Inference Web: Portable Explanations for the Web - CiteSeerX

1 downloads 0 Views 122KB Size Report
Richard Fikes, Jessica Jenkins, Gleb Frank, Eric H su, and Yulin Li for input on JTP, our ... McGuinness, Raphael Volz and Chris Welty. Web. Ontology Language ...
InferenceWeb: PortableExplanationsfortheWeb DeborahLMcGuinness . a ndPauloPinhei KnowledgeSystemsLaboratory

rodaSilva StanfordUniversity

Abstract

frame-like systems [McGuinness, 1996; BorgidaMcGuinness,1996]togeneraterequirements. Weals o obtainedrequirementsinputfromcontractorsinDAR PAsponsored programs concerning knowledge-based applications(theHighPerformanceKnowledgeBase 2 program1,RapidKnowledgeFormationProgram ,and 3 theDARPAAgentMarkupLanguageP rogram andmore 4 5 recently,theARDAAQUAINT and NIMD programs). We also obtained requirements from literature on explanationforexpertsystems,(e.g.,[Swartout,e t.al., 1991]), and usability of knowledge representation systems(e.g.,[McGuinness-Patel-Schneider,1998an d 2003]),andtheoremprovingexplanation(e.g.,[Fel tyMiller,1987]).

The World Wide Web lacks support for explaininginformationprovenance.W henweb applicationsreturnanswers,manyusersdonot knowwhatinformationsourceswereused,when theywereupdated,howreliablethesourcewas, whatinformationwaslookedupversusd erived, andifsomethingwasderived,howitwas derived. Inthisp aperweintroducetheI nference Web(IW)thataddressesthep roblemsassociated with opaque query answers by providing portable, combinable, and distributed explanations. The explanations include informationconcerningwhereanswerscame fromandhowtheywered educed(orretrieved). TheIWsolutionincludes:anextensiblewebbasedregistrycontainingdetailso ninformation sources and reasoners, a portable proof specification,andanexplanationb rowser.

1Introduction InferenceWeb(IW)aimstoenableapplicationsthat generateportableanddistributedexplanationsfor theiranswers. Therearemanyreasonsthatusersa agentsneedtounderstandtheprovenanceoifnforma that theyget backfromapplications. The main motivatingfactorsforusareinteroperability,reu trust.Interoperabilityisessentialifagentsare collaborate. Trustandreuseofretrievalanddedu processesisfacilitatedwhenexplanationsareavai Ultimately,ifusersand/oragentsareexpectedto informationandactionsofapplicationsandifthey expectedtouseandreuseapplicationresultspoten in combination with other information or other applicationresults,theymayneedtohaveaccesst kinds of information such as source, recency, authoritativeness,methodofreasoning,termmeanin andinterrelationships,etc. Thisworkbuildsonexperiencedesigningexplanatio componentsforreasoningsystems[McGuinness,1996; McGuinness-Borgida,1995;Borgida,et.al,1999,an 2000]andexperiencedesigningquerycomponentsfor

can anyof nd tion se,and to ction lable. trust are tially omany g n d

Ourgoalistoaddressneedsthatarisewithuseof systemsperformingreasoningandretrievaltasksin heterogeneousenvironmentssuchatsheweb. Users may obtaininformationfromindividualormultiplesour ces andtheymayneedtodetermine whichinformationto trust. Usersmayalsoobtainconflictinginformati onand theymayneedadditionalinformationtohelpevalua te whattobelieve. Theymayalsogatherinformation from complex and hybrid sources and they need help integratinganswersandsolutions. Aswebusagegr ows, abroaderandmoredistributedarrayofinformation services are available for use and the needs for explanations that can be shared across distributed environmentsgrow. In this paper, we include a list of explanation requirements gathered from past work and from surveying users. We present the Inference Web architectureandprovideadescriptionofthemajor componentsincludingtheportableproofspecificati theregistry(containinginformationaboutinferenc engines, proof methods, and ontologies), and the justificationbrowser.Wealsoprovidesomesimple 1 http://reliant.teknowledge.com/HPKB/ 2 http://reliant.teknowledge.com/RKF/ 3 http://www.daml.org 4 http://www.ic-arda.org/InfoExploit/aquaint/ 5 http://www.ic-arda.org/Novel_Intelligence/

