您的当前位置：首页 An Information State Approach to Natural Interactive Dialogue

An Information State Approach to Natural Interactive Dialogue

来源：九壹网

AnInformationStateApproachtoNaturalInteractiveDialogue

StaffanLarssonandRobinCooper

DepartmentoflinguisticsG¨oteborgUniversityBox200-295,HumanistenSE-40530G¨oteborg

SWEDEN

sl,cooper@ling.gu.se

Abstract

2Menusvs.Dialogue

Muchofourinteractionwithautomatedtaskorientedsys-temsiscurrentlymenudriven.Forexample,anautomated

systemtobookacinematicketcaninvolveﬁrstchoosingthecity,thentheﬁlm,thecinemaandthetime.Currenttechnologyusingatelephoneinterfaceinvolveslisteningtoanumberoptionsandrespondingbypressinganumber.Anexampleofamorecomplexmenustructureispresentintheprogrammingfacilitiesofmobilephoneswherethemenucanbeviewedasatreestructurewhichtheuserhastodescendstartingattherootnode.

IntheTRINDI,SDSandINDIprojectswehaveusedtheno-tionofdialogueinformationstatetoexplorethenatureofdia-loguemoves.Theseinformationstatesrepresenttheinformationavailabletoadialogueparticipantatagivenstageofthedia-logue.Inthispaper,wewilldiscusshowtaskorientedsystemsofseveralkinds-menudriven,informationseeking,andinstruc-tional-canbeimplementedusingtechniquesdevelopedintheseprojects,andhowthesetechniquessupportnaturalinteractivedialoguewheretheuserintuitivelycontrolswhatinformationispresentedandinwhatorder.

Whilemenuinterfacesareubiquitousinmoderntech-nologytheyareoftentediousandfrustrating.Currentdia-loguetechnologycanstraightforwardlyimplementmenu

1Introductionstructuresbuttheeffectcanbeevenmoretedioussince

theuserisforcedtousethenaturalandefﬁcientmodality

Inthispaper,wewilldiscusshowthenotionofdialogueofspeechinaveryunnaturalwaydescendingthemenuinformationstateprovidesabasisbothforformulatingstructureonenodeatatime.theoriesaboutdialogueandfortheimplementationtaskorientedsystemsofseveralkinds.WepresenttechniquesTRINDItechnologyoffersthepossibilityofallowingtosupportnaturalinteractivedialoguewheretheuserintu-theusertopresentseveralpiecesofrelevantinformationitivelycontrolswhatinformationispresentedandinwhatatonetimeortopresentinformationintheorderinwhichorder.Thepaperisstructuredaroundthreetypesofdia-theuserﬁndsmostnatural(eitherinspeechorwrittenlogueandcorrespondingdialoguesystems:menudriven,interfaces).Thesystemwillrequestinformationthatisinformationseeking,andinstructional.Intheﬁrstsec-missing.Ifinformationpresentedbytheuserisrelevanttion,thepossibilityofimprovingonmenu-driveninter-toaspeciﬁcbranchofthemenusystemthenthesystemfacesusingTRINDItechnologyispointedout.Inthewillbeabletojumpdirectlytothatbranchwithoutrequir-followingsection,anapproachtoinstructionaldialogueingtheusertostepthroughalltheinterveningnodes.Ifispresented,buildingonanotionofaccommodationoftheuserdoesnotknowwhattodothesystemcanpresentquestions,tasksandplans.Thisapproachhasbeenim-theoptions.Thismeansthatuserscanusetheirowncon-plementedintheGoDiSsystem.Intheﬁnalsection,weceptionoftheknowledgespaceandnotbelockedtothatpresentpreliminaryresearchbetweenwrittenmanualsandofthedesignerofthesystem.Thiswillallowformoreinstructionaldialogue.Asmallexperimentalsystemfornaturaldialogicinteractionwithcomputersinsituationshandlinginstructionalmonologueanddialogue,basedonwheretheconventionalinterfaceofkeyboardandscreenGoDiS,ispresented.isimpossibleorundesirable.

3Questionandtaskaccommoda-tionininformation-seekingdia-logue

Information-seekingsystemsissimilarinmanywaystomenu-drivensystems;basically,theuserneedstogivethesystemcertaininformationwhichenablestosystemtoperformitstask-bookingaticket,providingpricein-formation,makingacalletc.Therefore,thetechniquesdevelopedforinformation-seekingdialogueapplyalsotomenu-drivensystems.ThetechniquesdescribedinthissectionhavebeenimplementedinGoDiS[Bohlinetal.,1999],anexperimentaldialoguesystemcurrentlyadaptedforthetravelagencydomain.GoDiSisimplementedus-ingtheTRINDIKIT[Larssonetal.,1999],atoolkitforexperimentingwithinformationstatesanddialoguemoveenginesandforbuildingdialoguesystems.

