D3.3-final multilingual adaptive dialogue toolbox...Deliverable (D3.3). Release of the multilingual, adaptive dialogue toolbox 1. PURPOSE OF DOCUMENT This is the last deliverable for

ARIA Valuspa

EuropeanUnion’sHorizon2020researchandinnovationprogramme645378,ARIA-VALUSPA

September,2017

-1-

Deliverable(D3.3).Releaseofthemultilingual,adaptivedialoguetoolbox

ArtificialRetrievalofInformationAssistants–VirtualAgentswithLinguisticUnderstanding,Socialskills,andPersonalisedAspects

CollaborativeProject

Startdateofproject:01/01/2015 Duration:36months

(D3.3).Releaseofthemultilingual,adaptivedialoguetoolbox

Duedateofdeliverable:Month31 Actualsubmissiondate:28/09/2017

ARIAValuspa

ARIA Valuspa


September,2017

-2-


Projectco-fundedbytheEuropeanCommission

DisseminationLevel

PU Public X

PP Restrictedtootherprogrammeparticipants(includingtheCommissionServices)

RE Restrictedtoagroupspecifiedbytheconsortium(includingtheCommissionservices)

CO Confidential,onlyformembersoftheconsortium(includingtheCommissionServices)

STATUS:[FINAL]

DeliverableNature

R Report

P Prototype

D Demonstrator X

O Other

ParticipantNumber

Participantorganizationname Participantorg.shortname

Country

Coordinator1 UniversityofNottingham,MixedReality/Computer

VisionLab,SchoolofComputerScienceUN U.K.

OtherBeneficiaries2 Imperial College of Science, Technology and

MedicineIC U.K.

3 Centre National de la Recherche Scientifique,TélécomParisTech

CNRS-PT France

4 UniversitatAugsburg UA Germany5 UniversiteitTwente UT TheNetherlands6 CereprocLTD CEREPROC U.K.7 LaCantocheProductionSA CANTOCHE France

ARIA Valuspa


September,2017

-3-


ARIA Valuspa


September,2017

-4-


TableofContents1.PURPOSEOFDOCUMENT 62.DIALOGUEMANAGEMENT 7

2.1.1DIALOGUESTRUCTURE 92.1.2OPERATIONALDIALOGUEMANAGER 122.1.3DECISIONMAKING 132.1.4INTENTPLANNINGFORAGENTBEHAVIOURGENERATION 15

2.2EXAMPLESOFTEMPLATES 162.2.1EXAMPLEMANAGEMENTTEMPLATES 162.2.2EXAMPLEMOVETEMPLATES 17

2.3DIALOGUEENGINE:FLIPPER2.0 192.3.1STANDARDIZATION 192.3.2OPTIMIZATION 192.3.3USABILITYANDERRORHANDLING 192.3.4FEATURES 202.3.5PERSISTENCY 20

3.PROGRESSPERTASK 213.1MULTI-LINGUALNATURALLANGUAGEUNDERSTANDING 213.2TASK-ORIENTEDDIALOGUEMANAGEMENT 213.3USER-ADAPTIVEDIALOGUESTRATEGIES 223.3.1.TURNTAKING 223.3.2VERBALALIGNMENT 23

3.4REINFORCEMENTLEARNINGBASEDONUSERFEEDBACK 243.5DEALINGWITHUNEXPECTEDSITUATIONS 243.6GENERATIONOFDIALOGUESFORBOOKPERSONIFICATIONDEMONSTRATOR 253.6.1QUESTIONGENERATION. 253.6.2FOLLOW-UPQUESTIONSTRATEGY 263.6.3USERTESTANDNEXTSTEPS 27

3.7GENERATIONOFDIALOGUESFORINDUSTRYASSOCIATEDEMONSTRATOR 284.CONCLUSIONSANDPLANSFORNEXTPERIOD 295.OUTPUTS 30REFERENCES 31

ARIA Valuspa


September,2017

-5-


ARIA Valuspa


September,2017

-6-


1.PURPOSEOFDOCUMENTThisisthelastdeliverableforWorkPackage3:“Implementationofadaptivetask-baseddialogue system”. The leader of this work package is UT, with involvement of thefollowingpartners:UON,CNRS,ICL,UA,CNTandCRPC.Theobjectivesofthisworkpackagearetwo-fold.Thefirstistobuildadaptiveandtask-orienteddialogues inmultiple languagestoassist informationretrieval ingeneral,andforthetwoscenarios(thebookARIAandtheIndustryARIA)inparticular.Thesecondobjective is tobuilda framework for informationretrievalstyledialoguemanagementspecificationthatcanbeusedoutsidetheARIA-VALUSPAprojectforadaptivedialogueconstruction.Thisdocumentdescribesthelastversionofthedialoguemanager,DM2.0,andthestepstakentowardsthisversionofthedialoguemanager.TheDM2.0isthecorecomponentinthe(multilingual)adaptivedialoguetoolboxthatwillbereleasedattheendoftheARIAproject as part of the ARIA-VALUSPA Platform (AVP). We discuss the concepts andworkingsofthetoolbox.Thedeliverableisstructuredasfollows.Section2ofthisdeliverabledescribesDM2.0indetail.Section3describestheprogressonalltasksofthisworkpackagethatwerenotcoveredinSection2.InSection4themainconclusionsarepresentedandplansfortheremainderoftheprojectareoutlined.Section5describestheoutputsofthepastperiod.

ARIA Valuspa


September,2017

-7-


2.DIALOGUEMANAGEMENTInthissection,wedescribethenewversionofdialoguemanagementdevelopedinthelast period of the project (DM2.0).Within dialoguemanagementwe distinguish threeconcepts:

● DialogueManager(Managementtemplates)● Dialogues(i.e.scenariosconsistingofMovetemplates)● DialogueEngine(Flipper2.0)

TheDialogueManager, described in detail in Section 2.1, deals with how the agentbehaves in an interaction. The Dialogue Manager consists of Management Templatesthat, for example, define the organisation of information in the information state. SeeSection2.2.1forexamplemanagementtemplates.Dialogues are used to describe the scenario that the agent knows about and canconverseabout.TheyaredefinedinaDialogueStructureconsistingofMovetemplates,describedindetailinSection2.1.1.Section2.2.2showsexampleMovetemplates.TheDialogueEngine,describedindetail inSection2.3,consistsofthenewversionofFlipper. Flipper checks the dialogue templates that are used to define the DialogueManagementintheARIAagent.

ARIA Valuspa


September,2017

-8-


Figure1:DM2.0Structure

ARIA Valuspa


September,2017

-9-


2.1DIALOGUEMANAGER2.0WehavedesignedanewversionoftheDialogueManager(DM2.0)thattakesascenarioandsituation-drivenapproachtocreatingdialoguestructuresbasedonconversationalacts. It shares some properties with current tools for the development of dialogues,such as the use of dialogue trees fromDISCO (Rich and Sidner, 2012) and the use ofquestion-answermatchingfor informationretrieval fromtheNPCEditoroftheVirtualHumanToolkit (Leuski andTraum,2011). Similar to theFLoReSdialoguemanagerofMorbinietal.(2014),theDM2.0hasbeensetuptofacilitatethecreationofstructureddialogueswiththeuseofdomainexperts.BelowwefirstdescribethewaytheinformationintheDialogueManagerisstructured(the Dialogue Structure, Section 2.1.1) and what information is present in theoperationalsystem(Section2.1.2).Afterthis,wedescribehowthedecisiontoperformspecific agent behaviour is made (Section 2.1.3) and how intent planning is done(Section 2.1.4). The conceptual components of the Dialogue Manager and theinformationflowbetweenthemaredepictedinFigure1.Themodelhasbeendevelopedand implemented inFlipper2.0 (see Section2.3), andcurrentlyafewexampledialogueshavebeenauthored.DM2.0willbeusedasthebasisforthedevelopmentoftheindustryARIA(Section3.7).

