308
Handling Uncertainty during Plan Recognition
in Task-Oriented Consultation Systems
Bhavani Raskutti
ComputerScienceDepartment
MonashUniversity
Clayton,VICTORIA3168
AUSTRALIA
Abstract
Duringinteractionswithhumanconsultants,
people are used to providing partial and/ or
inaccurate information, and still be under
stood and assisted. We attempt to em
ulate this capablity of human onsultants
in computer consultation systems. In this
paper, we present a mechanism for han
dling uncertainty in plan recognition dur
ingtask-orientedconsultations. Theuncer
tainty arises while choosing an appropriate
nterretatioof a user'sstatementsamong
many possible interpretations. Our mecha
nismhandlesthisuncertaintybyusingproba
bilitytheorytoassesstheprobabilitiesofthe
interpretations,andcomplementsthisassess
mentbytakingintoaccountthe information
content oftheinterpretatons. Theinforma
tion content of an interpretation is a mea
sureofhowwelldefnedaninterpretationis
in terms of the actions to be performed on
the basis of the interpretation. This mea
sureisusedtoguidetheinfereneprocessto
wardsinterpretationswithahigherinforma
tioncontent. Theinformationcontentofan
interpretationdependsonthespecifcityand
thestrengthotheinferencesinit,wherethe
strengthofaninferencedependsonthereli
abilityoftheinformationonwhichtheinfer
enceisbased. Ourmechanismhasbeen de
velopedforuseintask-orientedconsultation
systems. The domain that we have chosen
forexplorationisthatofatravelagency.
1 INTRODUCTION
Duringtaskorientedconsultations,aconsultantneeds
toinferauser'srequirementsfromhis/herstatements
inordertoprovideassistance. Tothisefect, thecon
sultantneedstointerpretauser'sstatementscorrectly.
However,thistaskishinderedbythefactthatpeople
Ingrid Zukerman
ComputerScienceDepartment
MonashUniversity
Clayton,ICTORIA3168
AUSTRALIA
often provide partialand/or inaccurate inormation.
Thisrequires the consultantto fll in the missing in
formationbyusingdiferentinformationsources,such
as knowledge of discourse coherence, domain knowl
edge and knowledge about the user. However, these
informationsourcesarenotfullyreliable,requiringthe
consultanttodrawinferenceswhihareinherentlyn
certain. Asaresult,thestatementsissuebytheuser
maybeinterpretedinmorethanoneway. Hence,the
consultant needs to evaluatethe possible interpreta
tions and select the most probable one. In this pa
per wepresentamechanismforhandling theuncer
taintyarisingfromthekof relabilityofhevarios
information sources sed by the consultant, and for
discriminating between multiple interpretations of a
usersstatements.
Aninterpretationofa user'sstatementsconsistsof a
sequenceofplansthattheuserproposestocarryout,
and a plan consists of an action with a number of
parameters defning the action. For instance, in the
traveldomain, the proposal tofyfrom Melbourne to
SydneyonDecember1st, 1990,isaplan,whereflying
isthe action, andthe parameters origin, destination
and departure date are instantiated.
A number of researchers have used plan recognition
asameanstoresponsegenerationduringconsultation
(Grosz 977, Allenand Perrault 980, Sidner and s
rael 1981,Carberry1983,LitmanandAllen1987,Pol
lack 1990). However, the models ofplan recognition
developedbytheseresearcherscopeonlywthasingle
interpretation of a user'sactions or utterances. Car
berry (1990)addressestheproblemofmultiple inter
pretationsbyusingdefaultinferences,andbyapplying
Dempster-Shafertheoryofevidence(Section9.1,Pearl
1988) to compute plausibilityfactors ofalternate hy
potheses. However, indomains suchastrave,where
thedefaultassumptionsareweak,thisapproachalone
doesnotcopewiththeproblemofmultipleinterpreta
tions. KautzandAllen (1986)usecircumscription to
generate all possibleinterpretationsdring story un
derstanding. However, since all possibilities are con
structed, an thesearc islimitedonly by thestruc-