How Contexts Matter Understanding in Dialoguesaliensunmin.github.io/aii_workshop/2nd/slides/1.pdf · How Contexts Matter Understanding in Dialogues ... Contexts provide informative

HowContextsMatterUnderstandinginDialoguesYUN-NUNG (VIVIAN)CHEN

§ Word-LevelContextsinSentences§ LearningfromPriorKnowledge–

Knowledge-Guided StructuralAttentionNetworks(K-SAN)[Chenetal.,‘16]§ LearningfromObservations–

ModularizingUnsupervisedSenseEmbedding (MUSE)[Lee&Chen,‘17]

§ Sentence-LevelContextsinDialogues§ InvestigationofUnderstanding Impact–

ReinforcementLearningBasedNeuralDialogueSystem[Lietal.,‘17]

§ Conclusion

2

§ Dialoguesystemsareintelligentagentsthatareabletohelpusersfinishtasksmoreefficientlyviaconversationalinteractions.

§ Dialoguesystemsarebeingincorporatedintovariousdevices(smart-phones,smartTVs,in-carnavigatingsystem,etc).

3

JARVIS– IronMan’sPersonalAssistant Baymax – PersonalHealthcareCompanion

§ Word-levelcontext§ Priorknowledgesuchaslinguistic syntax

§ Collocatedwords

§ Sentence-levelcontext

4

Smartphone companiesincludingapple,blackberry,andsony willbeinvited.

showmetheflights fromseattle tosanfrancisco

(browsingactionmoviereviews…)Findmeagoodonethisweekend

LondonHasFalleniscurrentlythenumber1actionmovieinAmerica

request_movie(genre=action,date=thisweekend)

Howmisunderstanding influences thedialoguesystemperformance

Contextsprovide informativecuesforbetterunderstanding

5

Knowledge-GuidedStructuralAttentionNetwork(K-SAN)

Y.-N.Chen,D.Hakkani-Tur,G.Tur,A.Celikyilmaz, J.Gao,andL.Deng,“KnowledgeasaTeacher:Knowledge-GuidedStructuralAttentionNetworks,”preprintarXiv:1609.00777, 2016.

§ Syntax(DependencyTree) § Semantics(AMRGraph)

6

show

me

the

flights

from

seattle

to

san

francisco

ROOT

1.

3.

4.

2.

1.showme2.showflightsthe3.showflightsfromseattle4.showflightstofranciscosan

Sentence sshowmetheflightsfromseattle tosanfrancisco

Knowledge-GuidedSubstructurexi

(s/show:ARG0(y/you):ARG1(f/flight

:source(c/city:name(d/name:op1Seattle))

:destination(c2/city:name(s2/name:op1San:op2Francisco)))

:ARG2(i /I):modeimperative)

Knowledge-GuidedSubstructurexi1.showyou2.showflightseattle3.showflightsanfrancisco4.showi

show

you

flightI

1.

2.

4.

citycity

Seattle

SanFrancisco3..

Y.-N.Chen,D.Hakkani-Tur,G.Tur,A.Celikyilmaz, J.Gao,andL.Deng,“KnowledgeasaTeacher:Knowledge-GuidedStructuralAttentionNetworks,”preprintarXiv:1609.00777, 2016.

knowledge-guidedstructure{xi}

KnowledgeEncoding

SentenceEncoding

InnerProduct

u

mi

KnowledgeAttention Distributionpi

EncodedKnowledgeRepresentation WeightedSum

∑

h

o

Knowledge-GuidedRepresentation

slottaggingsequence

s

y

show me the flights from seattle to san francisco

ROOT

InputSentence

ht-1 ht+1htW W W W

wt-1

yt-1

U

Mwt

U

wt+1

U

Vyt

Vyt+1

V

MM

RNNTagger

KnowledgeEncodingModule

CNNkg

CNNin NNout

7Y.-N.Chen,D.Hakkani-Tur,G.Tur,A.Celikyilmaz, J.Gao,andL.Deng,“KnowledgeasaTeacher:Knowledge-GuidedStructuralAttentionNetworks,”preprintarXiv:1609.00777, 2016.

Themodelwillpaymoreattentiontomoreimportantsubstructures thatmaybecrucialforslottagging.

§ Darkerblocksandlinescorrespondtohigherattentionweights


§ Darkerblocksandlinescorrespondtohigherattentionweights

K-SANlearnsthesimilarattentiontosalientsubstructureswithlesstrainingdata


10

ModularizingUnsupervisedSenseEmbeddings (MUSE)

G.-H.LeeandY.-N.Chen, “MUSE:ModularizingUnsupervised SenseEmbeddings,” inEMNLP,2017.

