Commonsense AIMyth and Truth原版完整文件.pptx

《Commonsense AIMyth and Truth原版完整文件.pptx》由会员分享，可在线阅读，更多相关《Commonsense AIMyth and Truth原版完整文件.pptx（77页珍藏版）》请在得力文库 - 分享文档赚钱的网站上搜索。

1、Commonsense AI:Myth and Truthcheeseburger stabbing redux,2021 edition YejinChoiPaul G.Allen School of Computer Science&Engineering University of Washington&Allen Institute for Artificial IntelligenceCommonsense AI:Myth and Truthcheeseburger stabbing redux,2021 edition YejinChoiPaul G.Allen School of

2、 Computer Science&Engineering University of Washington&Allen Institute for Artificial IntelligenceYear 2020ACL Commonsense TutorialVered ShwartzMaarten SapAntoine BosselutYejin ChoiDan Roth 2nd most popular(among 8 tutorials)https:/homes.cs.washington.edu/msap/acl2020-commonsense/5k=#of the main con

3、ference registration1.3k=#of our tutorial registration3.4k=#of view counts on our recorded lecturesUW=EPFLUW=TBDAI2=TBDCirca 2017Shouldnt work on it since it wont workDont even say that wordThats a research topic of 70s&80sCommonsense?What is commonsense?Commonsense AI is an impossible goal(ever)Tha

4、ts a research topic of 70s and 80sMaybe only possible in the faraway futureIts what everyone knows and agrees on?language is irrelevant to commonsense?Too hard to define precisely,thus shouldnt work on itShould we or should we not?Truth or Myth?2017 2020Commonly held beliefsKnowledge and reasoning a

5、re distinct and exclusiveLanguage is in the way of reasoning;lets do formal logicsLanguage is not symbols.Words and numbers are,but not language at largeHumans acquire commonsense completely un-/self-supervised,thus so should machinesWhat is commonsense?Should we or should we not?Commonsense AI is a

6、n impossible goal(ever)Thats a research topic of 70s and 80sMaybe only possible in the faraway futureIts what everyone knows and agrees on?language is irrelevant to commonsense?Too hard to define precisely,thus shouldnt work on itTruth or Myth?The Curious Case of Cheeseburger Stabbing An example rep

7、eatedly appeared my talks between Mar 2017 and May 2018 The Curious Case of“Cheeseburger Stabbing”Someone stabbed a cheeseburger?A cheeseburger stabbed someone?A cheeseburger stabbed another cheeseburger?Someone stabbed someone else over a cheeseburger?The Curious Case of“Cheeseburger Stabbing”what

8、is said+what is not saidRPehaydsiincgalbCetowmeemnotnheselinnsees:not possible to stab using a burgerUnderstandSiongcial Commonsense:stabbing someone is badSomeone stabbed a cheeseburger?A cheeseburger stabbed someone?A cheeseburger stabbed another cheeseburger?Someone stabbed someone else over a ch

9、eeseburger?2018:“Stabbing is a crime punishable by cheeseburger”2020:“stabbing of a cheeseburger”per GPT-3The Curious Case of“Cheeseburger Stabbing”2021:Stay tuned for the 2021 edition!Path to commonsense?Brute force larger networks with deeper layers?You dont reach to the moonby making the tallest

10、building in the world tallerPath to commonsense?ObligatoryControversial Remarksof the Day1.the continuum betweenknowledge and reasoning2.the interplay between reasoningand language generation3.the blend between neural vssymbolic representationKnowledgeReasoningLanguageNeuralSymbolicInduction from sp

11、ecific to generalDeduction from general to specificAbduction “why something happened”(or reasoning about the the probable explanation)Counterfactual “what if something else happened”ReasoningLanguagePeirce 1965Unsupervised Backprop-based Decoding for Counterfactual and Abductive Commonsense Reasonin

12、gEMNLP 2020Antoine BosselutMeRonan LeBrasChandra BhagavatulaDELOREANBack to the Future:Peter WestJena HwangLianhui QinVered ShwartzWhat happened in between?Abductive Reasoning(Bhagavatula et al.,2019)Ray ran to his daughter to make sure she was okay.Future ObservationRay hung a tire on a rope to mak

13、e his daughter a swing.Past ObservationRay hung a tire on a rope to make his daughter a swing.Ray ran to his daughter to make sure she was okay.Abductive Reasoning(Bhagavatula et al.,2019)Past ObservationFuture ObservationShe hit the rope and the tire fell on top of her.HypothesisAbduction “why some