2.1.1DIALOGUESTRUCTUREDialoguesaredefinedintermsofhierarchicaldialogueacts inaDialogueStructure,asshowninFigure2:

● Dialogue structure: this is the root. The name should refer to the name of thecharacter.

● Episode:thefirsttierconsistsofepisodes.Anepisodecoversaphase(oraverygeneral topic)of a conversation (e.g. social,Q&A, readingalongwith theagent,unexpectedsituations)

● Episode.Exchanges: episodes are made up of exchanges that are related to aspecifictopicintheepisode(e.g.twoormoreutterancesaboutthesametopic)

● Episode.Exchanges.Moves: the lowest tier consists of moves. An exchange ismadeupofseveralmoves.Thesecorrespondtoconversationalacts,basedontheDIT++ standard (Bunt et al., 2012). Amove can be realizedwith an utterance,nonverbalbehaviour,oracombinationofthetwo.Amovehasagoalthatcanbeachieved by the behaviour and a status for that goal. Moves are selected forexecution based on how relevant they are to the situation in the conversation.Rulesdefinewhenamovebecomesrelevant.

ARIA Valuspa


September,2017

-10-


Figure2:Thehierarchicaldialoguestructure.Movesare theatomicunit (dialogue item) in theDialogueManager (DM).Amovecanrefer to amove (a dialogue act) of the user or the agent. Amove can be listening orspeaking, depending to its DIT++ category. We distinguish three types of moves,dependingontheinformationtheycarry:

● Content/DialogueAct(C):C-Movesdealwithinformationexchange.Theycontainthe content of what the agent or user said, for example if the user asked thequestion“whatcanyoutellmeaboutAliceinWonderland?”.

● Interaction Management (I): I-Moves contain information that is about theinteraction, for example turn-management or establishing contact between theuserandagent.TheycorrespondtotheInteractionManagementfunctionsoftheDIT++taxonomyofdialogueacts.

● SocioEmotional(S):S-Movesexpressthesocialandemotionalstateoftheagentor user, for example an increase in dominance or a decrease in sentiment. InDIT++ these are seen as qualifiers of dialogue acts; however we see them asmovessincetheycanhavespecific(oftennonverbal)behavioursassociatedwiththem.

Additionally,thecontentmovesfortheagentarefurtherdivideddependingonthetypeofcontentthatiscontainedwithinthemove.Thisdivisionhelpswithplanningtheagentmoves(forexamplecombiningcontentwithanopinion)andhelpsadialoguescenarioauthortokeeptrackofwhatpurposeamovehas:

ARIA Valuspa


September,2017

-11-


● Content (C-tag): This is the type of C-Move that contains information from thebook, for example what events happened andwho Alice hasmet. An examplewouldbe:“WhenIsawtheWhiteRabbit,Ichasedhimintoarabbithole.”

● Opinion(O-tag):AC-Movewithanopiniontagcontainsanopinion,forexampleAlice’sopinionabouteventsor characters from thebook, suchas: “I thought itwasratherstrangetoseeawhiterabbitwithawatch.”

● Meta information (M-tag): A C-Move with a meta information tag containsinformationonameta levelabout the interaction.Using suchmoves, theagentcantalkabout the interaction, forexampleduringa lull in theconversationtheagentmightsay“Doyouwanttocontinuetalkingaboutthis?”.

TheDMhasbeendesignedwithauthoringinmind.Anauthorcancreatedialogues(i.e.episodes, exchanges, moves of the agent) without needing expert knowledge aboutdialogue systems. In general, an author only needs to specify a situation (e.g. a newunknownuserappears, smiles, andsays “Hello”)andwhatanagent shoulddo in thatsituation(e.g.smileandsay“Hello,nicetomeetyou!”).Amovehasdifferentattributes,someofwhichneedtobefilledinbyanauthor,otherscanbecomputedautomatically.Thefollowingattributesneedtobeauthored:1

● Name: the path in the dialogue structure (e.g. kb1.episode1.exchange1.move1).Wesuggestbasingthenameonthegoalofthedialogueitem.

● AgentGoal:[DIT++].exchangename(e.g.info_request.QATopic)● AgentBehaviour:verbalornon-verbalbehaviour.● ContentType:meta|content|opinion● PreconditionRules(specific):

o Preconditionsthatcontroldialogueflow.Forexample,aftergreetingtheuser,itmakessensefortheagenttosuggesttotheuseranactiontotakeortointroduceatopic.

o SpecificUserUtterancestowhichthemovewouldbeanappropriateresponse.Forexample,theuserutterance“howareyoudoing”isagoodindicatorthatamovethathasAgentBehaviour“I’mdoinggreat,thankyouforasking”isrelevant.

Amovehasthefollowingattributesthatcanbecomputedautomatically(inmostcases)before an interaction. They can also be authored if needed, or computed during aninteraction:

1SomeoftheseattributesapplyonlytoC-moves.

ARIA Valuspa


September,2017

-12-


● GoalStatus:Thisreferstowhetherornotthegoalofthemovehasbeencompletedoraccomplished,andisupdatedduringtheinteractionforeverymove.

● RefersTo:Theowner(agent|user)ofthemove● Relevance:Avalue[0..1]calculatedduringtheinteraction.● AgentForwardGoal:[DIT++].exchangename(e.g.attention.QATopic).Thisisthe

agent’splanneddialogueactthatisusedtoplanthenextmovesandcanbeautomaticallydeterminedusingadjacencypairsandfloormanagementorturn-takingrules.Forexample,foramovewithasgoalaninformation_request(i.e.aquestion),theforwardgoaloftheagentwouldbeattention(i.e.listentotheanswer).

● UserExpectedGoal:[DIT++].exchangename(e.g.i.QATopic).Thisiswhattheagenttheuser’scurrentdialogueactwillbeandcanbeusedtomakeadialogueplan.Forexample,whentheagentistalking,weexpecttheusertobequiet.

● UserForwardExpectedGoal:[DIT++].exchangename(e.g.inform_answer.QATopic).Thisgoaliswhattheagentthinkstheuser’snextdialogueactwillbe.Thiscanbeusedtocreateadialogueplan.Forexample,aftertheagenthasaskedaquestion,weexpecttheusertostarttalking.

● PreconditionRules(general):o Generalpreconditions(rulesthatapplytomostmoves):Forexample,it

doesn’tmakesensefortheagenttotalkwhennouserispresent.o SituationSpecificpreconditions(rulesthatapplytosomemovesorto

movesinsomecase).Forexample,itdoesn’tmakesensefortheagenttospeakrightafterithasaskedaquestion.

AnexampleofaMoveTemplate,withexplanation,canbefoundinSection2.3.2.