IW on, e

usageexamples.W econcludewithadiscussionofo u workinthecontexto fexplanationworkandstateo contributionsintheareasoaf pplicationinteroper reuse,andtrust.

r ur ability,

• Termcoherence(isparticular a definitionincohere • Sourceconsistency(istheresupportinasystemfo bothAand~A) • Wereassumptionsusedinderivation? a Ifso,have assumptionschanged?

nt?) r the

2Requirements Ifhumansandagentsneedtomakeinformeddecision s aboutwhenandhowtouseanswersfromapplications , therearemanythingstoconsider.Decisionswill be basedonthequalityofthesourceinformation,the suitabilityandqualityofthereasoningengine,an dthe contextotfhesituation.Particularlyforuseo nt heweb, informationneedstobeavailableinadistributed environment and needs to be interoperable across applications. First, we consider issues concerning the source information.Evenwhensearchenginesordatabases simplyretrieveassertedo“r told”information,use rs(and agents)mayneedtounderstandwherethesource informationcamefromatvaryingdegreesodf etail. This information sometimes called provenance, may be viewedasmetainformationabouttoldinformation. Provenanceinformationmayinclude: • Sourcename(e.g.,CIAWorldFactBook) • Dateandauthor(s)oflastupdate • Author(s)oforiginalinformation • Authoritativenessofthesource(isthisknowledge storeconsideredorcertifiedasreliablebyathir d party?) • Degreeobelief f • Degreeocf ompleteness(Withinaparticularscope, is thesourceconsideredcomplete. Forexample,does thissourcehaveallotfheemployeesoafparticul ar organizationupuntilasomed ate? Ifso,notfind ing aparticularemployeewouldmeanthatt heyarenot employed,countingemployeeswouldbeanaccurate responsetonumberofemployees,etc.) Theinformationabovecouldbehandledwithmeta informationaboutcontentsourcesandaboutindivid ual assertions.Additionaltypesofinformationmaybe requirediusers f needtounderstandthemeaningo f terms orimplicationso fq ueryanswers. Ifapplications make deductionsorotherwisemanipulateinformation,use rs mayneedtounderstandhowdeductionsweremadeand whatmanipulationsweredone. Informationconcerni ng derivedomanipulated r informationmayinclude: • Termorphrasemeaning(innaturallanguageora formallanguage) • Term inter-relationships (ontological relations includingsubclass,superclass,part-of,etc.) • Thesourceofderivedinformation(reasonerused, reasonermethod,reasonerinferencerule,etc.) • Reasonerdescription(isthereasonerusedknownto be soundandcomplete?) • Termuniqueness(isJ.Smiththesameindividualas JohnSmith?)

3UseC ases Everycombinationofaquerylanguagewithaqueryansweringenvironmentisapotentialnewcontextfo rthe InferenceWeb.Weprovidetwomotivating scenarios. Considerthesituationwheresomeonehasanalyzeda situationpreviouslyandwants toretrievethisanalysis. In ordertopresentthefindings,theanalystmayneed todefend theconclusionsbyexposingthereasoningpathused along withthesourceotfheinformation. Inorderfort heanalyst toreusethepreviouswork,s/hewillalso needtodecideif thesourceinformationusedpreviouslyisstillval id(and possibly ithe f reasoning pathistill s valid). Anothersimplemotivatingexampleariseswhenuse a asks r forinformationfromawebapplicationandthennee dsto decidewhether toactontheinformation. Forexample,a usermightuse search a engineinterfaceoqaruery language 6 suchaD s QL for retrievinginformationsucha“szinfandels from NapaValley” or“winerecommendedforserving with aspicyredmeatmeal”(asexemplifiedinthewine agent exampleintheOWLguidedocument[Smithet.al.,20 03]). Ausermightaskforanexplanationofwhythepart icular wineswererecommendedaswellaswhyanyparticula r propertyofthewinewasrecommended(likeflavor, body, color,etc.). Theusermay alsowantinformationc oncerning whoserecommendationsthesewere(awinestoretryi ngto moveits inventory, w a inewriter,etc.). Inorderforthisscenariotobeoperationalized,w eneedto havethefollowing: • Awayforapplications(reasoners,retrievalengine s,etc.) todumpjustificationsfortheiranswersinforma a that t otherscanunderstand.Tosolvethisproblemwe introduce portable a proof specification. • Aplaceforreceiving,storing,manipulating,annot ating, comparing, andreturningmetainformationusedto enrichproofsandprooffragments.Toaddressthis requirement,weintroducetheInferenceWebRegistr y forstoringthemetainformationandtheInference Web Registrarwebapplicationfor handling theRegistry . • Awaytopresentjustificationstotheuser.Ason e solution tothis problem,weintroduce proof a brow ser.