14、thing happened”(or reasoning about the the probable explanation)Peirce 1965Counterfactual Reasoning(Qin et al.,2019)Ray hung a tire on a rope to make his daughter a swing.Future ObservationRay ran to his daughter to make sure she was okay.Past ObservationAbductive ReasoningShe hit the rope and the t

15、ire fell on top of her.(Bhagavatula et al.,2019)HypothesisZeke was throwing a party.All his friends were dressing up for this Halloween party.Zeke thought about being a vampire or a wizard.Then he decided on a scarier costume.Zeke dressed up like a skeleton.What if this is a Game of Thrones themed p

16、arty instead of a Halloween party?Story context changesAn example from the“TimeTravel”dataset(Qin et al.,EMNLP 2019)Zeke was throwing a party.All his friends were dressing up for this Halloween party.Zeke thought about being a vampire or a wizard.Then he decided on a scarier costume.Zeke dressed up

17、like a skeleton.What if this is a Game of Thrones themed party instead of a Halloween party?Story context changesAn example from the“TimeTravel”dataset(Qin et al.,EMNLP 2019)Story ending doesnt make sense nowZeke was throwing a party.All his friends were dressing up for this Halloween party.Zeke tho

18、ught about being a vampire or a wizard.Then he decided on a scarier costume.Zeke dressed up like a skeleton.Story context changesWhat if this is a Game of Thrones themed party instead of a Halloween party?Counterfactual Reasoning(Qin et al.,2019)Reasoning about the alternative futurebased on counter

19、factual past.An example from the“TimeTravel”dataset(Qin et al.,EMNLP 2019)Story ending doesnt make sense nowZeke was throwing a party.All his friends were dressing up for this Halloween party.Zeke thought about being a vampire or a wizard.Then he decided on a scarier costume.Zeke dressed up like a s

20、keleton.What if this is a Game of Thrones themed party instead of a Halloween party?Zeke was throwing a party.All his friends were dressing up for this Halloween party.Zeke thought about Lannister,but he didnt want to look like a Lannister.He wanted to look like a Stark.Zeke dressed up like a Stark.

21、An example from the“TimeTravel”dataset(Qin et al.,EMNLP 2019)Story ending doesnt make sense nowStory context changesConsistentnow!Only do minimal edit!Zeke thought about being a vampire or a wizard.Then he decided on a scarier costume.Zeke dressed up like a skeleton.Zeke thought about Lannister,but

22、he didnt want to look like a Lannister.He wanted to look like a Stark.Zeke dressed up like a Stark.All his friends were dressing up for this Game of Thrones themed party.Story ContextZeke was throwing a party.All his friends were dressing up for this Halloween party.Rewritten EndingOriginal EndingRa

23、y hung a tire on a rope to make his daughter a swing.Future ObservationRay ran to his daughter to make sure she was okay.Past ObservationAbductive ReasoningShe hit the rope and the tire fell on top of her.(Bhagavatula et al.,2019)HypothesisCounterfactual Reasoning(Qin et al.,2019)Abductive Reasoning

24、(Bhagavatula et al.,2019)Both involve nonmonotonic reasoning with past context X and future constraint ZInput:ZOutput:Input:XCounterfactual Reasoning(Qin et al.,2019)Pretrained Language Models are successful on many tasksHow are Pretrained LMs on the Nonmonotonic Reasoning?ZPre-trained GPT2Ray ran t

25、o his daughter to make sure she was okay.XRay hung a tire on a rope to make his daughter a swing.Its a big swing.The little girl liked it and was thrilled at it.YLets first see the abductive caseNot able to do right to left!Ray ran to his daughter to make sure she was okay.Z SXRay hung a tire on a r

26、ope to make his daughter a swing.But the swing didnt go off,so they moved down the slope towards the lake.Doesnt make sense!YWhy not just concatenate both direction?Pre-trained GPT2Ray ran to his daughter to make sure she was okay.Z SXRay hung a tire on a rope to make his daughter a swing.As the swi

27、ng moved,the girls cries sounded in his ears.Doesnt make sense!Try again?Pre-trained GPT2YZRay ran to his daughter to make sure she was okay.XRay hung a tire on a rope to make his daughter a swing.Something might have been missing hereBackpropagation!Pre-trained GPT2YStyle-Loss(Y,Styled Image)ConvNe