2.1.2OPERATIONALDIALOGUEMANAGERTheDialogueEngine(Flipper)storesallinformationtheagentknowsintheinformationstate.InformationcomesfromvarioussourcesandisrepresentedintheformofMoves.During an interaction, themoves of the user are created by the system via the InputProcessingcomponent(seeFigure1).Someexamplesofusermovesare:

● The user has started speaking, detected by theVoiceActivityDetection: an I-type usermove is generated stating that the user has started speaking (andpossiblythatthisisaninterruptionwhentheagentisalsospeaking).

● The user has spoken, and the automatic speech recognition (ASR) outputs awordstring:aC-typeusermoveisgeneratedholdingtheASRoutput.

● Theuserhasstartedsmiling,theSSIupdatesthevalenceoftheuser’saffectivestate:anS-typeusermoveisgeneratedholdingthevalence.

ARIA Valuspa


September,2017

-13-


Additionally, information about theagent’s actions is received as input (i.e. feedback)from the behaviour realiser. The behaviour realiser (e.g. Greta) sends continuousfeedbackaboutwhatbehaviourhasbeencarriedout.Feedbackcanbe:

● BMLCallbacks:theBMLrealisersendsinformationaboutwhichbehaviour(BMLblock)hasstarted,ended,orhasbeenstopped.

● TimeMarkersCallbacks:duringtheagentbehaviour,therealisersendsfeedbackontheexacttimingofeachbehaviourthat isexecuted.This isdoneusingtime-markers(seeSection2.2.4).Foragentutterances,thisisdoneonwordlevel.

ThefeedbackallowstheInputProcessingtokeeptrackofthefloor(i.e.turn-taking)andthecompletionofthegoalsoftheagent.Forexample,knowingfromthefeedbackwhentheagenthasstoppedspeakingandknowing fromtheASRwhen theuserhasstartedspeakingallowsustodeterminewhetherthereisoverlappingspeechandthuswhethertheuser interrupted the agent.Additionally, the timemarkers allowus toknowwhatpart of the agent utterance has been said uninterrupted (and thus was heard by theuser) and what part was not heard because it was interrupted. The agent mightconcatenatemoves, forexampleaC-taggedC-move (“TheWhiteRabbithadawatch”)with an O-tagged C-move (“I liked that watch”). Time marker feedback is used todetermine whether the goal of the agent’s move was accomplished: if the agent wasinterrupted before it could complete the utterance, the goal of the move (e.g. ofconveying this information) is not accomplished. Thismightmean that the agentwillrepeatitself.

2.1.3DECISIONMAKINGMoveshaverulesthatdeterminewhenthemovebecomesrelevant.Theagentselectsitsmovesbasedontheirrelevance.Weviewrelevanceastheutilityvalueofamove,wherethe agent is trying tomaximize the utilities ofmoves and selects themoveswith thehighest relevance above a certain threshold. This threshold is dynamic anddecreaseswhenforacertainamountoftimenomovehasbeenperformedandincreaseswhentheagentisspeaking.TherelevanceofamovegetsupdatedbytheAgentMoveUpdater.Relevanceisbasedontherulesinthemove.Whenarule(definedinthepreconditionsoftheFlippertemplate)ismet,therelevanceofthemoveincreases.Additionally,therelevanceofamovegoesupwhentheuserutterancematchesoneoftheutterancestowhichthismovewouldbean appropriate response, as predefined in the move. This is an extension of the QA(question-answer)Matcher introducedinthepreviouslydeliveredversionof theARIAsoftware (see also Section 3.6). Furthermore, relevance of amove increases if closelyrelated moves (e.g. moves in the same exchange) become more relevant. We useManagementtemplatesinFlippertoupdatetherelevanceofthedialoguestructure.

ARIA Valuspa


September,2017

-14-


TheAgentMoveSelectorkeepstrackof therelevanceofall themoves intheDialogueStructure.Itselectsthemovewiththehighestrelevanceabovethethresholdandsendsthismove to theMove Planner. Additionally, it sends the selectedmove to theMoveGenerator(seeFigure1)forexecutionbytheagentembodiment(seeSection2.2.4).TheMovePlannerkeepstrackofthecurrentagentmoveandtheplannedagentmove.ItgetsinformationfromthemoveselectorandtheInputProcessingmodulesforobservedand predicted user moves. The agent keeps track of five moves at any time for thedecision-makingprocess:

● Thecurrentagentmove.Thisisthemovewhichtheagentiscurrentlycarryingout,forexampleaskingaquestionorshowingattentivelisteningbehaviour.

● The planned agent move. This represents the agent’s forward-looking move.Theagenthasapredictionofwhatitsnextmovemaybeandputsthismoveinitsbehaviourplan.Thisisplannedbasedonthecurrentagentmove,usingadjacencypairs.Forexample,theagentwillmostlikelyperformlisteningbehaviourafterithasaskedtheuseraquestion.

● Theexpectedusermove.Basedonthecurrentagentmove,weexpecttheuserto behave in some way during the agent’s move. Expectations can be based,among other things, on turn-taking rules (e.g. one speaker at a time). Forexample,whenwearetalkingweexpecttheusertolisten.

● Theobservedusermove.This iscreatedviatheInputProcessingmodule.Forexample,ifinthepreviousturntheagenthasaskedaquestionandvoiceactivityisdetected,itismostlikelythattheuserisnowinformingusoftheanswer.

● Theexpecteduser’snextmove.Thisexpectationofwhattheuserwilldonextiscreated based on the agent’s current move and the observed user move. Anexampleiswhentheagentisaskingaquestion,observesthattheuserispayingattention,andexpectstheusertogiveananswerinthenextturn.

The Move Planner checks whether the plan is still correct: are the expected andobservedusermove the same?When this isnot the case, a re-plan isneededand theMovePlannerrequeststheAgentMoveUpdatertore-evaluateallmoveswiththenewsituation.Summarizing,decidingwhatmoveoftheagentshoulddoisdonebythreecomponents,seeFigure1:

● TheAgentMoveUpdaterusesall theavailable information in the informationstate toupdate the relevanceof each agentmove in thedialogue scenario (seeSection2.2.1)whentheMovePlannercallsforanupdate.

● TheAgentMoveSelectorselectsthemostrelevantmove,aftertheAgentMoveshave been updated and one ormoremoves exceed the relevance threshold. It

ARIA Valuspa


September,2017

-15-


sendsthemostrelevantmovetotheMovePlanner,andpreparesforitsexecution(seeSection2.2.4).

● The Move Planner keeps a plan of the moves that have been selected forexecutionbytheagentanddetermineswhichmovesareexpectedfromtheuser.Whentheplanisfinished,ortheexpectedusermoveisnotobserved,theMovePlannercallsforanAgentMoveUpdate(whichconstitutesare-plan).

2.1.4INTENTPLANNINGFORAGENTBEHAVIOURGENERATIONOnceanagentmovehasbeenselectedandputintheMovePlanner,theMoveGeneratortranslates this move to FML-APML. First, the agent’s verbal utterance (if present) isextracted from the selected move, and time markers are added to it. Secondly, theemotionoftheagentisset,basedonthecurrentemotionalstateoftheagentstoredintheinformationstate.Furthermore,additionalparameters(e.g.backchannel,stance)inthemove can be used to fill the placeholders in the FML-templates. Finally, we haveimplementedanaturallanguagegenerationmodulethatcanadapttheverbalcontentofan agent move to align it to the user’s word choice. The module takes the dialoguehistory into account and tries to align the agent’s utterance to the user’s utterances.MoredetailsonthismodulecanbefoundinSection3.3.