Onofthetechniqueswehaveinvestigatedisanotionofaccommodationbasedonaninformationupdateperspec-tiveonLewis’notionofaccommodation[Lewis,1979].Most(ifnotall)accountsofaccommodationhavedealtwithreferentsorpropositions;whatweproposetodoistoextendthenotionofaccommodationtoquestions.Weﬁrstgiveabriefintroductiontothekindofinfor-mationstatesweareassuming.Wethendiscussquestionaccommodationandshowhowtoformulateitintermsofinformationstateupdaterules.Suchrulesprovideanin-tuitivewayofformalisingquestionaccommodation,andcanbeusedindialoguesystemstoimprovenaturalnessofhuman-machinecommunication.

ThenotionofinformationstateweareputtingforwardhereisbasicallyaversionofthedialoguegameboardwhichhasbeenproposedbyGinzburg[Ginzburg,1998].Werepresentinformationstatesofadialogueparticipantasarecordofthetypeshowninﬁgure1.

Themaindivisionintheinformationstateisbetweeninformationwhichisprivatetotheagentandthatwhichissharedbetweenthedialogueparticipants.TheprivatepartoftheinformationstatecontainsaPLANﬁeldhold-ingadialogueplan,i.e.isalistofdialogueactionsthattheagentwishestocarryout.TheAGENDAﬁeld,ontheotherhand,containstheshorttermgoalsorobligationsthattheagenthas,i.e.whattheagentisgoingtodonext.WehaveincludedaﬁeldTMPthatmirrorsthesharedﬁelds.Thisﬁeldkeepstrackofsharedinformationthathasnotyetbeengrounded,i.e.conﬁrmedashavingbeenunder-stoodbytheotherdialogueparticipant.TheSHAREDﬁeldisdividedintothree.Onesubﬁeldisasetofpropositionswhichtheagentassumesforthesakeoftheconversation.Theothersubﬁeldisforastackofquestionsunderdis-cussion(QUD).Thesearequestionsthathavebeenraisedandarecurrentlyunderdiscussioninthedialogue.TheLUﬁeldcontainsinformationaboutthelatestutterance.

Wedeﬁneupdaterulesforupdatingtheinformationstatebasedontherecognisedmove(s).Therulesarede-ﬁnedintermsofpreconditionsandeffectsontheinfor-mationstate;theeffectsarealistofoperationstobeex-ecutedifthepreconditionsaretrue.Asanexample,theupdateruleforintegratingananswer(asimpliﬁedver-sionofGinzburg’sQUDdowndaterule)isshownbelow.

RULE:

QUD-downdate

member(SHARED.LU.MOVES,answer())fst(SHARED.QUD,),PRE:

answer

weuse“he”torefertothedialoguesystemandthehuman

travelagentmoreorlessinterchangeably.Thisisnotmeanttoentailany

1Here

PLANAGENDABEL

::::

PRIVATE

TMP

StackSet(Action)Stack(Action)Set(Prop)

BEL:Set(Prop)QUD:Stack(Question)

SPEAKER

MOVES

::Participant

AssocSet(Move,Bool)

SHARED

BELQUDLU

:::

Set(Prop)

Stack(Question)

SPEAKERMOVES

::Participant

AssocSet(Move,Bool)

Figure1:Thetypeofinformationstateweareassuming

andthencontinueswiththeintegrationoftheinformationexpressedbyascheapaspossibleasnormal.Thus,wemakeaconnectionbetweentherelevantquestionbeinginthedialogueplanandbeingabletointerprettheellipsis.Notethatifsuchaquestionisnotavailablethentheellipsiscannotbeinterpretedasinthisdialogue:

A.WhattimeareyoucomingtopickupMaria?B.Around6p.m.Ascheaplyaspossible.

dialogue,theonlyitemontheagent’splanistoraisethequestionofwhatthecustomerwants,i.e.thequestionaboutthetaskofthedialogue.Aftercontacthasbeenes-tablished,a“normal”dialoguewouldcontinuewiththeagentasking“WhatcanIdoforyou?”andthecustomerreplyingbyspecifyingthetasktobeachieved,suchasﬁndingoutthepriceofatrip.However,afteraninitialex-changeforestablishingcontacttheﬁrstthingthatsaystothetravelagentinourdialogueis:

ThisdialogueisincoherentifwhatisbeingdiscussediswhenthechildMariaisgoingtobepickedupfromherfriend’shouse(atleastunderstandarddialogueplansthatwemighthaveforsuchaconversation).

Theupdateruleforquestionaccommodation,QuAcc,

isshownbelow.Wheninterpretingthelatestutterancebytheotherparticipant,thesystemmakestheassumptionthatitwasareplymovewithcon-tent.Thisassumptionrequiresaccommodatingsomequestionsuchthatisarelevantanswerto.Thecheckoperator“answer-to()”istrueifisarelevantanswertogiventhecurrentin-formationstate,accordingtosome(possiblydomain-dependent)deﬁnitionofquestion-answerrelevance.