4InferenceW eb Webeginwithashortdescriptionodf ifferentcate InferenceWebusers. Theseusersalongwiththeus 6 http://www.daml.org/2002/08/dql/.

goriesof age

examplesabove motivatethemaincomponentsoInference f Web:portableproofsandtheirparsers,registrya ndits registrar,andproof browsers. Theprimeusers of inferencewebare: • Applicationdevelopers(authorsofreasoners,sear ch engines,database systems,etc.) whowouldliketo justifywhytheiranswerstoqueriesshouldbebeli eved orwhowouldliketostateunderwhatconditionsth eir systemsarebestused. Thesepeopleareinterested in allowingtheirsystemtonotonlyanswerqueriesbu t alsoprovidemetainformationabouttheanswer. Th e portableproofspecificationinInferenceWeballow s applicationdeveloperstostorethisinformationin a sharableformat. • Authorsofhybridsolutionsprogramsinterestedin combining multiple answering systems and/or knowledgebases.Thesepeopleneedtounderstand howtermsrelatetoeachotherandhowanswerswere derivedandmightbeintegrated. Examplesofsuch peopleincludeontologybuilderswhoaremerging ontologiesoer xtendingontologies,crawlerow r rap per authors,peoplecombiningdatabasesorknowledge basedsystems,etc.TheregistryinInferenceWeb provides a store of information about inference methods,inferenceengines,ontologies,andsources thathelps address theseissues. • Humansoragents needingtodecideiftheycantrust eitherretrievedinformationoinference r processes used toretrieveinformation.Thebrowserininferencew eb addressestheseissuesbyallowinguserstoviewpa rtial or completejustificationsfor answers. InferenceWebcontainsboth datausedforproofgeneration andpresentationand toolsforbuilding,maintaining, presenting,andmanipulatingproofs. InferenceWeb data includesproofsandprooffragmentspublishedanywh ereon theweb. InferenceWebdataalsoincludesacentra lized repository of meta-data including sources, inferenc e engines,inferencerulesandontologies.Inference Webtools includearegistrarforinteractingwiththeregist ry,aparser forproofI/O,abrowserfordisplayingproofs,and planned futuretoolssuchasproofweb-searchengines,proo f verifiers(possiblyutilizingtoolssuchaSs pecwar e,etc). In thispaper,welimitourdiscussiontotheportable proofs (andanassociatedparser),theregistry(andthea ssociated registrar tools),andthebrowser.

4.1 PortableProof s Systemsthatmaybeaskedtoreturnajustification answeralongwithananswerneedtoexposeprovenan informationalongwiththeirdeductiveprocessposs includingmetainformationaboutthesystemitself. provideaspecificationwritteninthewebmarkupl DAML+OIL[Connollyet.al.,2001].Proofsdumpedin portableproofformatbecomeaportionoftheInfer Webdatausedforpresentingproofs. Ourportable specificationincludesfourmajorcomponentsoIfW trees:inferencerules,inferencesteps,wellforme

foran ce ibly We anguage the ence proof proof dformulae

(WFFs),andreferencedontologies. Inferencerules modusponens)canbeusedtodeduceaconsequent(a formedformula)fromanynumberofantecedents(als formedformulae).Aninferencestepiassingleapp ofaninferencerule. Theinferencestepwillbea withtheconsequentWFFandiwill t containpointer antecedentWFFs,theinferenceruleused,andanyv bindingsusedintheinferenceruleapplication.Th antecedentWFFsmaycomefromotherinferencesteps existingontologies,extractionfromdocuments,or beassumptions.Figure1presents taypicaldumpof

(suchas well owell lication ssociated tsothe ariable e , theymay W a FF.