§ Wordembeddings aretrainedonacorpusinanunsupervisedmanner

§ Usingthesameembeddings fordifferentsenses forNLPtasks,e.g.NLU,POStagging

11

FinallyIchoseGoogle insteadofApple.

Canyoubuymeabagof apples,oranges,andbananas?


Wordswithdifferent sensesshouldcorrespond differentembeddings

Smartphone companiesincluding blackberry,andsony willbeinvited.

§ Input:unannotatedtextcorpus

§ Twokeymechanisms§ Senseselection givenatextcontext§ Senserepresentation toembedstatisticalcharacteristicsofsenseidentity


apple

apple-1 apple-2senseselection

senseembedding

12

§ Senseselection§ Policy-based

§ Value-based

13

Corpus:{Smartphonecompaniesincludingapple blackberry,andsonywillbeinvited.}

senseselection←

rewardsignal←

senseselection→

samplecollocation1

2

2

3

Senseselectionforcollocatedword𝐶$%

SenseSelectionModule

…𝐶$' = 𝑤*𝐶$'+,

𝑞(𝑧*,|𝐶$') 𝑞(𝑧*2|𝐶$') 𝑞(𝑧*3|𝐶$')

matrix𝑄*

matrix𝑃

… 𝐶$'6,apple andincluding sonyblackberry

𝑧7,

SenseRepresentationModule

…𝑃(𝑧*2|𝑧7,) 𝑃(𝑧89|𝑧7,)

negativesampling

matrix𝑉

matrix𝑈

§ Senserepresentationlearning

§ Skip-gramapproximation

SenseSelectionModule

…𝐶$ = 𝑤7𝐶$+,

𝑞(𝑧7,|𝐶$< ) 𝑞(𝑧72|𝐶$< ) 𝑞(𝑧73|𝐶$< )

Senseselectionfortargetword𝐶$

matrix𝑄7

matrix𝑃

… 𝐶$6,including apple blackberrycompanies and

Collocatedlikelihood servesasarewardsignaltooptimizethesenseselectionmodule.

§ Dataset:SCWSformulti-senseembeddingevaluation

14

Approach MaxSimC AvgSimCHuangetal.,2012 26.1 65.7Neelakantan etal.,2014 60.1 69.3Tianetal.,2014 63.6 65.4Li&Jurafsky,2015 66.6 66.8Bartunov etal.,2016 53.8 61.2Qiu etal.,2016 64.9 66.1MUSE-Policy 66.1 67.4MUSE-Greedy 66.3 68.3MUSE-ε-Greedy 67.4+ 68.6

Heborrowedthemoney frombanks. Iliveneartoariver. correlation=?

15

Context …bravesfinishtheseasonintiewiththelosangelesdodgers…

…hislateryearsproudlyworetiewiththechinese charactersfor…

k-NN scorelessotl shootout6-6hingis 3-37-70-0

pantstrousersshirtjuventusblazersocksanfield

Figure

16

Context …ofthemulberryortheblackberry andminos senthimto…

…ofthelargenumberofblackberry usersintheusfederal…

k-NN cranberriesmaplevacciniumapricotapple

smartphonessapmicrosoft ipv6smartphone

Figure

17

Context …shellsand/orhighexplosivesquashhead and/oranti-tank…

… headwasshaventoprevent head liceseriousthreatbackthen…

…appointjohnpoperepublicanashead ofthenewarmyof…

k-NN venterthoraxneckspearmillimetersfusiform

shavedthatcherlokithoraxmao luther chest

multi-partyappointsunicameralberiaappointed

Figure

MUSElearnssenseembeddings inanunsupervised wayandachievesthefirstpurelysense-level representation learningsystemwithlinear-timesenseselection

RL-BasedNeuralDialogueSystems

X.Li,Y.-N.Chen, L.Li,J.Gao,andA.Celikyilmaz, “End-to-EndTask-Completion NeuralDialogueSystems,”inIJCNLP,2017.

§ Dialoguemanagement isframedasareinforcementlearning task

§ Agentlearnstoselectactionstomaximizetheexpectedreward

19

Environment

Observation

Action

Reward

Ifbookingarightticket,reward=+30

Iffailing,reward=-30

Otherwise,reward=-1

Agent


§ Dialoguemanagement isframedasareinforcementlearning task

§ Agentlearnstoselectactionstomaximizetheexpectedreward

20

Environment

Observation

Action

Agent

NaturalLanguageGeneration

UserAgendaModeling

UserSimulator

LanguageUnderstanding

DialogueManagement

NeuralDialogueSystem

TextInput:Arethereanyactionmovies toseethisweekend?