ARIA Valuspa


September,2017

-16-


2.2EXAMPLESOFTEMPLATES

2.2.1EXAMPLEMANAGEMENTTEMPLATESHereweshowtwoexamplesofmanagementtemplatesinFlipper2.0.ThefirstexampleisatemplatethatpassesauserutterancetotheQAmatchingmodule,whentheuserhasmadeacontentmove:<template id=”001” name=”answerQuestion> <preconditions>

<condition>is.states.user.lastMove.type == c </condition> </preconditions> <effects> <method persistent=”is.states.modules.QAMatcher” is=”is.states.agent.say.answer” class=”eu.aria.dm.modules.QAMatcher” name=”findBestAnswer”> <arguments> <value class=”String” is=”is.states.user.say.utterance” </arguments> </method> </effects> </template>

The second example is a template that makes the agent listen after it has asked aquestion.Thistemplate ispartofa largercollectionof interactionstatetemplatesthatindicatewhentheagentshouldperformalisteningortalkingbehaviour: <template id=”101” name=”updateInteractionState> <preconditions> <condition>is.states.agent.say.askedQuestion</condition> </preconditions> <effects> <assign is=”is.states.agent.interactionState”>listen</assign> </effects> </template>

ARIA Valuspa


September,2017

-17-


2.2.2EXAMPLEMOVETEMPLATESBelowweshowanexampleofapossibledialoguemovebytheagent. Inthis instance,theagentmakessuggestionsforatopic.

{"Episode":"Goal":"Social",{"Exchange": "Goal":"TopicManagement" {"Move":

"Goal":"Agent_topic_suggestions", "UU":[

"Whatcanyoudo?", "Whatcanyoutellme?", "Whatelsecanyoutellme?", "Doyouknowotherthings?" ], "Rules":"silence>5||

QA.$prev_exchange==completed&&prev_episode==QA",

"AB":[ "Well,I'dliketotalkabout$!prev_topic.OK?", "Whatdoyouthinkabout$!prev_topic?", "Doyouknow$!prev_topic?" ], "Type":"M"

}}}

Thismove is part of episode “Social”, and part of exchange “TopicManagement”. Themove is called Social.TopicManagement.Agent_topic_suggestions (which is aconcatenation of the goals of the hierarchy). UU refers toUserUtterances thatwould

ARIA Valuspa


September,2017

-18-


makethismoverelevant(i.e.towhichitwouldbeanappropriateresponse).TheRulesrefer tosituations inwhich thismove is relevant. In this case,either therehasbeenasilence that lasted longer than 5 seconds, or the previous exchange was part of theepisode “QA” and it has been exhausted (i.e. all moves in this exchange have beenexecuted).ABreferstotheAgentBehaviourthatwouldbeappropriatebehaviourinthegiven context. The Type refers to what sort of content thismove has, in this caseM(MetaInformation):thismovetalksabouttheconversationonametalevel,inthiscasedeterminingwhattotalkabout.Anothertypeofepisodewehavedesigned,andwhichcanbereusedinmanydifferentdialoguescenarios,iscalled“SafetyZone”.Inthisepisode,theagenttriestorestoretheconversation after it has broken down, for example because the agent does notunderstandtheuserorbecausenouseractivityisdetected.Theagentwillstay`safe’inthisepisodeuntilitdiscoversarulethatmovesittoanotherepisode.TheSafetyZoneisdesigned to keep the user engaged and to prevent the agent from making repeatedstatementssuchas“I’msorry,Idon’tknow”or“I’msorry,Idonotunderstand”.Anexampleofamoveinthisepisodeisshownbelow.Whenevertheagentdoesnotseetheuseranymore,ortheuserhasbeenlookingawayforsometime,theagentwilltrytore-establishcontact.{"Episode":"Goal":"SafetyZone",{"Exchange":"Goal":"Contact" {"Move": "Goal":"reestablish_contact", "UU":[], "Rules":["face_presence=false", "agent_turn=true", "engagement=true", “history=true” ],“AB”:“Hello,stillthere?” }}}

ARIA Valuspa


September,2017

-19-


2.3DIALOGUEENGINE:FLIPPER2.0TocheckthedialoguetemplatesthatdefinetheDialogueManagementintheARIAagentwe make use of Flipper (ter Maat and Heylen, 2011; see also D3.2). Flipper is aninformation-statebaseddialogueenginethatupdatesthedialoguecontextanddecideson which action the agent should take, similar to the implementation of FLoReS(Morbini et al (2014). In the past months, we have developed a new version of thedialogue engine (Flipper 2.0). This new version improves the original in severalrespects.Tomakeitmorerobust,versatileandeasytouse,thefollowingnewfeatureshavebeenaddedinFlipper2.0:

● Standardization(JSON,JavaScript)● Optimization● Usabilityanderrorhandling● Features● Persistency

2.3.1STANDARDIZATIONWehavechanged the representationof the information state froma customdesignedJava-model to JSON, hence making it easier for different modules of the DialogueManagertoretrieveinformationfromtheinformationstate.Furthermore, we have replaced Flipper’s old system for evaluating templatepreconditions by JavaScript evaluation, making it possible for users to use both self-written JavaScript or load JavaScript libraries.We kept the XML format as similar aspossibletothepreviousFlipper,becausewefoundthatXMListhebestreadableformatfornon-technicalpeoplewhostillwanttowritecustomtemplates.Finally, input modules send user input in standardized formats (XML or JSON). Thismakesiteasiertoputtheuserinputintotheinformationstate.

2.3.2OPTIMIZATIONThepreviousversionofFlippermadevery frequentuseofparsingandevaluationbutdid not use many standard libraries for this. We have upgraded this to run moreefficiently(lazyevaluationanduseofstandardlibrariesforparsingXMLandJSON).Thisreducedthecodebyatleast30%andmadeitalsomorereadableformaintenance.

2.3.3USABILITYANDERRORHANDLINGThetemplatesofthefirstversionofFlippercanstillbeusedinthenewversion,astheversionsarebackwardscompatible.Thecomponentsusedbyaparticulartemplate(e.g.partsof the informationstateor functionscalled)arealwaysdefinedorreferredto inthetemplates,sothattheauthoronlyneedstocheckaparticulartemplate.Hedoesnot

ARIA Valuspa


September,2017

-20-


have to search through all the code and templates to find out what the impact is ofchangingsaidtemplate.Error handling has been implemented with precise notifications if something wentwrong, making it easier for other developers to find errors in their connected Java-modulesorXML-templates.Also,itispossibletoretrievecertainsystemparametersforcheckingwhichtemplatehasbeenused,whichisusefulfortestingpurposes.

2.3.4FEATURESIn the first version of Flipper, it was not easy to use external modules (e.g. naturallanguageunderstandingandgeneration)andallreasoningandlogicfortheagentwasinthetemplates.Wecannowcreatetemplatesthatcanuseexternalmodulestodomorecomplexcalculationswithout cluttering the templates,whichshouldonlydescribe thedialoguestructureonahigherlevel.Asmentionedbefore,JavaScriptisnowapartofFlipper2.0.Thisfeatureimprovestheflexibility of defining preconditions and manipulating the information state.Furthermore,existingJavaScriptlibrariescanbeaddedeasily.