28、tBackpropagation!YInspired by“Image Style Transfer”(Gatys et al,2016)Styled ImageZRay ran to his daughter to make sure she was okay.XRay hung a tire on a rope to make his daughter a swing.YBack to our case Backpropagation!Coherence-Loss(XY,Z)Pre-trained GPT2DELOREAN(DEcoding for nonmonotonic LOgical

29、 REAsoNing)Ray ran to his daughter to make sure she was okay.Zx1x2xNXRay hung a tire on a rope to make his daughter a swing.XYy 2y Ny 1InitializationJust as how you do regular decodingDELOREANLMRay hung a tire on a rope to make his daughter a swing.XRay ran to his daughter to make sure she was okay.

30、ZYRayrantoSz1z2z3zNZy Ny 2y 1Backward PassBackpropagate future informationDELOREANLMRay hung a tire on a rope to make his daughter a swing.XRay ran to his daughter to make sure she was okay.ZYRayrantoSzNZz1z2z3y b1y b2y bNy 1y 2y NDELOREANBackward PassBackpropagate future information Loss(Z|X,Y)LMBa

31、ckpropagationL(X,Y,Z):=艺NZn=1nlog PLM(z|X,Y,Z1:n1)Ray hung a tire on a rope to make his daughter a swing.XRay ran to his daughter to make sure she was okay.ZYBackpropagationL(X,Y,Z):=艺NZlog PL M(zn|X,Y,Z1:n1)n=1RayrantoSz1z2z3zNZy b1y b2y bNForward PassMix both past and future informationDELOREANy N

32、y 2y 1LMx1x2xNXRay hung a tire on a rope to make his daughter a swing.XRay ran to his daughter to make sure she was okay.ZYBackpropagationL(X,Y,Z):=艺NZlog PL M(zn|X,Y,Z1:n1)n=1RayrantoSz1z2z3zNZy b1y b2y bNy Ny 2y 1y f1y f2y fNForward PassMix both past and future informationDELOREANLMx1x2xNXRay hung

33、 a tire on a rope to make his daughter a swing.XRay ran to his daughter to make sure she was okay.ZRayrantoSzNZz1z2z3Backpropagationy b1y b2y bNy Ny 2y 1y f1y f2y fNYLMDELOREANRepeatT timesL(X,Y,Z):=艺NZn=1n1:n1log PLM(z|X,Y,Z)x1x2xNXRay hung a tire on a rope to make his daughter a swing.XRay ran to

34、his daughter to make sure she was okay.ZRayrantoSzNZz1z2z3Backpropagationy b1y by b2Ny Ny 2y 1y f1y f2y fNOutput:SYhe hit the rope and the tire fell on top of her.LMDELOREANSamplingL(X,Y,Z):=艺NZn=1log PLM(zn|X,Y,Z1:n1)x1x2xNXXZYzNZz1z2z3Backpropagationy b1y b2y bNy Ny 2y 1Counterfactual Reasoning?y

35、f1y f2y fNLMDELOREAN?x1x2xNXXZYzNZz1z2z3Backpropagationy b1y b2y bNy Ny 2y 1y f1y f2y fNLMDELOREANCounterfactual Reasoning?Distance-Loss:(Y,Z)L(X,Y,Z):=KL?(Zl/softmax(Y/)x1x2xNXYzNZz1z2z3Backpropagationy b1y b2y bNy Ny 2y 1y f1y f2y fNLMDELOREANCounterfactual Reasoning?Distance-Loss:(Y,Z)L(X,Y,Z):=K

36、L(Zl/softmax(Y/)Zeke was throwing a party.Counterfactual All his friends were dressing up for this Game of Thrones themed party.XZeke thought about beinga vampire or a wizard.Thenhe decided on a scariercostume.Zeke dressed uplike a skeleton.ZZekethoughtaboutSx1x2xNXYzNZz1z2z3Backpropagation12y by by

37、 bNy Ny 2y 1y f1y f2y fNLMDELOREANCounterfactual Reasoning?Distance-Loss:(Y,Z)L(X,Y,Z):=KL(Zl/softmax(Y/)Zeke was throwing a party.Counterfactual All his friends were dressing up for this Game of Thrones themed party.XZeke thought about beinga vampire or a wizard.Thenhe decided on a scariercostume.Z

38、eke dressed uplike a skeleton.ZZekethoughtaboutSZeke thought about Lannister,but he didnt want to look like a Lannister.He wanted to look like a Stark.Zeke dressed up like a Stark.Abductive ReasoningCoherence02.557.510Coherece(X,Y,Z)7.832.972.862.36Zeroshot+RankingSupervised+COMeT(Bhagavatula et al.