(a WFF is stored as a predicate logic sentence) type (…) (a WFF can be associated to a set of Inference steps) (…) (inference step antecedents are IW files with their own URIs) (…)

Figure1An . InferenceWebProof TherewecanseeaninstanceoaWFF, f aninference step, andaninferencerule.Thereins o ntologyassocia tedwith thisWFFsinceitisderived. Ifithadbeenasser ted,it wouldrequireaanssociation totheontology thatc ontains it. Aproofcanthenbedefinedasatreeofinference steps explainingtheprocessodeducing f theconsequentW FF.In InferenceWeb,proofsare treesopf rooffragments rather thansinglemonolithicproofs.Withrespecttoaqu ery,a logicalstartingpointfor pa roofinInferenceWeb is proof a fragmentthatcontainsthelastinferencestepused toderive aWFFthatisananswerforthequery. Anyinferen cestep canbepresentedasastandalone,meaningfulproof

fragmentasicontains t theinferenceruleusedwit hlinksto itsantecedentsandvariablebindings.Thegenerati onof prooffragmentsisastraightforwardtaskonceinfe rence enginedatastructuresstoringproofelementsarei dentified asIWcomponents. Tofacilitatethegenerationof proofs, theInferenceWebprovidesaparserinJavathatdu mps proofsfromIWcomponentsanduploadsIWcomponents fromproofs.Thedevelopmentoaf nIWparserinLIS Pis under consideration. TheIWinfrastructurecanautomaticallygeneratefo llow-up questionsfor anyprooffragmentbyaskinghoweach antecedent WFF was derived. The individual proof fragmentsmay bceomposedtogether togenerate co a mplete proof,i.e.,asetofinferencestepsculminatingi ninference steps containing only asserted (rather than derived ) antecedents.. WhenanantecedentWFFisasserted, there arenoadditionalfollow-upquestionsrequiredand thatends thecompleteproof generation. AWFFmaybetheconsequentofanynumberoifnfere nce steps. IWcanbeusedtosupportmultiplejustific ationsfor any particularWFF. WFFsmay notbetheconsequent ofan inferencestepiftheyareassumptionsormerelyas serted informationinanontologythattheuserisreferen cing. The specificationoIW f conceptsusedinFigure1isav ailableat http://www.ksl.stanford.edu/software/IW/spec.

ofarule. Thus,reasoner-specificrulescanbex theRegistrybeforethereasonerisactuallyusedt IWproofs. Inferencewebthusprovidesawaytous reasoner toexplainanother reasoner’s inferenceru wasthestrategyusedin[Borgidaeal, t 1999]for Thisstrategymaybe useful forexplaining heavily optimizedinferenceengines.InferenceWeb’sregis whenfullypopulated,willcontaininferencerules manycommonreasoningsystems.Usersmayview inferencerulesetstohelpthemdecidewhetherto particular inferenceengine.

plainedin ogenerate eone les. (This example.) try, etsfor usea

4.2 Registry TheIWregistryiscurrentlyacentralizedreposito ryof informationusedtoenrichexplanationswithdetail sabout authoritativesources , ontologies, inferenceengines, and inferencerules In .thefuture,wemaystoreonlypointersto registryentriespublishedelsewhereontheweb.Ou r registryincludestemplateinformationabouteacho fthe classesintheregistry. Forexample,inferenceen ginesmay havethefollowingpropertiesassociatedwiththem: name, URL,author(s),date,versionnumber,organization, etc. The current demonstration registry is available at: http://belo.stanford.edu:8080/iwregistry/BrowseRegistry.jsp Informationintheregistrycontainstheinformatio nlinked tointheproofs.Everyinferencestepshouldhave alinkto atleastoneinferenceenginethatwasresponsible for instantiating theinferencestepitself,as shown i F n igure1. Thedescriptionofinferencerulesisoneofthemo st importantfeaturesotfheRegistry.Registeredrule scanbe atomicocan r bdeerivedfrom other registeredrule s. InordertointeractwiththeIWRegistry,thereis R a egistrar webapplicationallowinguserstoupdateorbrowse the registry. AscreenshotfromtheRegistrarinterf acefor inferencerulesisincludedinFigure2. Thisdisp laysa listingoftheatomicinferencerulesfortheJTPm odeleliminationreasoneratStanford.Eachoftheinfe rence rulesincludesaname,description,optionalexampl e,and optionalformalspecification. Manyreasonersalsouseasetofderivedrulesthat maybe usefulforoptimization,forexample.Oneindividu al reasonermaynotbeabletoprovideaproofotfhe derived rulesbutonereasonermaypointtoanotherreasone r’sproof