C:flightstoParis

RULE:

QuAcc

member(SHARED.LU.MOVES,answer()),val(SHARED.LU.SPEAKER,usr),

PRE:

member(PRIVATE.PLAN,raise())answer

Thisisagainanellipsiswhichonouranalysishastobeinterpretedastheanswertoaquestioninordertohavecontent.Noquestionshavebeenraisedyetinthedialogue,andtheagentcannotﬁndtheappropriatequestiononhisplan.Furthermore,asthisistheﬁrstindicationofwhatthecustomerwants,thetravelagentdoesnothaveaplanwithdetailedquestions.Weas-sumethatthetravelagenthasvariousplantypesforhan-dlingdifferenttasksinhisdomainknowledgedetermin-ingwhatkindofconversationsheisabletohave.E.g.heisabletobooktripsbyvariousmodesoftravel,heisabletohandlecomplaints,bookhotelsetc.Whatheneedstodoistakethecustomer’sutteranceandtrytomatchitagainstquestionsintheplantypesinhisdo-mainknowledge.Whenheﬁndsasuitablematchhewillaccommodatethetaskcorrespondingtothematch-ingplanbyaddingapropositionindicatingthetasktothesharedbeliefs.ThisdonebytheTaskAccrule.

RULE:

TaskAccmember(

PRE:

match

SHARED.LU.MOVES,

)

claimsofpsychologicalrealityofthekindsofinformationstateweareassuming.

2Notethatthisstrategyrequiresthat(a)theagenthasaplan,and(b)thereisasinglequestion,theraisingofwhichispartoftheplan,whichmatchestheanswer.Incaseswherethereareseveralcandidatequestions,adialogueparticipantcouldeithertakeachanceonthemostlikelycandidate,orsomehowindicatethathecannotintegratethean-swer.

interpreted.Inthemonologueplan,theycorrespondtosimplyinformingtheuseroftheconditional.Indialoguemode,however,thesystemraisesthequestionwhethertheconditionholds.Whenthesystemﬁndsoutifthecondi-tionholds,itwillinstructtheusertoexecutetheappropri-ateguardedaction.

Inshort,here’showconditionalsaretreatedbythesys-temindialoguemode:Whenthesystemhasfoundout

4Frommanualstoinstructionaldi-whattheuser’staskis,itwillloadtheappropriatedialogue

planintothePRIVATE.PLANﬁeldoftheinformationstate.alogue

Itwillthenexecutetheactionsintheappropriateorderby

Inthissection,wepresentpreliminaryresearchonthemovingthemtotheagendaandgeneratingappropriateut-relationbetweenwrittenmanualsandinstructionaldia-terances.Whenaconditionalstatementistopmostonthelogue.Weoutlinehowamanualcanbeconvertedintoplan,IMDiSwillcheckwhetherithasbeenestablishedaformatwhichcanbeusedasdomainknowledgebyathattheconditionholds(bycheckingtheSHARED.BELdialoguesystem,capableofgeneratingbothnaturalin-ﬁeld).Sincethesystemhaspreviouslyaskedtheuserandstructionaldialogueandmonologue.Startingfromashorttheuserhasanswered,eithertheconditionoritsnegationsampletextfromamanual,weusetheTRINDIinforma-willbeinthesetofestablishedpropositions.Ifthecondi-tionstateapproach[Traumetal.,1999]tobuildanexper-tionoritsnegationholds,theconditionalwillbepoppedimentaldialoguesystemcapableofinstructingausertoofftheplanandreplacedbytheﬁrstorsecondguardedperformthetask.IMDiS,asmallexperimentalimplemen-action(respectively).ThisprocessisimplementedinthetationbasedontheGoDiSdialoguesystem,ispresented.formofinformationstateupdaterules.Welookatsamplemonologueanddialogueoutputand

discusstheadvantagesprovidedbythedialoguemodein4.2MonologueanddialoguebehaviourIMDiS.Oneofthemainadvantagesisthattheusercan

controlthedialoguetomakethesystemprovideexactlyInthemonologuemodeinIMDiS,thecontrolmodule

doesnotcalltheinputandinterpretationmodules.Thetheinformationneeded.

Thetextbelowistakenfromausermanualforthetextisoutput“movebymove”asasequenceofutterances

fromthesystem.Comparedtothemonologuemode,theHomecentre,alowendXeroxMultiFunctionalDevice.

dialoguemodeoffersseveraladvantages:

Reinstallingtheprinthead

Caution:MakesurethatthegreencarriagelockleverisSTILLmovedallthewayforwardbeforeyoureinstalltheprinthead.

1.Lineuptheholeintheprintheadwiththegreenpostontheprintercarriage.

