97
Lecture 16: Dialogue Alan Ri4er (many slides from Greg Durrett)

Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

  • Upload
    others

  • View
    0

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Lecture16:Dialogue

AlanRi4er(many slides from Greg Durrett)

Page 2: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

ThisLecture

‣ Chatbotdialoguesystems

‣ Task-orienteddialogue

‣ OtherdialogueapplicaAons

Page 3: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Chatbots

Page 4: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

TuringTest(1950)‣ ImitaAongame:AandBarelockedinroomsandanswerC’squesAonsviatypewriter.BotharetryingtoactlikeB

Page 5: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

TuringTest(1950)‣ ImitaAongame:AandBarelockedinroomsandanswerC’squesAonsviatypewriter.BotharetryingtoactlikeB

OriginalInterpretaAon:

Page 6: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

TuringTest(1950)‣ ImitaAongame:AandBarelockedinroomsandanswerC’squesAonsviatypewriter.BotharetryingtoactlikeB

A B

C

B B

trainedjudge

OriginalInterpretaAon:

Page 7: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

TuringTest(1950)‣ ImitaAongame:AandBarelockedinroomsandanswerC’squesAonsviatypewriter.BotharetryingtoactlikeB

A B

C

B B

trainedjudge

OriginalInterpretaAon:

Page 8: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

TuringTest(1950)‣ ImitaAongame:AandBarelockedinroomsandanswerC’squesAonsviatypewriter.BotharetryingtoactlikeB

A B

C

B B

trainedjudgeC trainedjudge

OriginalInterpretaAon: StandardInterpretaAon:

Page 9: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

TuringTest(1950)‣ ImitaAongame:AandBarelockedinroomsandanswerC’squesAonsviatypewriter.BotharetryingtoactlikeB

A B

C

B B

trainedjudgeC trainedjudge

OriginalInterpretaAon: StandardInterpretaAon:

‣ Thetestisnot“doesthiscomputerseemhuman-liketorandompeoplewithawebbrowser?”

Page 10: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

ELIZA‣ Created1964-1966atMIT,heavilyscripted

‣ DOCTORscriptwasmostsuccessful:repeatsuser’sinput,asksinanequesAons

Weizenbaum(1966)

Page 11: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

ELIZA

‣ IdenAfykeyword,idenAfycontext,applytransformaAonrule

Page 12: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

ELIZA

(.*)you(.*)me

WhydoyouthinkI$2you?

‣ IdenAfykeyword,idenAfycontext,applytransformaAonrule

Page 13: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

ELIZA

(.*)you(.*)me

WhydoyouthinkI$2you?

‣ IdenAfykeyword,idenAfycontext,applytransformaAonrule

My(.)(.*)

Whatelsecomestomindwhenyou thinkofyour$1?

Page 14: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

ELIZA

(.*)you(.*)me

WhydoyouthinkI$2you?

‣ IdenAfykeyword,idenAfycontext,applytransformaAonrule

‣ Veryli4leneedtogeneratenewcontent,butcanonlyhaveonetypeofconversaAon

My(.)(.*)

Whatelsecomestomindwhenyou thinkofyour$1?

Page 15: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Cleverbot‣ Carpenter(1986),onlinesystembuiltin2006

User:Hitherecleverbot.Cleverbot:Howareyou?User:Doingokay.whatplansdoyouhavetoday?Cleverbot:Idorevision.User:Whatareyourevisingfor?Cleverbot:Mathshistory.User:Howoldareyou?Cleverbot:That'snoneofyourbusiness.User:Whatismathshistory?Cleverbot:Idon’tknow,farmingmaybe

Page 16: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Cleverbot‣ Carpenter(1986),onlinesystembuiltin2006

‣ “Nearestneighbors”:humansaysstatementA,findahumanresponseinhuman-humanorhuman-computerchatstostatementA,repeatthat

User:Hitherecleverbot.Cleverbot:Howareyou?User:Doingokay.whatplansdoyouhavetoday?Cleverbot:Idorevision.User:Whatareyourevisingfor?Cleverbot:Mathshistory.User:Howoldareyou?Cleverbot:That'snoneofyourbusiness.User:Whatismathshistory?Cleverbot:Idon’tknow,farmingmaybe

Page 17: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Cleverbot‣ Carpenter(1986),onlinesystembuiltin2006