2.3.5PERSISTENCYInFlipper2.0,theinformationstatecanonlybechangedfromwithinFlippertemplates,no longer from outside classes. A database connection has been made possible withPostgreSQL.We commit and store the information state in this database, so that theinformationstatecanberetrievedorcheckedwhenithaschanged(e.g.forcheckinglogsof agent behaviours) or be restored when data becomes corrupted (e.g. when theInternetconnectionfails).Thisiscalledpersistency.Itmakesthesystemmorerobusttounwantedbehaviour.Becausealltheinformationthat is intheinformationstateatsomemomentduringaninteractioniskeptinthedatabase,wecoulduseittorestorethedialoguetoapreviousstate,forexample,thestatewhenthecurrentuserhadapreviousconversationwiththesystem.Thismakesitpossiblefortheagenttorefertothispreviousconversationorfortheuser and agent to continue their conversationwhere they left off. (Implementingthistypeofbehaviourisfuturework.)Furthermore,wecanstorevariouskindsofbackgroundknowledgeinthedatabase.Wecanuseittostorealllong-terminformation,comparedtoshort-terminformationintheinformation state. For example, the full dialogue history and all previous userdemographicinformationcanbestoredinthedatabase.

ARIA Valuspa


September,2017

-21-


3.PROGRESSPERTASKIn this section, we discuss the progress per task thatwas not covered earlier in thisdocument.

− Task3.1,Multi-lingualnaturallanguageunderstanding− Task3.2,Task-orienteddialoguemanagement− Task3.3,User-adaptivedialoguestrategies− Task3.4,Reinforcementlearningbasedonuserfeedback− Task3.5,Dealingwithunexpectedsituations− Task3.6,GenerationofDialoguesforBookPersonificationdemonstrator− Task3.7,GenerationofDialoguesforIndustryAssociatedemonstrator

3.1MULTI-LINGUALNATURALLANGUAGEUNDERSTANDINGThistaskconcernstheextensionoftheARIA-VALUSPAagent'sshallownaturallanguageunderstanding skills to handle the three natural languages targeted by the project:English,FrenchandGerman.ThreedifferentAutomatedSpeechRecognition(ASR)moduleshavebeendevelopedtorecognizespeechinEnglish,GermanandFrench.Recognizedwordstringsareputintheinformation state, and the Input Processing component of the dialogue system isinformedofthenewuserutterance.TheInputProcessingcomponentthentriestomaptheuserutterancetotheuserutterancesdefinedinthemoves(seeUUintheexampletemplate from Section 2.3.2), which are currently only in English. To make thismultilingual,theutterancesofboththeagentanduserwillbetranslatedtotheforeignlanguages.Themappingofrecognizeduserutterancestouserutterancesspecifiedinthetemplatesis currently done in a language-independent fashion by measuring unigram overlapbetweenthestrings;itdoesnotinvolvesyntacticorsemanticparsingandthereforedoesnot require any language-dependent resources. Therefore simply translating thetemplates is sufficient to have dialogues in French or German in addition to English.Future users of the system can author the dialogue templates in their preferredlanguage,providedthatanASRcomponentandaTTScomponentareavailableforthatlanguage.Fornon-verbalbehaviour,wemakethesimplifiedassumptionthatthereisnodifferenceamongstthedifferentlanguages.

3.2TASK-ORIENTEDDIALOGUEMANAGEMENTThis task involves the implementation of an information state-based architecture fordialoguemanagementthatcanrundifferentdialoguemanagementdimensionsinparallel, focusing on interaction management aspects such as turn taking, repairmechanisms,andfloormanagement.TheworkinthistaskhasbeenextensivelycoveredinSection2.

ARIA Valuspa


September,2017

-22-


3.3USER-ADAPTIVEDIALOGUESTRATEGIESAspecialnewfeatureoftheARIA-VALUSPAdialoguemanagementisadaptationtotheuser.Adaptation,oralignment,innaturaldialoguesappearsinmanyaspectsofdialogue.Thistaskinvolvestheimplementationofdifferentadaptivestrategiesthatarerelevanttotheapplication.Onesimple formofuseradaptation isbymakinguseof the inputprovidedby theSSIframework.TheSSIcandetectdemographicsofauser(e.g.thegender)andthiscanbeusedbytheagenttocreateanappropriateofresponses,forexample,“HelloSir”versus“Hello Madam”. Other, more advanced forms of adaptation we have focused on areadaptation in termsof turn-takingandalignmentof theagent’swordchoice to thatoftheuser.

3.3.1.TURNTAKINGKnowingwhentheuserspeaksandadaptingtheagent’sbehaviourtothisisimportantfor a smooth dialogue. To make this possible, we developed an Interaction StateManager,whichislocatedintheInputProcessingcomponentofDM2.0(seeFigure1).Itcollects information about a user being present and the current floor and turn-takingsituation in the dialogue. This information is passed to the information state, fromwhereitcanbeusedbytherestoftheDM.TheInteractionStateManagerkeepstrackoftheagent’sinteractionstate:ifitisidle,tryingtoengageordisengage,orifitisengagedwithauser.Additionally,theagentusesamodelofwhentheagentistalking,listening,interrupting,yieldingaturn,orwaiting,showninFigure3.Thissupportsdynamicturn-takingandcanbeusedtoupdatetherelevanceofmovesaswell.

ARIA Valuspa


September,2017

-23-


Figure3:InteractionStateModel,basedontheworkofLoch(2011)

3.3.2VERBALALIGNMENTInadditiontotheverbalalignmentdescribedinSection2.3inD4.3,wehavedevelopedrules foradaptingagentutterancesto thoseof theuser.Thismakes itpossible for theagent to use referring expressions from the book that are preferred by the user. Anexampleistalkingaboutthe‘tinygoldenkey’foundbyAliceinthebook.Wecouldrefertoitintheagentutteranceas’key’,‘goldenkey’or‘tinykey’.Iftheuserusesaparticulardescription,theagentisabletomirrorthis.Toachievethealignment,weadoptatext-to-text generation approach. This means we automatically modify the manuallyspecified agent utterances, instead of generating aligned referring expressions fromscratch,asdoneintheworkonalignmentofreferringexpressionsbyBuschmeieretal.(2009)andGattetal.(2011).First,todeterminetheuser’spreferredexpression,theuserinputisstemmed,part-of-speechtaggedanddividedintochunks.Stemmingandchunkingwerechosenovertheirmoresophisticatedcounterparts(i.e.lemmatizationandfullsyntacticparsing)tomakethe whole process more resilient to potential inaccuracies of the ASR. The sameprocessingisdoneforthesentencedefinedintheFML,i.e.theagentanswer,butinthiscase we can also obtain a syntactic tree. For stemming, part-of-speech tagging and

ARIA Valuspa


September,2017

-24-


parsing theStanfordCoreNLPtoolkit isused,while theTextPro2suitedivides the textintochunks.Wethenproceed toadjusteverynounphrase(NP) in theagentsentencewhoseheadstemmatchesanounintheusersentence,addingandremovingadjectivesandadverbsso that thephraseof theagentmatches thechunkof theuser.Additionalconstraintsmakesure thephrasesarenotmodifiedwhen it is inappropriate todoso,e.g. inexistentialquestions(whentheuserasks,“arethereanycatsinthisstory?”,theagentshouldnotanswer“yes,thereareanycats”)orinsidequotations(whentheagentisreportingsomethingsaidbyothers).BymodifyingeveryNP,wearecreatingasentencethat-byconstruction-maximizesthealignmentwiththeuser.However, theprocessalso involvesanumberof intermediatesteps where not every noun phrase is aligned (overgeneration). This means that,dependingonitsgoal,theagentcanalsochoosedifferentsentences,e.g.shorteronestobeconcise,orthosethatminimizethealignmenttosee if theuserperceivestheagentdifferently.Wearecurrentlyworkingonextendingthisprocesswithsynonymsandantonyms.Forexample,iftheuserreferencesthesameobjectusingdifferentsynonyms,wemighthavetheagentapplying this strategyaswell.Finallyweare in theprocessof settingupanexperiment to see to what extent the difference between unaligned and aligneddialoguesisnoticeablebyusers,andifthiscontributespositivelytothedialogue.

