00各類自然語言處理算法快速發(fā)展，在很多任務(wù)上甚至超越人類00各類自然語言處理算法快速發(fā)展，在很多任務(wù)上甚至超越人類4400算法在實(shí)際應(yīng)用中的效果卻不盡如人意00算法在實(shí)際應(yīng)用中的效果卻不盡如人意00不經(jīng)過魯棒性評估會面臨巨大風(fēng)險7700模型對測試數(shù)據(jù)的微小變化非常敏感9900模型對測試數(shù)據(jù)的微小變化非常敏感00模型對測試數(shù)據(jù)的微小變化非常敏感00為什么會這樣？問題1:為什么基于基準(zhǔn)測試集合和常用評價指標(biāo)的模式不能反映上述問題?問題2:深度神經(jīng)網(wǎng)絡(luò)模型到底學(xué)習(xí)到了什么?問題3:現(xiàn)階段自然語言處理算法魯棒性究竟怎么樣?00為什么會這樣？問題1:為什么基于基準(zhǔn)測試集合和常用評價指標(biāo)的模式不能反映上述問題?問題2:深度神經(jīng)網(wǎng)絡(luò)模型到底學(xué)習(xí)到了什么?問題3:現(xiàn)階段自然語言處理算法魯棒性究竟怎么樣?11數(shù)據(jù)集上存在偏置–WINOGRANDEHaveneurallanguagemoverestimatingthetruecapabilitiesofmachinecommonsense?11數(shù)據(jù)集上存在偏置–WINOGRANDE11數(shù)據(jù)集上存在偏置–WINOGRANDE1.RoBERTafine-tunedonasmallsubsetofthedataset.2.Anensembleoflinearclassifiers(logisticregressions)3.Trainedonrandomsubsetsofthedata4.Determinewhethertherepresentationisstronglyindicativeofthecorrectansweroption5.Discardthecorrespondinginstances11數(shù)據(jù)集上存在偏置–WINOGRANDE11數(shù)據(jù)集上存在偏置–WINOGRANDEFuetal.,RethinkingGeneralizationofNeuralModels:ANamedEntityRecognitionCaseStudy，AAAI2020Liuetal.,EXPLAINABOARD:AnExplainableLeaderboard1.基準(zhǔn)集合構(gòu)建時通常存在數(shù)據(jù)偏置2.粗粒度的評測指標(biāo)不能夠全面反映模型特性00為什么會這樣？問題1:為什么基于基準(zhǔn)測試集合和常用評價指標(biāo)的模式不能反映上述問題?問題2:深度神經(jīng)網(wǎng)絡(luò)模型到底學(xué)習(xí)到了什么?問題3:現(xiàn)階段自然語言處理算法魯棒性究竟怎么樣?Sundararajanetal.,Axiomaticattributionfordeepnetworks.2017基于Bert的用戶檢索詞---文章語義匹配模型AttentionheadsexhibitingpatternsAttentionheadscorrespondingtolinguisticphenomenaThebestperformingattentionsheadsofBERTonWSJdependencyparsingoftenexhibitingsimilarCertainattentionheadscorreAttention-basedprobingclassifierdemonstratedthatsubstantialsyntacticinformationcoulHowupweightingaparticulartrainingexample(xi,yi)intheHowthischangeinthemodelparameterswouldtrainingset{(x1,y1),…,((xn,yn)}byε!wouldchangetheinturnaffectthelossofthetestinputlearnedmodelparametersθHanetal.,Explainingblackboxpredictionsandunveiling00問題2：深度神經(jīng)網(wǎng)絡(luò)模型到底學(xué)習(xí)到了什么？非常初步的猜想,大規(guī)模數(shù)據(jù)分析和實(shí)驗(yàn)中3.預(yù)訓(xùn)練語言模型學(xué)習(xí)到了部分復(fù)述(paraphrase)的相似表示覆蓋了人工構(gòu)造的基礎(chǔ)特征,以及人工很難構(gòu)造的特征高階綜合00為什么會這樣？問題1:為什么基于基準(zhǔn)測試集合和常用評價指標(biāo)的模式不能反映上述問題?問題2:深度神經(jīng)網(wǎng)絡(luò)模型到底學(xué)習(xí)到了什么?問題3:現(xiàn)階段自然語言處理算法魯棒性究竟怎么樣?Hauseretal.,BERTisRobust!ACaseAgainstSynonym-BasedAdversarialExamplesinTextClassificatiSietal.,BenchmarkingRobustneSietal.,BenchmarkingRobustnessofMachineReadingCompr….