‣ “Nearestneighbors”:humansaysstatementA,findahumanresponseinhuman-humanorhuman-computerchatstostatementA,repeatthat

User:Hitherecleverbot.Cleverbot:Howareyou?User:Doingokay.whatplansdoyouhavetoday?Cleverbot:Idorevision.User:Whatareyourevisingfor?Cleverbot:Mathshistory.User:Howoldareyou?Cleverbot:That'snoneofyourbusiness.User:Whatismathshistory?Cleverbot:Idon’tknow,farmingmaybe

‣ Canofengivesensibleanswers,butthebotdoesn’treallyimposehigh-level discoursestructure

Page 18: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Data-DrivenApproaches‣ CantreatasamachinetranslaAonproblem:“translate”fromcurrentu4erancetonextone

Ri4eretal.(2011)

Page 19: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Data-DrivenApproaches‣ CantreatasamachinetranslaAonproblem:“translate”fromcurrentu4erancetonextone

Ri4eretal.(2011)

Page 20: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Data-DrivenApproaches‣ CantreatasamachinetranslaAonproblem:“translate”fromcurrentu4erancetonextone

‣ Filterthedata,usestaAsAcalmeasurestopruneextractedphrasestogetbe4erperformance

Ri4eretal.(2011)

Page 21: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Data-DrivenApproaches

Ri4eretal.(2011)

Page 22: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Seq2seqmodels

Whatareyoudoing

I

<s>

am going home [STOP]

‣ JustlikeconvenAonalMT,cantrainseq2seqmodelsforthistask

Page 23: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Seq2seqmodels

Whatareyoudoing

I

<s>

am going home [STOP]

‣ JustlikeconvenAonalMT,cantrainseq2seqmodelsforthistask

‣Whymightthismodelperformpoorly?Whatmightitbebadat?

Page 24: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Seq2seqmodels

Whatareyoudoing

I

<s>

am going home [STOP]

‣ JustlikeconvenAonalMT,cantrainseq2seqmodelsforthistask

‣Whymightthismodelperformpoorly?Whatmightitbebadat?

‣ Hardtoevaluate:

Page 25: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

LackofDiversity

Lietal.(2016)

‣ Trainingtomaximizelikelihoodgivesasystemthatpreferscommonresponses:

Page 26: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

LackofDiversity

Lietal.(2016)

‣ SoluAon:mutualinformaAoncriterion;responseRshouldbepredicAveofuseru4eranceUaswell

Page 27: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

LackofDiversity

Lietal.(2016)

‣ SoluAon:mutualinformaAoncriterion;responseRshouldbepredicAveofuseru4eranceUaswell

‣ StandardcondiAonallikelihood: logP (R|U)

Page 28: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

LackofDiversity

Lietal.(2016)

‣ SoluAon:mutualinformaAoncriterion;responseRshouldbepredicAveofuseru4eranceUaswell

‣MutualinformaAon:

‣ StandardcondiAonallikelihood: logP (R|U)

logP (R,U)

P (R)P (U)= logP (R|U)� logP (R)

‣ logP(R)canreflectprobabiliAesunderalanguagemodel

Page 29: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

LackofDiversity

Lietal.(2016)

‣ OpenSubAtlesdata

Page 30: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Futureofchatbots‣ HowdeepcanaconversaAonbewithoutmoresemanAcgrounding?Basicfactsaren’tevenconsistent…

Lietal.(2016)Persona…

Page 31: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Futureofchatbots‣ HowdeepcanaconversaAonbewithoutmoresemanAcgrounding?Basicfactsaren’tevenconsistent…

‣ Canforcechatbotstogiveconsistentanswers,butsAllprobablynotveryinteresAng

Lietal.(2016)Persona…

Page 32: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Futureofchatbots

‣ XiaoIce:MicrosofchatbotinChinese,20Musers,averageuserinteracts60Ames/month

‣ HowdeepcanaconversaAonbewithoutmoresemanAcgrounding?Basicfactsaren’tevenconsistent…

‣ Canforcechatbotstogiveconsistentanswers,butsAllprobablynotveryinteresAng

Lietal.(2016)Persona…

Page 33: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Futureofchatbots

‣ XiaoIce:MicrosofchatbotinChinese,20Musers,averageuserinteracts60Ames/month

‣ Peopledoseemtoliketalkingtothem…?