Figure2Sample : InferenceWebRegistrarEntry Ontologies areanother componentin theIWregistry Ontologies arestores of assertions thatmay buese proofs.Itcan biemportanttobaebletporesenti such aontology s source,date,version,URL (for br etc. Figure3contains saampleontology registry theontology usedionurwineexamples.

. din nformation owsing), entry for

theaxiomisdefined. InFigure4selecting , the consequent wouldpresentinformationaboutJTP-theinference engine usedtoderive Selecting it. GMP –theinferencerule,would presentinformationaboutJTP’sGeneralizedModusP onens rule. Selecting statement a such a“sbeef curry is sapicy red meat”o“spicy r redmeatcoursesrequirefull-bodie dwines” presentsinformationaboutthewinesontology. Sel ectinga derivedocached r inferencerulepresentsinformati onabout theinferencerule.(JTPusesasetofspecialpur pose axiomsformoreefficientlyreasoningwiththeDAML +OIL language and those inferences maybe used inan explanation). Anexampleothis f processcanbe seenfrom the Inference Web web pages at: http://www.ksl.stanford.edu/software/iw/Ex1/.

5

Figure Sample 3: InferenceWebOntologyEntry

4.3 Brow ser InferenceWebincludesabrowserthatdisplaysproo f fragmentsinanumberof ormats. Initially,wein clude 7 formatsforrestrictedEnglish,KIF and , conjunctivenormal form.Wealsoexpectthatsomeapplicationsmay implementtheir own displaysusing theIWAPI. Theprototypebrowserallows user a toseeaninfer encerule usedalongwiththederivedsentenceandtheantece dent sentences.Thebrowserimplementsalensmetaphor responsibleforrenderingafixednumberoflevels of inferencestepsdependingonthelensmagnitudeset ting. Figure4presentsaninferencestepforonewineus ceasein Section3.Priortothisview,theprogramhasaske dwhat winetoservewith spicy-red-meat a course. In Fig ure4one , canseethatNEW-COURSE12,which istheselectedmeal course,requiresadrinkthathasafullbodysince itisa spicyredmeatcourse. Thesentencesareformatted inKIF andthelensmagnitudeios ne,thusthebrowserdis playsthe inferencestepusedtoderiveiincluding t itsante cedents. A lenssettingoftwowouldalsoincludetheantecede nt’s derivations. Webelievethatoneofthekeystopresentationof justificationsibsreakingproofsintoseparablepi eces. Since wepresentfragments,automaticfollow-upquestion support isacriticalfunctionoftheIWbrowser. Everyel ementin theviewinglenscantriggerabrowseraction.The selection ofanantecedentthatisderivedre-focusesthelen sonan antecedent’sinferencestep.Forotherlenselement s, associatedactionspresentRegistrymeta-informatio ninthe TrustDisclosurePanel .Theselectionoftheconsequent presents details abouttheinferenceengineusedto derivethe actualtheorem.Theselectionoan finferencerule presentsa descriptionoftherule.Theselection ofasentencethatis assertedinformationpresentsdetailsaboutontolog ieswhere 7 http://logic.stanford.edu/kif/kif.html.

RelatedWorkandContributions

Recognitionoftheimportanceofexplanationcompon ents forreasoningsystemshasexistedinanumberofi eldsfor manyyears. Forexample,fromtheearlyexperienc eswith MYCIN[Shortliffe,1976], expert systems researchers understoodtheneedforsystemsthatunderstoodthe ir reasoningprocessesandcouldgenerateexplanations ina languageunderstandabletoitsusers.Inferencew eb attemptstostandontheshouldersopf astworkin expert systems,suchaM s YCINandtheexplainableexperts ystem ongeneratingexplanationsusingboththeirleaning osnhow togenerateexplanationsandinteroperatingwithne xt generationsystems thatgenerateexplanations.IWalso buildsonthelearningsoef xplanationindescripti onlogics (e.g.,[McGuinness,1996;Borgida,et.al,2000])th at attemptstoprovidealogicalinfrastructureforse parating piecesoflogicalproofsandautomaticallygenerati ng follow-questionsbasedonthelogicalformat. Ita lsolooked tothetheoremprovingcommunitywithworksuchas [Felty-Miller, 1987]) that attempts to provide understandableandflexibleexplanationsotheorem f provers andfoundationalsystemssucha[sBoyer,et.al,19 95]that providessomeexplanationsodeductions f alongwith WFFs notproven.