DialoguePolicy:request_location



wi

<slot>

wi+1

O

EOS

<intent>

wi

<slot>

wi+1

O

EOS

<intent>

§ LU,DST(neuraldialoguesystem),andNLG(usersimulation)aretrainedinsupervisedway

§ End-to-endtrainingfordialoguepolicylearning

KnowledgeDatabase

NeuralDialogueSystem

UserModel

UserSimulation

DialoguePolicy

NaturalLanguage

w0

w1 w2

NLGEOS

UserGoal

wi

<slot>

wi+1

O

EOS

<intent>

LU

𝑠$DST

𝑠, 𝑠2 𝑠>

𝑎, 𝑎2 𝑎@

……

…

DialoguePolicyLearning

21

Usergoal – Twoticketsfor“Deadpool” tomorrow9PMatAMCPacificPlace11theater,Seattle.

RULE BASEDSYSTEMUser: Doyouhaveanyticketformovie‘Deadpool’at9:00pm?Agent:Whichtheaterdoyouwant?User: AMCpacificplace11theater,please!Agent:Howmanyticketsdoyouwant?User: Two!Agent:Whatdate?User: Tomorrow!Agent:Whichcitydoyouwant?User: Seattle!Agent:Doyouhaveanyspecification forzipcode?User: No,Idonotcare.Agent:Doyouhaveanydistanceconstraint?User: No,Idonotcare.Agent:Howaboutvideoformat?2d?Imax3d?User: Anykindisokay.Agent:Doyouhaveanyconstrainttoprice?User: No!Agent:Okay,Iwill issue2ticketsforyou,tomorrow9:00pmatAMCpacificplace11theater,Seattle,movie‘Deadpool’.User: Okay,thanks.Agent:Thanks!

REINFORCEMENT LEARNINGSYSTEMUser: Iwanttwoticketsformovie‘Deadpool’.Agent:Whatisthestarttime youprefer?User: 9:00PMAgent:Whichday?User: TomorrowAgent:Whichcity?User: SeattleAgent:Whichtheaterdoyouprefer?User: AMCPacificPlace11theater.Agent:Okay,Iwill issue2ticketsforyou,tomorrow9:00pmatAMCpacificplace11theater,Seattle,movie‘Deadpool’.User: Okay,thanks.Agent:Thanks!

Skiptherequeststheusermaynotcareabout toimproveefficiency

Issue1:nonotionaboutwhatrequestscanbeskipped

22

Usergoal – Twoticketsfor“Hail Caesar”tomorrow7:15PMatAMCLowesOaktree6,Seattle.

Automaticallylearnwhenandwhattoconfirmforbetterrobustness

REINFORCEMENT LEARNINGSYSTEMAgent:Whatstarttimedoyouprefer?User: 7:15pmTomorrowAgent:Whichcityyouarelookingfor?User: SeattleAgent:Okay,Iwill issueyouticketsformovie‘HailCaeser ’tomorrow7:50pmattheaterAMCLowesOakTree6,Seattle.User: Oh,no.Agent: Itis7:50pm,right?User: Twoticketsfor7:15pm,please.Agent:Okay,Iwill issueyoutwoticketsformovie‘HailCaeser ’tomorrow7:15pmattheaterAMCLowesOakTree6,Seattle.User: Thanks!Agent:Thanks!

RULE BASEDSYSTEMUser: Doyouanyticketformovie‘HailCaesar ’attheaterAMCLowesOakTree6?Agent:Whatstarttimedoyouprefer?User: 7:15pm TomorrowAgent:Whichcityyouarelookingfor?User: SeattleAgent:Howmanyticketsdoyouwant?User: TwoAgent:Okay,Ihaveissuedtwotickets formovie‘HailCaeser ’tomorrow7:50pmattheater AMCLowesOakTree6,Seattle.User: Oh,no.Agent:Thanks!

Issue2:noprinciplewaytodealwithLUerrors/uncertainty

23

§ Userspresentedwithapredefinedusergoal

§ Userrating:1(worst)to5(best)basedonbothnaturalness andcoherence ofthedialogue

24

§ Word-levelcontextsinsentenceshelpunderstandwordmeanings§ LearningfromPriorKnowledge–

K-SANachievesbetterLUviaknownknowledge [Chenetal.,‘16]§ LearningfromObservations–

MUSElearnssenseembeddings withefficientsenseselection[Lee&Chen,‘17]

§ Sentence-levelcontextshavedifferentimpactsondialogueperformance§ InvestigationofUnderstanding Impact–

Sloterrorsdegradesystemperformancemorethanintenterrors[Lietal.,‘17]

§ Contextsfromdifferentlevelsprovidecuesforbetterunderstandinginsupervisedandunsupervisedways

25

QA

How Contexts Matter Understanding in Dialoguesaliensunmin.github.io/aii_workshop/2nd/slides/1.pdf · How Contexts Matter Understanding in Dialogues ... Contexts provide informative

Documents