Lowertheprintheaddowngentlyintoposition.

2.Gentlypushthegreencartridgelockleverupuntilitsnapsintoplace.

Thissecurestheprinthead.

3.Closethetopcoverandreattachthescanner.4.PressandreleasetheyellowLEDbutton.

Theprinterwillpreparethecartridgeforprinting.Note:Ifthecarriagedoesnotmovefromthecenterpositionafteryoupressthecartridgechangebutton,removeandreinstalltheprinthead.

servestodrivethedialogueforward.Thatis,themecha-nismbywhichtheagentinterpretsthisellipsis,giveshimaplanforasubstantialpartoftherestofthedialogue.ThisisawayofcapturingtheintuitionthatsayingﬂightstoParistoatravelagentimmediatelymakesanumberofquestionsbecomerelevant.

UserattentionandcontrolTheusercandirectherat-tentiontothemachineanddoesnothavetolookatthemanual.Thismeansthattheuserdoesnothavetokeeptrackofthecurrentpointinthemanual.Also,theuserde-cideswhentomoveontothenextaction,byconﬁrmingthatthepreviousactionisdone.

AvoidirrelevantinformationWhentheactiontobetakendependsonacondition,thesystemdoesnotgiveirrelevantinformation.

S:Hasthecarriagemovedfromthecenterposi-tion?

4.1Monologueanddialogueplans

Fromthistext,onecan(re)constructadomainplanforU:yes

reinstallingtheprinthead.SuchaplanmayberepresentedS:Theprintheadisnowinstalledasinﬁgure2.Notethatthisisaconditionalplan,i.e.itcontainsbranchingconditions.Fromthisplan,IMDiS

generatestwoplans:amonologueplanandadialogueControloverlevelofdetailIftheuserdoesnotknowplan.Thisisdoneusingthe“translationschema”intablehowtoperformasubstep,shecanaskthesystemformore

detailedinstructions.1.

Thedifferencebetweenthetextplanandthedialogue

planisinthewaythatconditionalsinthedomainplanareS:securetheprinthead

NAMEPREDECreinstall(print_head)moved_forward(carriage_lock)secure(print_head)NAMEPREDECsecure(print:head)line_up(hole, post)lower(print_head)close(top_cover)push(lever)reattach(scanner)press_and_release(yellow_button)yesmoved_from_center(print_head)noEFFsecured(print_head)remove(print_head)actioncomplex action / planfinal statereinstall(print_head)EFFreinstalled(print_head)Figure2:Sampledomainplan

DOMAIN

DIALOGUE

then

InstructCheck()InstructExec()Inform(ifelse(,,

))

then

Inform()

S:Hasthecarriagemovedfromthecenterposi-tion?U:what?

S:Hasthecarriagemovedfromthecenterposi-tion?

5Conclusion

Webelieve,andhaveattemptedtoshowinthispaper,thattheinformationstateapproachprovidesafruitfulfoun-dationforexploringthenatureofnaturalinteractivedi-alogueandforimplementingdialoguesystemswheretheuserintuitivelycontrolswhatinformationispresentedandinwhatorder.

Acknowledgments

ThispaperreportsworkcarriedoutinthetheTRINDI(TaskOrientedInstructionalDialogue),ECProjectLE4-8314),SDS(SwedishDialogueSystems,NUTEK/HSFRLanguageTechnologyProjectF1472/1997)andINDI(InformationExchangeinDialogue,RiksbankensJu-bileumsfond1997-0134)projects.

References

[Bohlinetal.,1999]P.Bohlin,R.Cooper,E.Engdahl,andS.Larsson.Informationstatesanddialoguemoveengines.InJ.Alexandersson,editor,IJCAI-99Work-shoponKnowledgeandReasoninginPracticalDia-logueSystems,1999.[Ginzburg,1998]J.Ginzburg.Clarifyingutterances.InJ.HulstijnandA.Niholt,editors,Proc.oftheTwenteWorkshopontheFormalSemanticsandPragmaticsofDialogues,pages11–30,Enschede,1998.UniversiteitTwente,FaculteitInformatica.[Larssonetal.,1999]StaffanLarsson,PeterBohlin,Jo-hanBos,andDavidTraum.Trindikitmanual.Techni-calReportDeliverableD2.2-Manual,Trindi,1999.[Lewis,1979]D.K.Lewis.Scorekeepinginalanguagegame.JournalofPhilosophicalLogic,8:339–359,1979.[Traumetal.,1999]DavidTraum,JohanBos,RobinCooper,StaffanLarsson,IanLewin,ColinMatheson,andMassimoPoesio.Amodelofdialoguemovesandinformationstaterevision.TechnicalReportDeliver-ableD2.1,Trindi,1999.

因篇幅问题不能全部显示，请点此查看更多更全内容

查看全文