39、,2019)DELOREANHumanEven better than supervised methodHuman Evaluation ResultsAutomatic Evaluation ResultsZeroshot+RankingSupervised+COMeTDELOREANHuman03570105140BLEU4ROUGE-LBERTSCOREPlease check the paper for more baselines Outperforms unsupervised baselines over all metrics.Reasoning as generative

40、tasksAs opposed to discriminative tasks(i.e.,categorization)Because the space of reasoning in language is infiniteInductionDeductionAbductionCounterfactual“thinking out loud”We often think as we speak,on the fly,word-by-word without enumerating all possible alternative sentencesReasoningLanguageReas

41、oning serves the purpose of communicationNEUROLOGIC DECODINGYejin ChoiChandra BhagavatulaRowan ZellersRonan LeBrasPeter West(Un)supervised Neural Text Generation with Predicate LogicConstraintsNAACL 2021Ximing LuSeq2SeqMachine TranslationThe physician told the baker that she had cancer.Der Arzt sagt

42、e dem Bckerin,dass er Krebs habe.XDialogue ResponseThere are 182 hotels if you do not care whether dogs are allowed.typehotelcount182dogs alloweddont careX food,table,sit,front The man sat with his food at the front of the table.XCOMMONGENImage CaptioningMan in blue wetsuit is surfing on wave.XXLang

43、uage ModelXFine-tuned Language Model board,lose,ride,fall,balance COMMONGEN(Lin et al.EMNLP 2020)missing keyword lose,ride A man is trying to keep his balance as he falls off a board.should use all givenkeywords.XLanguage ModelYDecodingLogical Constraint in CNF formNEUROLOGIC DECODINGC1CmD1 _ D2 _ D

44、i Dk _ Dk+1 _ DlAdvanced beam search with diverse partial solutions of CNF in consideration offour dynamic states of clauses:reversible unsatisfaction irreversible unsatisfaction reversible satisfaction irreversible satisfactionCOMMONGEN(Zero-shot)ROUGE-L33.0044.0041.2538.5035.75distillbase medium l

45、argeXLbeam search(supervised)NeuroLogic(supervised)NeuroLogic(zero-shot)METEOR20.0031.0028.2525.5022.75distillbase medium largeXLCoverage60.00100.0090.0080.0070.00distillbase medium largeXLUnsupervised NeuroLogicoutperformssupervised approachesUnsupervised NeuroLogic on smallernetworks outperformssu

46、pervised approaches on larger networks!Path to commonsense?ObligatoryControversial Remarksof the Day1.the continuum betweenknowledge and reasoning2.the interplay between reasoningand language generation3.the blend between neural vssymbolic representationKnowledgeReasoningLanguageNeuralSymbolicPath t

47、o commonsense?ObligatoryControversial Remarksof the Day1.the continuum betweenknowledge and reasoning2.the interplay between reasoningand language generation3.the blend between neural vssymbolic representationKnowledgeReasoningLanguageNeuralSymbolic(COMET-):On Symbolic and Neural Commonsense Knowled

48、ge Graphs wait,doesnt GPT-3 know everything?AAAI 2021Jeff DaRonan Le BrasJena HwangMeKeisuke SakaguchiAntoine BosseultChandra BhagavatulaATOMIC2020Symbolic commonsense knowledge graphNeural commonsense modelLanguage models!=knowledge modelsCall Uber for a ridePay the billAs a result,X wants toMainta

49、in their carBecause X wanted to ATOMICA mechanicMoneyBefore,X needs Fold into origamiPaying repairsUsed forby workingHas propertyPaperIs made ofUsed forPhysical-Entity CommonsenseEarnedSocial-Interaction CommonsenseThe car costs too muchXs car is totaled completelyCan behindered byX gets Xs car repa

50、iredX drives an old carHappens afterX spends a fortuneHappens beforeX likes driving nowEvent-Centered CommonsenseATOMIC2020ATOMIC2020Call Uber for a ridePay the billAs a result,X wants toMaintain their carBecause X wanted to A mechanicMoneyBefore,X needs Fold into origamiPaying repairsUsed forEarned