308 Handling Uncertainty during Plan Recognition in Task-Oriented Consultation Systems Bhavani Raskutti ComputerScienceDepartment MonashUniversity Clayton,VICTORIA3168 AUSTRALIA Abstract Duringinteractionswithhumanconsultants, people are used to providing partial and/ or inaccurate information, and still be under stood and assisted. We attempt to em ulate this capablity of human onsultants in computer consultation systems. In this paper, we present a mechanism for han dling uncertainty in plan recognition dur ingtask-orientedconsultations. Theuncer tainty arises while choosing an appropriate nterretatioof a user'sstatementsamong many possible interpretations. Our mecha nismhandlesthisuncertaintybyusingproba bilitytheorytoassesstheprobabilitiesofthe interpretations,andcomplementsthisassess mentbytakingintoaccountthe information content oftheinterpretatons. Theinforma tion content of an interpretation is a mea sureofhowwelldefnedaninterpretationis in terms of the actions to be performed on the basis of the interpretation. This mea sureisusedtoguidetheinfereneprocessto wardsinterpretationswithahigherinforma tioncontent. Theinformationcontentofan interpretationdependsonthespecifcityand thestrengthotheinferencesinit,wherethe strengthofaninferencedependsonthereli abilityoftheinformationonwhichtheinfer enceisbased. Ourmechanismhasbeen de velopedforuseintask-orientedconsultation systems. The domain that we have chosen forexplorationisthatofatravelagency. 1 INTRODUCTION Duringtaskorientedconsultations,aconsultantneeds toinferauser'srequirementsfromhis/herstatements inordertoprovideassistance. Tothisefect, thecon sultantneedstointerpretauser'sstatementscorrectly. However,thistaskishinderedbythefactthatpeople Ingrid Zukerman ComputerScienceDepartment MonashUniversity Clayton,ICTORIA3168 AUSTRALIA often provide partialand/or inaccurate inormation. Thisrequires the consultantto fll in the missing in formationbyusingdiferentinformationsources,such as knowledge of discourse coherence, domain knowl edge and knowledge about the user. However, these informationsourcesarenotfullyreliable,requiringthe consultanttodrawinferenceswhihareinherentlyn certain. Asaresult,thestatementsissuebytheuser maybeinterpretedinmorethanoneway. Hence,the consultant needs to evaluatethe possible interpreta tions and select the most probable one. In this pa per wepresentamechanismforhandling theuncer taintyarisingfromthekof relabilityofhevarios information sources sed by the consultant, and for discriminating between multiple interpretations of a usersstatements. Aninterpretationofa user'sstatementsconsistsof a sequenceofplansthattheuserproposestocarryout, and a plan consists of an action with a number of parameters defning the action. For instance, in the traveldomain, the proposal tofyfrom Melbourne to SydneyonDecember1st, 1990,isaplan,whereflying isthe action, andthe parameters origin, destination and departure date are instantiated. A number of researchers have used plan recognition asameanstoresponsegenerationduringconsultation (Grosz 977, Allenand Perrault 980, Sidner and s rael 1981,Carberry1983,LitmanandAllen1987,Pol lack 1990). However, the models ofplan recognition developedbytheseresearcherscopeonlywthasingle interpretation of a user'sactions or utterances. Car berry (1990)addressestheproblemofmultiple inter pretationsbyusingdefaultinferences,andbyapplying Dempster-Shafertheoryofevidence(Section9.1,Pearl 1988) to compute plausibilityfactors ofalternate hy potheses. However, indomains suchastrave,where thedefaultassumptionsareweak,thisapproachalone doesnotcopewiththeproblemofmultipleinterpreta tions. KautzandAllen (1986)usecircumscription to generate all possibleinterpretationsdring story un derstanding. However, since all possibilities are con structed, an thesearc islimitedonly by thestruc-