‣ HowdeepcanaconversaAonbewithoutmoresemanAcgrounding?Basicfactsaren’tevenconsistent…

‣ Canforcechatbotstogiveconsistentanswers,butsAllprobablynotveryinteresAng

Lietal.(2016)Persona…

Page 34: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue

Page 35: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue‣ QuesAonanswering/search:

Page 36: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue‣ QuesAonanswering/search:

Page 37: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue

Google,what’sthemostvaluable

Americancompany?

‣ QuesAonanswering/search:

Page 38: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue

Google,what’sthemostvaluable

Americancompany?

Apple

‣ QuesAonanswering/search:

Page 39: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue

Google,what’sthemostvaluable

Americancompany?

Apple

WhoisitsCEO?

‣ QuesAonanswering/search:

Page 40: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue

Google,what’sthemostvaluable

Americancompany?

Apple

WhoisitsCEO?

TimCook

‣ QuesAonanswering/search:

Page 41: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue‣ Personalassistants/APIfront-ends:

Page 42: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue

Siri,findmeagoodsushi restaurantinChelsea

‣ Personalassistants/APIfront-ends:

Page 43: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue

Siri,findmeagoodsushi restaurantinChelsea

SushiSekiChelseaisasushirestaurantinChelseawith4.4stars

onGoogle

‣ Personalassistants/APIfront-ends:

Page 44: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue

Siri,findmeagoodsushi restaurantinChelsea

SushiSekiChelseaisasushirestaurantinChelseawith4.4stars

onGoogle

‣ Personalassistants/APIfront-ends:

Howexpensiveisit?

Page 45: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue

Siri,findmeagoodsushi restaurantinChelsea

SushiSekiChelseaisasushirestaurantinChelseawith4.4stars

onGoogle

‣ Personalassistants/APIfront-ends:

Howexpensiveisit?

Entreesarearound$30each

Page 46: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue

Siri,findmeagoodsushi restaurantinChelsea

SushiSekiChelseaisasushirestaurantinChelseawith4.4stars

onGoogle

‣ Personalassistants/APIfront-ends:

Howexpensiveisit?

Entreesarearound$30each

Findmesomethingcheaper

Page 47: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue‣ Personalassistants/APIfront-ends:

Page 48: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue

HeyAlexa,whyisn’tmyAmazon orderhere?

‣ Personalassistants/APIfront-ends:

Page 49: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue

HeyAlexa,whyisn’tmyAmazon orderhere?

Letmeretrieveyourorder. Yourorderwasscheduledtoarrive

at4pmtoday.

‣ Personalassistants/APIfront-ends:

Page 50: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue

HeyAlexa,whyisn’tmyAmazon orderhere?

Letmeretrieveyourorder. Yourorderwasscheduledtoarrive

at4pmtoday.

‣ Personalassistants/APIfront-ends:

Itnevercame

Page 51: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Task-OrientedDialogue

HeyAlexa,whyisn’tmyAmazon orderhere?

Letmeretrieveyourorder. Yourorderwasscheduledtoarrive

at4pmtoday.

‣ Personalassistants/APIfront-ends:

Itnevercame

Okay,Icanputyouthroughtocustomerservice.

Page 52: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

AirTravelInformaAonService(ATIS)‣ Givenanu4erance,predictadomain-specificsemanAcinterpretaAon

DARPA(early1990s),FigurefromTuretal.(2010)

‣ CanformulateassemanAcparsing,butsimpleslot-fillingsoluAons(classifiers)workwelltoo

Page 53: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask‣ Parsing/languageunderstandingisjustonepieceofasystem

Youngetal.(2013)

Page 54: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask‣ Parsing/languageunderstandingisjustonepieceofasystem

Youngetal.(2013)

‣ Dialoguestate:reflectsanyinformaAonabouttheconversaAon(e.g.,searchhistory)

Page 55: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask‣ Parsing/languageunderstandingisjustonepieceofasystem

Youngetal.(2013)

‣ Dialoguestate:reflectsanyinformaAonabouttheconversaAon(e.g.,searchhistory)

‣ Useru4erance->updatedialoguestate->takeacAon(e.g.,querytherestaurantdatabase)->saysomething

Page 56: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask

Page 57: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask

FindmeagoodsushirestaurantinChelsea

Page 58: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

Page 59: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

location <- Chelsea

Page 60: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

location <- Chelsea

curr_result <- execute_search()

Page 61: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

location <- Chelsea

SushiSekiChelseaisasushirestaurantinChelseawith4.4starsonGoogle

curr_result <- execute_search()

Page 62: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

location <- Chelsea

SushiSekiChelseaisasushirestaurantinChelseawith4.4starsonGoogle

curr_result <- execute_search()

Howexpensiveisit?

