE)Microsoft——DEEPCODER:LEARNINGTOWRITEPROGRAMS

Facebook——HierarchicalGenerative

ConvolutionalNeuralNetworks(HGCNN)對比我們在2015年的工作…深度神經(jīng)網(wǎng)絡程序生成模型深度神經(jīng)網(wǎng)絡程序生成模型訓練生成深度神經(jīng)網(wǎng)絡程序生成模型訓練生成深度神經(jīng)網(wǎng)絡程序生成模型深度神經(jīng)網(wǎng)絡程序生成模型深度神經(jīng)網(wǎng)絡程序生成模型Our

recent

workOur

recent

work35Isitpossibletobridgethegap

betweenNaturalLanguageTextsandPrograms?IntelligentProgrammingHelperProgramunderstandingandFunctionLearningNaturalLanguageProcessingandSemanticsLearningNaturalLanguageText—ProgramMatchingEngineIntelligentProgrammingHelperProgramunderstandingandFunctionLearningNaturalLanguageProcessingandSemanticsLearningNaturalLanguageText—ProgramMatchingEngineIsitpossibletobridgethegap

betweenNaturalLanguageTextsandPrograms?36Thisfunctionimplementsthebubblealgorithmthatsortsanarrayinascendingorder.Wecouldfindamapbetweenasnippetofprogramcodeandapieceofnaturallanguagedescription.KeyProblemsofLearningbasedCodeGenerationHow

torepresenttheprogram?Canweencodethelogicinformationintheprogram?Howtousetheprogramknowledgeincodegeneration?Howtorepresenttheknowledgeusedincodingperiod?Howtodesigntheneuralmodelforprogramlearning?Istheprobabilistic

model

enoughforprogram?Howtodealwiththedefectsofneuralnetworkmodel?Probabilisticdependencyproblemsoverlongdistances.回應一個問題VincentJ.Hellendoorn,PremkumarT.Devanbu:

Aredeepneuralnetworksthebestchoiceformodelingsourcecode?ESEC/SIGSOFTFSE2017:763-773謝謝！ChallengetoAIThetruechallengetoartificialintelligenceprovedtobesolvingthetasksthatareeasyforpeopletoperformbuthardforpeopletodescribeformally;solvingtheproblemsthatwesolveintuitively,thatfeelautomatic;likerecognizingspokenwordsorfacesinimages;Thefigureshowstwoofthethreehistoricalwavesofartificialneuralnetsresearch,asmeasuredbythefrequencyofthephrases“cybernetics”and“connectionism”or“neuralnetworks”accordingtoGoogleBooksKey

Trends

about

Deep

LearningKey

Trends

about

Deep

Learning1.

Deeplearninghasbecomemoreusefulastheamount