@AmericanAirservicewas完備性-20種通用變形、60種任務(wù)特有變形、數(shù)千種變形組合可接受-所有變形基于語言學(xué)知識分析功能-對評測結(jié)果給出可視化分析報告“HelovesNLP”istransformedint拼寫錯誤反義詞“HewasborninChina”“HewasborninLlanfairpwllgwyngyllgogeryc看“看看,”“看一看,”“看了看,”and“看了一看.”“Thereisanappleonthedesk”“Thereisanimponderableonthedesk”原始集合Shebecameanurseandworkedinahospital.ItoldJohntocomeearly,buthefailed.TheriverderivesfromsouthernAmerica.Marrywouldliketoteachkidsinthekindergarten.Thestormdestroyedmanyhousesinthevillage.√√√?Plausibility(Lambertetal.,2010)measureswhetherthetextisreasonableandwrittenbynativespeakers.Sentencesordocumentsthatarenatural,appropriate,logicallycorrect,andmeaningfulinthecontextwillreceiveahigherplausibilityscore.Textsthatarelogicallyorsemanticallyinconsistentorcontaininappropriatevocabularywillreceivealowerplausibilityscore.?Grammaticality(Newmeyer,1983)measureswhetherthetextcontainssyntaxerrors.Itreferstotheconformityofthetexttotherulesdefinedbythespecificgrammarofalanguage.Gui,Tao,etal."Textflint:UnifiedmultilingualrobustnessevaluationtoELMO+BiLSTM+CRFBert+BiLSTM+CRFUANETS-LSTM meta-taggerLM-LSTM-CRFBILSTM-LAN BiLSTM-aux CRF++CNN+BiLSTM+CRF xlnet-large-cased xlnet-base-cased roberta-largeroberta-basebert-large-uncasedbert-base-uncased albert-xxlarge-v2albert-base-v2Match-lstm LCF-BERTBERT-SPCBERT-BASEAEN-BERTTNetTD-LSTM MGANLSTMATAE-LSTMMemNetAEN-GloveParadigmaticRelationMorphologySyntaxPragmaticsParadigmaticRelationMorphologyABSAABSAContractionKeyboardOcrSpellingErrorTyposWordCase-lowerWordCase-titleWordCase-upperInsertAdvSwapNamedEntAddPuncAppendIrrTwitterTypeMLMSuggestionSwapNumSwapSyn-WordNetContractionKeyboardOcrSpellingErrorTyposWordCase-lowerWordCase-titleWordCase-upperInsertAdvSwapNamedEntAddPuncAppendIrrTwitterTypeMLMSuggestionSwapNumSwapSyn-WordNetSwapSyn-WordEmbeddingSwapSyn-WordEmbeddingMRC-SQuAD2.0MRC-SQuAD1.1SM-qqpSM-mrpcNLI-SNLINLI-MNLI-mmNLI-MNLI-mCOREF-OntonotesABSA-SemEval2014-RestaurantABSA-SemEval2014-LaptopSA-Yelp-BinarySA-IMDBDP-PTBPOS-WSJParadigmaticRelationMorphologySyntaxPragmaticsParadigmaticRelationMorphologyContractionKeyboardOcrSpellingErrorWordCase-lowerWordCase-titleWordCase-upperInsertAdvSwapNamedEntAddPuncAppendIrrTwitterTypeMLMSuggestionSwapNumSwapSyn-WordEmbeddingSwapSyn-WordNetContractionKeyboardOcrSpellingErrorWordCase-lowerWordCase-titleWordCase-upperInsertAdvSwapNamedEntAddPuncAppendIrrTwitterTypeMLMSuggestionSwapNumSwapSyn-WordEmbeddingSwapSyn-WordNetGlobalSemanticsGlobalSemanti