Page 63: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

location <- Chelsea

SushiSekiChelseaisasushirestaurantinChelseawith4.4starsonGoogle

curr_result <- execute_search()

Howexpensiveisit?get_value(cost, curr_result)

Page 64: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

location <- Chelsea

SushiSekiChelseaisasushirestaurantinChelseawith4.4starsonGoogle

curr_result <- execute_search()

Howexpensiveisit?get_value(cost, curr_result)

Entreesarearound$30each

Page 65: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

POMDP-basedDialogueSystems

Youngetal.(2013)

‣ POMDP:useristhe“environment,”anu4eranceisanoisysignalofstate

Page 66: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

POMDP-basedDialogueSystems

Youngetal.(2013)

‣ Dialoguemodel:canlooklikeaparseroranykindofencodermodel

‣ POMDP:useristhe“environment,”anu4eranceisanoisysignalofstate

Page 67: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

POMDP-basedDialogueSystems

Youngetal.(2013)

‣ Dialoguemodel:canlooklikeaparseroranykindofencodermodel

‣ POMDP:useristhe“environment,”anu4eranceisanoisysignalofstate

‣ Generator:usetemplatesorseq2seqmodel

Page 68: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

POMDP-basedDialogueSystems

Youngetal.(2013)

‣ Dialoguemodel:canlooklikeaparseroranykindofencodermodel

‣ POMDP:useristhe“environment,”anu4eranceisanoisysignalofstate

‣ Generator:usetemplatesorseq2seqmodel

‣Wheredorewardscomefrom?

Page 69: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

RewardforcompleAngtask?

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

location <- Chelsea

SushiSekiChelseaisasushirestaurantinChelseawith4.4starsonGoogle

make_reservation(curr_result)

Howexpensiveisit?

+1

OkaymakemeareservaAon!

curr_result <- execute_search()

Page 70: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

RewardforcompleAngtask?

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

location <- Chelsea

SushiSekiChelseaisasushirestaurantinChelseawith4.4starsonGoogle

make_reservation(curr_result)

Howexpensiveisit?

+1

OkaymakemeareservaAon!

curr_result <- execute_search()

Veryindirectsignal ofwhatshould happenuphere

Page 71: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Usergivesreward?

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

location <- Chelsea

SushiSekiChelseaisasushirestaurantinChelseawith4.4starsonGoogle

curr_result <- execute_search()

Howexpensiveisit?get_value(cost, curr_result)

Entreesarearound$30each

+1

+1

Page 72: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Usergivesreward?

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

location <- Chelsea

SushiSekiChelseaisasushirestaurantinChelseawith4.4starsonGoogle

curr_result <- execute_search()

Howexpensiveisit?get_value(cost, curr_result)

Entreesarearound$30each

+1

+1

Howdoestheuserknowtherightsearchhappened?

Page 73: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Wizard-of-Oz

Kelley(early1980s),FordandSmith(1982)

‣ LearningfromdemonstraAons:“wizard”pullstheleversandmakesthedialoguesystemupdateitsstateandtakeacAons

Page 74: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask

FindmeagoodsushirestaurantinChelsea

Page 75: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

location <- Chelsea

curr_result <- execute_search(){wizardenters

these

Page 76: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

location <- Chelsea

curr_result <- execute_search(){wizardenters

these

SushiSekiChelseaisasushirestaurantinChelseawith4.4starsonGoogle{wizardtypesthis

outorinvokes templates

Page 77: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

location <- Chelsea

curr_result <- execute_search(){wizardenters

these

SushiSekiChelseaisasushirestaurantinChelseawith4.4starsonGoogle{wizardtypesthis

outorinvokes templates

‣Wizardcanbeatrainedexpertandknowexactlywhatthedialoguesystemsissupposedtodo

Page 78: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

LearningfromStaAcTraces

Bordesetal.(2017)