Figure An 4:InferenceWebBrowserScreen Wearenotawareow f orkthathasattemptedtoprov idean infrastructureforproviding,storing,andmanipula ting interoperable explanations of heterogeneous reasoni ng systems. Beyondjustexplainingasinglesystem,I nference

Webattemptstoincorporatebestinclassexplanati onsand provideawayofcombiningandpresentingjustifica tions thatareavailable. Itdoesnottakeonestanceon theform of theexplanationsinceitallowsdeductiveenginest odump singleormultipleexplanationsofanydeductionin deductivelanguageotheir f choice. Itprovides th ueserwith flexibilityinviewingfragmentsofsingleormulti ple explanationsinmultipleformats.IWsimplyrequir es inferenceruleregistration andportableproof form at. InferenceWebprovides thefollowing contributions: • An architecture supporting interoperability between agentsusingportableproofs.Portableproofsare specifiedintheemergingwebstandardDAML+OILso as to leverage XML-, RDF-, and DAML-based informationservices. Prooffragmentsaw s ellase ntire proofsmay bienterchanged. • Lightweightproofbrowsingusingthelens-basedIW proofbrowsersupportingeitherprunedjustificatio nsor guidedviewing ocafompletereasoning path. • Supportforalternativejustificationsusingmultip le inference steps. This allows derivations to be performedbyperformancereasonerswithexplanation s beinggeneratedbyalternativereasonersaimedat human consumption. • Registry oinference f engines,ontologies,andsour ces. 8 WearecurrentlyextendingtheStanford’sJTP theorem provertoproduceportableproofsandsimultaneousl y populatingtheIW registrywithJTPinformation. Future workincludesexpandingtoincludeotherinference engines. Wealsointendtoprovidespecializedsupportforw hy-not questions expanding upon [Chalupsky-Russ,2002] and [McGuinness,1996].Wearealsolookingataddition al supportfor proof browsing andpruning.

6Conclusion Inferencewebenablesapplicationsthatcangenerat e portableexplanationsoftheirconclusions. Wedes cribed themajorcomponentsofIW –theportableproof specificationbasedontheemergingweblanguage-DA ML (soontobuepdatedtoOWL),theregistry,andthe IWproof browser. Wefacilitateduseindistributed a envir onmentby providingIWtoolsforregisteringandmanipulating proofs, prooffragments,inferenceengines,ontologies,and source information. We also facilitated interoperability by specifyingtheportableproofformatandproviding toolsfor manipulatingproofsandfragments. Wehaveimpleme nted theIWapproachforoneinferenceengine(JTP)and arein discussionswithadditionalreasonerauthorstoinc ludemore reasoningengines. Wehavepresentedtheworkats ome governmentsponsoredprogrammeetings(RKF,DAML,an d AQUAINT) to gather input from other reasoner authors/usersandhaveobtainedfeedbackandintere st.We areinitiatingadditionalreasoner,ontology,ands ource registrations. 8 http://www.ksl.stanford.edu/software/jtp/.

Acknowledgements Manypeoplehavep rovidedvaluableinputtoourwor ThanksinparticulargotocolleaguesaK t SLinclud RichardFikes,J essicaJenkins, GlebFrank,EricH andYulin LiforinputonJTP,ourspecificationor applications. Alsothanksg otoanumberocf ollea insomegovernmentprogramswhoprovidedinput includingH ansChalupsky,PeterClark,KenForbus, Murray,andSteveReed, Allerrors,o fcourseare responsibility.