3.4REINFORCEMENTLEARNINGBASEDONUSERFEEDBACKThis task refers to training the system's adaptive dialogue strategies usingreinforcementlearning.Itwasoriginallyplannedforthemonths31-36,butnotupdatedcorrectlywhenthedeliverydateofD3.3wasmovedforwardfrommonth36tomonth31.Thistaskhasthusjuststartedtobeactivelypursued.

3.5DEALINGWITHUNEXPECTEDSITUATIONSThis task concerns enabling the ARIA agents to deal with unexpected situations thatoccurduringaninteraction.Twotypesofunexpectedsituationsthatwehaveaddressedinvolve turn-taking.We have designed amodel for dealingwith interruptions by theuser: the agent can employ re-planning, holding and halt behaviour (see the maindescriptioninD4.3).Additionally,theagentcaninterrupttheuser.Thisisachievedbymoves thatbecomerelevantwhile theuser is speakingandbyusing incrementalASRoutput.Usingmoveswithdynamicrelevanceupdatingallowsustointerrupttheuseratanopportunemoment.Forexample,iftheagentwantstodisagreewithsomethingtheuser is saying, a move with rules describing exactly this situation will get a high2http://textpro.fbk.eu

ARIA Valuspa


September,2017

-25-


relevance. The agent will start talking, even though it has not gotten the turn, thusinterruptingtheuser.Additionally, we have defined an episode called ‘unexpected situations’. This episodecontains exchanges and moves dealing specifically with unexpected situations(ironicallymakingthemexpected).Forexample,whenweobservetheusernotlookingatuswhile theagent is talking, theagent can respond to this situationwitha remark(e.g.“Well,Ithinkitispolitetopayattention…”,seeSection2.2).

3.6GENERATIONOFDIALOGUESFORBOOKPERSONIFICATIONDEMONSTRATORThis task is devoted to the generation of dialogues for the book personificationapplication.Theaimis tocreateasetofdialoguestructures thatcovers thethemes inthe book and avoids open domain conversation. A first version of the bookpersonificationdemonstratorwasdescribed inD3.2.Since then, thedemonstratorhasbeenextendedwithquestion-answer(QA)matchingfunctionalitycomparabletothatofthe VH Toolkit (Hartholt et al, 2013). A question by the user ismapped to themostsimilarquestionfoundinadatabasewithquestion-answerpairs,andthecorrespondinganswerisreturned.CurrentlythedemonstratorusesacombinationofQAmatchingfunctionalityandDM1.0templates(tobereplacedwithDM2.0episodesandmovesasdescribedinSection2.1).TheDM2.0versionofthedemonstratorwillbedeliveredlaterthisyear.

3.6.1QUESTIONGENERATION.To easily expand the range of user questions that the Book Personification agent cananswer through QAmatching, we have developed a question generation system thattakes text as input and generates a large number of QA pairs from it. This form ofquestiongenerationcanbeusedtocomplementorreplacedatacollectionwithhumanusers, which we have used in the past to populate the QA database (see D3.2). Forexample, based on the input sentence (from a summary of Alice’s Adventures inWonderland) “Her giant tears forma pool at her feet” the followingquestions canbegenerated: “What happens at her feet?” Or: “What happens to her giant tears at herfeet?”Theanswertobothquestionsis“Hergianttearsformapool”.Most question generation systems are used in educational applications, such as skilldevelopment assessment and knowledge assessment. Applications in conversationalcharactersarerare.OneoftheexceptionsistheworkofYaoetal.(2012).Theirquestiongeneration approach is based on syntactic transformation from declarative sentencesintoquestions.AssourcetextsforquestiongenerationtheyusedWikipedia.Incontrast,our approach isbasedon semantic information (semantic role labels augmentedwithdependencystructures)andweusenarrativetextsasasource.

ARIA Valuspa


September,2017

-26-


So far,wehavebeenusingasmallcorpusofonlinesummariesofAlice’sAdventures inWonderland as source texts for question generation, but note that our approach issufficiently general to take all kinds of texts as source material. We have usedsummaries instead of the book itself, to avoid potential problems with characterdialogueintheformofdirectspeechandthenarrativestyleusedinthebook.OurQuestionGenerationsystemhasbeen inspiredby theworkofMazidiandNielsen(2014;2015).Weusetwotypesofinformationaboutasentenceasabasisforquestiongeneration:(1)thesemanticrolelabelsassignedtothesentenceand(2)thedependencystructure of the sentence. Semantic role labeling parses a sentence into a predicate-argumentstructurewithconsistentargumentlabels.FollowingMazidiandNielsen,weusethetoolSENNA(Collobertetal.,2011)toproducethesemanticrolelabelsforeachinput sentence. SENNA produces a separate predicate-argument structure for eachclause in the sentence. For determining the sentence’s dependency structure, wecurrently use PyStanfordDependencies,3 a Python interface for parsing to StanfordDependencies.The firststepofquestiongeneration involvesmatchingthesemanticrole labels in theinput sentence to a set of previously established patterns. For instance, our examplesentenceabovematchesthepatternA0-V-A1-LOC,whereA0correspondsto theagentrole,A1thepatiensrole,andLOCisalocativemodifier:Hergianttears(A0)form(V)apool (A1) at her feet (LOC). In the next step, based on the detected patterns (wherenecessary combined with dependency information), question and answer pairs areformed from each clause. Currently, we distinguish around 20 different patterns forquestiongeneration; thesearebasedonthesemanticrolecombinations thatoccurredmost frequently in our corpus of summaries. These patterns have been refined andimprovedaftertestswithnew(previouslyunseen)inputdata.Ingeneral,aQApaircanbecreated ifasentencehasapredicate (V), twoormoreverbargumentsandzeroormore modifiers: temporal (TMP), locative (LOC), adverbial (ADV) or Manner (MNR).Sentenceswith fewer than two basic arguments are currently excluded, because theytendtoberelativelyuninformative,andthusnotagoodbasisforquestionsandanswers.On average, the question generation system produces 4 QA pairs per sentence in theinputdocument.