‣ Usingeitherwizard-of-OzorotherannotaAons,cancollectstaActracesandtrainfromthese

Page 79: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

location <- Chelsea

curr_result <- execute_search()

Page 80: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

location <- Chelsea

curr_result <- execute_search()stars <- 4+

Page 81: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

location <- Chelsea

curr_result <- execute_search()

‣ Useraskedfora“good”restaurant—doesthatmeanweshouldfilterbystarraAng?Whatdoes“good”mean?

stars <- 4+

Page 82: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

FullDialogueTask

FindmeagoodsushirestaurantinChelsea

restaurant_type <- sushi

location <- Chelsea

curr_result <- execute_search()

‣ Useraskedfora“good”restaurant—doesthatmeanweshouldfilterbystarraAng?Whatdoes“good”mean?

‣ HardtochangesystembehavioriftrainingfromstaActraces,especiallyifsystemcapabiliAesordesiredbehaviorchange

stars <- 4+

Page 83: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Goal-orientedDialogue

‣ BigCompanies:AppleSiri(VocalIQ),GoogleAllo,AmazonAlexa,MicrosofCortana,FacebookM,SamsungBixby,TencentWeChat

‣ Startups:

‣ Lotsofcoolworkthat’snotpublicyet

‣ Tonsofindustryinterest!

Page 84: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

OtherDialogueApplicaAons

Page 85: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Search/QAasDialogue

‣ “HasChrisPra4wonanOscar?”/“HashewonanOscar”

Page 86: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

QAasDialogue‣ DialogueisaverynaturalwaytofindinformaAonfromasearchengineoraQAsystem

Iyyeretal.(2017)

Page 87: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

QAasDialogue‣ DialogueisaverynaturalwaytofindinformaAonfromasearchengineoraQAsystem

Iyyeretal.(2017)

‣ QAishardenoughonitsown

‣ Challenges:

Page 88: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

QAasDialogue‣ DialogueisaverynaturalwaytofindinformaAonfromasearchengineoraQAsystem

Iyyeretal.(2017)

‣ QAishardenoughonitsown

‣ Usersmovethegoalposts

‣ Challenges:

Page 89: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

QAasDialogue‣ UWQuACdataset:QuesAonAnsweringinContext

Choietal.(2018)

Page 90: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

SearchasDialogue

‣ Googlecandealwithmisspellings,somoremisspellingshappen—Googlehastodomore!

Page 91: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

DialogueMissionCreep

System

Erroranalysis

Be4ermodel

Data

MostNLPtasks

Page 92: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

DialogueMissionCreep

System

Erroranalysis

Be4ermodel

‣ FixeddistribuAon(e.g.,naturallanguagesentences),errorrate->0

Data

MostNLPtasks

Page 93: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

DialogueMissionCreep

System

Erroranalysis

Be4ermodel

‣ FixeddistribuAon(e.g.,naturallanguagesentences),errorrate->0

Data

MostNLPtasks

System

Erroranalysis

Be4ermodel

Data

Dialogue/Search/QA

Page 94: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

DialogueMissionCreep

System

Erroranalysis

Be4ermodel

‣ FixeddistribuAon(e.g.,naturallanguagesentences),errorrate->0

DataHarderData

MostNLPtasks

System

Erroranalysis

Be4ermodel

Data

Dialogue/Search/QA

???

Page 95: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

DialogueMissionCreep

System

Erroranalysis

Be4ermodel

‣ FixeddistribuAon(e.g.,naturallanguagesentences),errorrate->0

Data

‣ Errorrate->???;“missioncreep”fromHCIelement

HarderData

MostNLPtasks

System

Erroranalysis

Be4ermodel

Data

Dialogue/Search/QA

???

Page 96: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

DialogueMissionCreep

‣ Highvisibility—yourproducthastoworkreallywell!

Page 97: Lecture 16: Dialoguearitter.github.io/courses/5525_slides_v2/lec16-dialogue.pdfFull Dialogue Task ‣ Parsing / language understanding is just one piece of a system Young et al. (2013)

Takeaways

‣ Somedecentchatbots,applicaAons:predicAvetextinput,…

‣ Task-orienteddialoguesystemsaregrowinginscopeandcomplexity

‣Moreandmoreproblemsarebeingformulatedasdialogue—interesAngapplicaAonsbutchallengingtogetworkingwell