k. ing su, gues Ken our

References [Borgidaet.al,2000]AlexBorgida,EnricoFranconi and , IanHorrocks.Explaining ALCsubsumption.I n Proc.of the14th European Conf.on ArtificialIntelligence (ECAI'2000),pages2 09-213.IOSP ress,2000. [Borgidaeal., t 1999]AlexB orgida,EnricoFrancon i,Ian Horrocks,Deborah McGuinness,andPeterPatelSchneider.``ExplainingALCsubsumption”Proceeding s of the International Workshop on Description Logics(DL-99),Linköping,Sweden,July1 999,pp3336. [Boyeret.al,1995] RobertBoyer,MattKaufmann,a ndJ . Moore.TheBoyer-MooreTheoremProverandIts InteractiveEnhancement, ComputersandMathematics withApplications , 29(2),1995,pp.27-62. [ChalupskyandRuss,2 002]H ansChalupskyandTom Russ.WhyNot:DebuggingFailedQueriesinLarge KnowledgeBases.In ProceedingsoftheFourteenth Innovative Applications of Artificial Intelligence Conference(IAAI-02) pages , 870-877. [Connollyet.al,2001]DanConnolly,Frankvan Harmelen,I anHorrocks,DeborahL.M cGuinness,Pete r F.Patel-Schneider,andLynnAndreaStein. DAML+OIL (March2001)ReferenceDescription.W3CNote18 December, 2001. http://www.w3.org/TR/daml+oilreference. [Deanet.al.,2002]MikeDean,DanConnolly,Frank van Harmelen, James Hendler, Ian Horrocks, Deborah McGuinness,PeterPatel-Schneider,andLynnAndrea Stein.OWLWebOntologyLanguage1.0Reference. WorldWideW ebConsortium(W3C)WorkingDraft29 July 2002. Latest version is available at http://www.w3.org/TR/owl-ref/. [FeltyandMiller,1987]AmyFeltyandDaleMiller. ProofExplanationandRevision.UniversityofPenn, Tech. Report MSCIS8817, 1987 http://cm.belllabs.com/who/felty/abstracts/proof87.html. [Fikesetal.,2002]RichardFikes,PatHayes,Ian Horrocks, editors. DAML Query Language(DQL) Abstract Specification. http://www.daml.org/2002/08/dql/. [McGuinness,1996]DeborahL.McGuinness.1996. Explaining ReasoninginDescriptionLogics. Ph.D. Thesis,RutgersUniversity,TechnicalReportLSCR-T R277.

[McGuinness and Borgida, 1995] Deborah L. McGuinnessandAlexBorgida. ExplainingSubsumption inDescriptionLogics In. Proc.14thInternationalJoint Conf.onArtificialIntelligence Montreal, , Canada.1995. [McGuinness and Patel-Schneider, 2003] Deborah McGuinness and Peter Patel-Schneider. ``From DescriptionLogicP roverstoKnowledgeRepresentati on Systems''.FranzBaader,DeborahMcGuinness,Daniel e Nardi,andP eterPatel-Schneider,editorsT heDescr iption Logic Handbook: Theory, Implementation, and Applications.CambridgeUniversityPress,2003. [McGuinness and Patel-Schneider, 1998] Deborah McGuinnessandP eterPatel-Schneider.``UsabilityI ssues inKnowledgeRepresentationSystems''.I n Proceedings of the Fifteenth National Conference on Artificial Intelligence,M adison,Wisconsin,J uly,1 998.Updated version of ``Usability Issues in Description Logic Systems''publishedin Proc.oIf nternationalWorkshop onDescriptionLogics ,GifsurYvette,(Paris),France, Sept,1997. [Shortliffe,1976] EdwardHanceShortliffe. Compu terBasedMedicalConsultations:M YCIN. Elsevier/North Holland,NewYork,1976. [Smith et al., 2003] Michael Smith, Deborah L. McGuinness,RaphaelVolzandChrisWelty.Web OntologyLanguage(OWL)GuideVersion1.0.World WideW ebConsortium(W3C)WorkingDraft.Available athttp://www.w3.org/TR/owl-guide. [Specware,2001] http://www.specware.org/ [Swartoutetal.,1991]W.S wartout,C. Paris,andJ. Moore.“ ExplanationsinK nowledgeSystems:Designfor ExplainableExpertSystems”. InIEEEIntelligent Systems,J une1991. http://www.computer.org/intelligent/ ex199/x3058abs.htm.