3.6.2FOLLOW-UPQUESTIONSTRATEGYAnimportantchallengewithusingquestiongenerationforconversationalagentsisthatthere is a high probability of mismatches between the generated questions and theactualquestionsthattheusersask(Yaoetal.,2012).Moreover,aninherentlimitationofthe question generation approach is that it can only generate questions towhich theanswers can be found in the source text (the book or a summary of the book). Toaddressthesechallenges,wedevelopedaninnovativedialoguestrategythatnudgesthe

3https://github.com/dmcc/PyStanfordDependencies

ARIA Valuspa


September,2017

-27-


users towardsasking follow-upquestions that canbematched to thequestions in theQAdatabase.ToincreasethelikelihoodthattheuseroftheARIAsystemwillactuallyaskthistypeofquestions,wedesigned and implemented the following strategy.The systemdoesnotjustpresentananswer,basedonaspecificsentenceinthebook,butfollowsitupwithahintabout thenext sentence, toencourage theuser toaska follow-upquestionaboutthis.Thishinttakestheformoftheclue"thensomethinghappens"plusaphrasefromthe next sentence (if applicable). The following example is an excerpt from an actualdialoguethattookplaceduringourusertest,whichisdescribedbelow:

Agent:Shefollowstherabbitdownthehole.ThensomethinghappenswhenShefallsforalongtime...User:Whathappensaftershefallsforalongtime?Agent:Shefallsforalongtimeandfindsherselfinalonghallwayfullofdoors

When the user takes up the hint for the follow-up question and repeats it in full, theuser's question can be easily matched with the corresponding question in the QAdatabase, as shown in the example. However, if the user takes the hint but asks anellipticquestion("Whathappens?") thisquestionmatchesmanydifferentquestions inthedatabase.Topreventincorrectmatches,thesystemexpandsellipticquestionsbasedonthehintithasjustgiven,convertingthemtofullquestionsthathaveadirectmatchintheQAdatabase.Thefollowingdialogueexcerptfromourusertestillustratesthis.Here,the user question “What happen there” is converted to “What happens when thecaterpillarandAlicegetintoanargument”andmatchedcorrectlytothecorrespondinganswer.

Agent:Shemeetsacaterpillarsittingonamushroomandsmokingahookah(i.e.,awaterpipe).ThensomethinghappenswhenThecaterpillarandAlicegetintoanargument...User:Whathappenthere?Agent:ThecaterpillarandAliceget intoanargumentbutbeforetheCaterpillarcrawlsaway indisgust ,he tellsAlice thatdifferentpartsof themushroomwillmakehergrowandshrink.ThensomethinghappenswhenAlicetastesapartofthemushroom...

3.6.3USERTESTANDNEXTSTEPSThe follow-up question strategy was implemented and tested in a small-scale userexperimentinwhichfourusersinteractedwithastandaloneversionoftheQAmatcher,withtypedtextinputandoutput.TheQAdatabaseusedforthetestconsistedof691QApairs, generated from three summaries of Alice's Adventures in Wonderland. Thedialogues with the test users have not yet been analysed in detail, but the firstimpression is that the follow-up question strategyworkswell to give the users some

ARIA Valuspa


September,2017

-28-


idea ofwhat to ask, and to suggest questions to them that can actually be answered.However,usersalsofrequentlyaskedquestionsthathadnomatchinthedatabase(e.g.,"Who is themad hatter" or "What is the King's name?"). Tomake it possible for thesystemtoanswersuchquestionsaswell,weneedtouseadditionalsourcesforquestiongeneration, e.g., Wikipedia entries about the book and its characters. The questiongenerationapproachcouldbeeasilyexpandedtoincludesuchsources.Inaddition,thequestiongenerationsystemneedstobeextendedtogeneratemultiplevariationsofthesamequestion,toensurethatacorrectmatchcanbefoundevenwhentheuserdeviatesfromthephrasingthatwasusedinthesourcetextandinthesystem’shints.The next step will be to expand the QA database used in the Book PersonificationdemonstratorwiththealreadygeneratedQApairsfromthebooksummaries4,andwithadditionalQApairstobegeneratedfromothersourcesassuggestedabove.Thefollow-upquestionstrategyalsoneedstobeaddedtotheBookPersonificationdemonstrator.Alinktothesoftwarecanbefoundatthebottomofthisdocument.

3.7GENERATIONOFDIALOGUESFORINDUSTRYASSOCIATEDEMONSTRATORThistask isdevotedtotheIndustryAssociateapplication. ItencompassesdefiningtheexactmissionoftheVirtualAssistant,creatingadialoguescenariotoachievethisgoal,andbuildingadaptiveandtask-orienteddialoguesinmultiplelanguages.Due to legal issues between the industry partner and the project, there has not beenmuchprogressonthistask.These issueshavebeensortedoutonlyveryrecently,andwearecurrentlyawaitingtheindustrypartner’sinputonthedesireddialoguessothatwecanstartimplementingtheseintheDM2.0.

4BeforetheycanbeusedintheBookPersonificationdemonstrator,thegeneratedquestionsandanswersstillneedtobeconvertedfromthirdtofirstandsecondpersonrespectively,whenreferringtoAlice.

ARIA Valuspa


September,2017

-29-


4.CONCLUSIONSANDPLANSFORNEXTPERIOD4.1CONCLUSIONSWehave developed a new dialogue engine, Flipper 2.0, and designed a new dialoguemanager (DM2.0) thatwe are developing around this system. This dialoguemanagerprovides a dialogue structure based on theDIT++ standard, thatmakes it possible tospecify domain-specific as well as domain-independent dialogue moves that covermultipleconversationaldimensions.ThefullDM2.0isnotyetavailable,butisplannedtobemadeavailablewithinthenextperiod,includingdialoguesforboththebook-ARIAandindustry-ARIA.Wewillprovideanauthoringtoolfordevelopingdialoguesforotherdomainsaswell.Inadditiontoourworkonthedialoguemanager,wehavedevelopednewtechniquesforthe agent’s verbal alignment to the user and for populating a QA database forinformationretrievalwithautomaticallygeneratedquestion-answerpairs.4.2PLANSFORNEXTPERIODThemostcrucialtaskinthefinalmonthsinthisprojectwillbefinishingtheDM2.0(i.e.implementing themoves, linking the behaviours, finishing the authoring tool) and itsdocumentation. This will allow others to use the software to create their own ARIAagents.Wewill finish the implementation anddefinitionof dialogue scenarios for thebook-ARIAandtheindustryARIA.ThesewillbegoodshowcasesofthesystemanditscapabilitiesandtheycanhelpotherstounderstandhowtocreatetheirownARIAs.Theopen-sourcenatureofthisprojectmeansthatinterestedpartieswillbeallowedtocreateanARIAontheirownaftertheprojectwithoursoftware.Thedeliveredsoftwarecomeswithsufficientdocumentationtoallow industrypartners(andexternalparties)tocreateacompellingconversationalagent.Authoring of dialogues will be easier with the ARIA DM2.0 than it is with other DMsystems,becausewherepossibleanauthoronlyneedstodescribeasituationandwhattheagentshoulddo in thatcontext.Tomakeauthoringevensimpler,wehavestartedwiththedevelopmentofavisualsoftwaretoolforauthoringofmoves.Onthepartofreinforcementlearningwewillusetherelevanceofamoveasabasisforlearning.Wewillsetupasmallexperiment inwhichweshowagent-humandialogues(or dialogue fragments) to observers, who have to rate the quality of the agent’sresponses.Thesescoreswillbeusedtolearnabetterrelevanceupdatepolicy.Already clear is that the system will continue to be developed after the project hasended.Wewilladdatopicplanner,whichwillcalculatepossibletopic transitionsthattheagentcanusefortalkingabouttopicstheagentwantstotalkaboutortopicstheuser

ARIA Valuspa


September,2017

-30-


findsinteresting.Currently,topicrecognitionindialoguesoftenworksontheutterancelevel(selectthenounphraseinthesubjectorobjectposition),butitcouldbeextendedtodeterminethetopicofadialoguesegment(Rats,1996;Mei,BansalandWalter,2017).Tolearnhowanagentshouldperformsegment-leveltopicrecognition,wewillsetupanexperiment inwhichobservers indicatetopics inthedialoguesofthecorpus.Withtheannotation of the corpus on topic transitions, we will investigate cues for topictransitionsandcreateatopicmanagementmodulebasedonouranalysis.Finally, we want to have users talking to a version of Alice that contains topicmanagement and can recognize higher-level topics and using human-like topictransition strategies. To evaluate these dialogues, observers will rate the recordedconversationsusingmetricssuchashuman-likenessandcompetence(Glas,2015).

5.OUTPUTSThe outputs with pertinence to this deliverable that have been published (or are inpress)orhavebeenotherwisedisseminatedsofarinthisproject.Publications:Bowden, K., Nilsson, T., Spencer, C., Cengiz, K., Ghitulescu, A. and vanWaterschoot, J.(2017) I Probe, Therefore I Am: Designing a Virtual Journalist with Human Emotions.Proceedingsofthe12thSummerWorkshoponMultimodalInterfaces(eNTERFACE’16),pp.47-53,July18–August12,2016Enschede,TheNetherlands.Cafaro, A., Bruijnes, M., van Waterschoot, J., Pelachaud, C., Theune, M., & Heylen, D.(2017).Selecting andExpressing Communicative Functions in a SAIBA-CompliantAgentFramework.InproceedingsofIntelligentVirtualAgents2017.LinaFasya,E.(2017).Automaticquestiongenerationforvirtualhumans.Masterthesis,UniversityofTwente,August2017.Kolkmeier,J.,Bruijnes,M.andReidsma,D.(2017)AdemonstrationoftheASAPRealizer–Unity3DBridge forVirtualandMixedRealityApplications. Inproceedingsof IntelligentVirtualAgents2017.vanWaterschoot,J.andTheune,M.(2017)Topic-BasedPersonalizationofDialogueswitha Virtual Coach. In proceedings of the workshop on Persuasive Embodied Agents forBehaviorChange,atIntelligentVirtualAgents2017.Posters:

ARIA Valuspa


September,2017

-31-


vanWaterschoot,J.(2017)Topicrecognitionandmanagementinconversationalagents.Poster presented at the Young Researchers’ Roundtable on Spoken Dialog Systems,August13-14,Saarbrücken,Germany.Workshops:AgentsinPractice-DesigningforDialogues.SIKSworkshop/tutorial.2017,MarchWorkshop on Conversational Interruptions in Human-Agent Interactions. At IntelligentVirtualAgents,2017.Invitedtalks:Heylen, D.K.J. Invited Talk at Workshop on Conversational Interruptions in Human-AgentInteractions.AtIntelligentVirtualAgents,2017.Heylen, D.K.J. A Sentimental Journal. Emotions in Daily Life. Invited talk at the 3rdInternational Workshop on Emotion and Sentiment in Social and Expressive Media(ESSEM 2017): User Engagement and Interaction. At the International Conference onAffectiveComputingandIntelligentInteraction,2017.Software:Flipper2.0:https://github.com/hmi-utwente/Flipper-2.0Alice-QuestionAnswerGeneration:https://github.com/evania/alice-qgFullAria-ValuspaPlatform(AVP):https://github.com/ARIA-VALUSPA/AVP

REFERENCESH.Bunt,J.Alexandersson,J.W.Choe,A.C.Fang,A.C.,K.Hasida,V.Petukhova...andD.R.Traum. ISO 24617-2: A semantically-based standard for dialogue annotation. InLRECpp.430-437,2012.H. Buschmeier, K., Bergmann, and S. Kopp. An alignment-capable microplanner forNaturalLanguageGeneration. InProc.12thEuropeanWorkshoponNaturalLanguageGeneration(ENLG’09)pp.82–89,2009.R. Collobert, J. Weston, L. Bottou, M. Karlen, K. Kavukcuoglu, and P. Kuksa. Naturallanguageprocessing(almost) fromscratch. JournalofMachineLearningResearch,vol.12,pp.2493–2537,2011.A. Gatt, M. Goudbeek and E. Krahmer. Attribute preference and priming in referenceproduction: Experimental evidence and computational modeling. Proceedings of theAnnualMeetingoftheCognitiveScienceSociety,vol.33,pp.2627-2632.2011.

ARIA Valuspa


September,2017

-32-


N.Glas.Engagementdriven topic selection for an information-givingagent.WorkshopontheSemanticsandPragmaticsofDialogue,SemDial2015,Gothenburg,Sweden.A.Hartholt,D.Traum,S.C.Marsella,A.Shapiro,G.Stratou,A.Leuski,L.-P.MorencyandJ.Gratch, All Together Now: Introducing the Virtual Human Toolkit. In InternationalConferenceonIntelligentVirtualHumans,Edinburgh,2013.A.LeuskiandD.Traum.NPCEditor:CreatingVirtualHumanDialogueUsingInformationRetrievalTechniques,InAIMagazine,volume32,2011.F.Loch.ComputationalModels forTurnManagementusingStatecharts.MasterThesis,UniversityofTwente,2011M. ter Maat and D.K.J. Heylen. Flipper: An Information State Component for SpokenDialogue Systems. In: Vilhjálmsson H.H., Kopp S., Marsella S., Thórisson K.R. (eds)Intelligent Virtual Agents. IVA 2011. Lecture Notes in Computer Science, vol 6895.Springer,Berlin,Heidelberg,2011.K.MazidiandR.D.Nielsen.Linguisticconsiderationsinautomaticquestiongeneration.IntheProceedingsofACL2014,pp.321–326.K.Mazidi and R. D. Nielsen. Leveragingmultiple views of text for automatic questiongeneration. In:Proceedingsof the InternationalConferenceonArtificial Intelligence inEducation(2015),pp.257–266.H.Mei,M.BansalandM.R.Walter.CoherentDialoguewithAttention-BasedLanguageModels.In:AAAI.2017.p.3252-3258.F.Morbinietal."FLoReS:aforwardlooking,rewardseeking,dialoguemanager."Naturalinteractionwith robots, knowbots and smartphones. Springer, New York, NY, 2014. pp313-325.C. Rich and C. Sidner. Using collaborative discourse theory to partially automatedialogue tree authoring. In Intelligent Virtual Agents (pp. 327-340). SpringerBerlin/Heidelberg,2012X.Yao,E.Tosch,G.Chen,E.Nouri,R.Artstein,A.Leuski,K.Sagae,andD.Traum.Creatingconversationalcharactersusingquestiongenerationtools.Dialogue&Discourse,vol.3,no.2,pp.125–146,2012.

Documents

D3.3-final multilingual adaptive dialogue toolbox...Deliverable (D3.3). Release of the multilingual, adaptive dialogue toolbox 1. PURPOSE OF DOCUMENT This is the last deliverable for