Implementing OWL Defaults VladimirKolovski1andBijanParsia2andYardenKatz1 1 DepartmentofComputerScience, UniversityofMaryland,CollegePark,MD20742,USA, [email protected],[email protected] 2 SchoolofComputerScience TheUniversityofManchester, [email protected] Abstract. WhileithasbeenarguedthatknowledgerepresentationfortheWorld WideWebmustrespecttheopenworldassumptionduetothe“open”natureof theWeb,usersoftheWebOntologyLanguage(OWL)haveoftenrequestedsome formofnon-monotonicreasoning.Inthispaper,wepresentpreliminaryoptimiza- tionsandanimplementationofarestrictedversionofReiter’sdefaultlogicasan extensiontothedescriptionlogicfragmentofOWL,OWLDL.Weimplementthe decisionprocedureforthislogicintheOWLDLreasonerPellet,andexploitits supportforincrementalreasoningthroughupdatetoimproveperformance.We also extend an open source ontology editor SWOOP with default rules editing support. 1 Introduction TheWebOntWorkingGroupidentifiedtwoobjectives3concerning,atleastpotentially, non-monotonicity,andyettheresultingWebOntologyLanguage(OWL)isthoroughly monotonic.Bothobjectives(someformofdefaults,i.e.,“defaultpropertyvalues”,and an unspecified version of the closed world assumption) could be plausibly met with Reiter’sdefaultlogic,whichisoneofthemoststudiednon-monotoniclogics. Reiter’s default logic (or simply default logic henceforth), while very expressive, is, like many non-monotonic formalisms, known to be computationally difficult even in the propositional case. In [1], Baader and Hollunder showed that even a restricted form of defaults coupled with a description logic that contains a smaller set of class constructorsthanOWL-DL4isundecidable.Theyalsoshowedthatifonerestrictedthe defaults to apply to only to named individuals (or, equivalently, restricted the logic to closeddefaults),thenarobustdecidabilityensued.Thedecisionprocedureproposedin thatpaperproveddifficulttoimplementwithreasonableefficiencyandtheproposalhas languishedeversince. However,thestateoftheartindescriptionlogicreasonershasadvanced,andthere have been recent developments in the OWL-DL reasoner Pellet that suggest that a 3SeeO2andO3intheusecasesandrequirementsdocumentathttp://www.w3.org/TR/webont- req/ 4ThelogicthatBaaderandHollunderuse(ALCF)incorporatesfeatureagreements,whichare notpartofOWL-DL,sowecannotclaimitisasubset. Report Documentation Page Form Approved OMB No. 0704-0188 Public reporting burden for the collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information, including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to a penalty for failing to comply with a collection of information if it does not display a currently valid OMB control number. 1. REPORT DATE 3. DATES COVERED 2006 2. REPORT TYPE 00-00-2006 to 00-00-2006 4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER Implementing OWL Defaults 5b. GRANT NUMBER 5c. PROGRAM ELEMENT NUMBER 6. AUTHOR(S) 5d. PROJECT NUMBER 5e. TASK NUMBER 5f. WORK UNIT NUMBER 7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION University of Maryland,Department of Computer Science,College REPORT NUMBER Park,MD,20742 9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR’S ACRONYM(S) 11. SPONSOR/MONITOR’S REPORT NUMBER(S) 12. DISTRIBUTION/AVAILABILITY STATEMENT Approved for public release; distribution unlimited 13. SUPPLEMENTARY NOTES The original document contains color images. 14. ABSTRACT 15. SUBJECT TERMS 16. SECURITY CLASSIFICATION OF: 17. LIMITATION OF 18. NUMBER 19a. NAME OF ABSTRACT OF PAGES RESPONSIBLE PERSON a. REPORT b. ABSTRACT c. THIS PAGE 10 unclassified unclassified unclassified Standard Form 298 (Rev. 8-98) Prescribed by ANSI Std Z39-18 2 reasonable implementation of a default logic reasoner can be achieved. These devel- opments include tableau tracing for the description logic SHOIN and incremental reasoningsupport.Inthispaper,wepresentaprototypeimplementationofthe“termi- nologicaldefaults”ofBaaderandHollunderbasedontheabovedevelopments.Tableau tracingallowsustouseSHOIN astheunderlyingdescriptionlogicoftheterminolog- icaldefaults,hencethetermOWLdefaults.WeprovideUIforthisdefaultsreasonerby extendingopensourceontologyeditorSwoop[10].Weshowthroughempiricalevalu- ationthatthisprototypeperformsreasonablyoverall,butalsodoesfairlywellinsome caseswhencomparedtothepropositionaldefaultlogicreasoner,XRay. 2 Background 2.1 DefaultLogic Reiter’sdefaultlogicisanonmonotonicformalismforexpressingcommonsenserules ofreasoning.Theserules,calleddefaultrules(orsimplydefaults),areoftheform: α : β γ where α,β,γ are first-order formulae. We say α is the prerequisite of the rule, β is thejustificationandγ theconsequent.Intuitively,adefaultrulecanbereadas:ifIcan provetheprerequisitefromwhatIbelieve,andthejustificationisconsistentwithwhat Ibelieve,thenaddtheconsequenttomysetofbeliefs.ForasetofdefaultrulesD,we denote the sets of formulae occurring as prerequisites, justifications, and consequents inDbyPre(D),Jus(D),andCon(D),respectively. Definition1 AdefaulttheoryisapairhW,DiwhereW isasetofclosedfirst-order formulae (containing the initial world description) and D is a set of default rules. A defaulttheoryisclosediftherearenofreevariablesinitsdefaultrules. Possiblesetsofconclusionsfromadefaulttheoryaredefinedintermsofextensions ofthetheory.Extensionsaredeductivelyclosedsetsofformulaethatalsoincludethe original set of facts from the world description. Extensions are also closed under the application of defaults in D - we keep applying default rules as long as possible to generateanextension. Default rules can conflict. A simple example is when two defaults d and d are 1 2 applicableyettheconsequentofd isinconsistentwiththeconsequentofd .Wethen 1 2 typically end up with two extensions: one where the consequent of d holds, and one 1 wheretheconsequentofd holds. 2 2.2 TerminologicalDefaultTheories Inthissection,wepresentsomedefinitionsfrom[1]andtheiralgorithmforcomputing allextensionsofadefaulttheory(whichwasthebasisforourimplementation). 3 Definition2(TerminologicalDefaultTheory) A terminological default theory is a pair hW,Di where W is a SHOIN knowledge base and D is a finite set of default ruleswhoseprerequisites,justifications,andconsequentsareconceptexpressions. Aterminologicaldefaulttheoryisnotnecessarilydecidable-infact,[1]showsthata defaulttheoryhW,DiwhereW isanALCF logicisundecidableduetoskolemization in Reiter’s treatment of open defaults. In order to retain decidability, [1] proposes a restrictedsemantics:defaultrulesareonlyapplicabletonamedindividuals.Thus,aset ofdefaultsDisinterpretedasrepresentingthecloseddefaultsobtainedbyinstantiating the free variable by all individual names occurring in W. Baader and Hollunder also show that as long as the underlying logic language is decidable, with this restricted semanticstheproblemofcomputingalloftheextensionsisalsodecidable. Reasoning with Terminological Defaults For sake of completeness, we describe in thissectionthealgorithmfrom[1]forcomputingalloftheextensionsofaterminolog- icaldefaulttheory.Ourimplementationandoptimizationsarebasedonthisalgorithm. Definition3(GroundedSetofDefaults) LetWbeasetofclosedformulae,andDbe asetofcloseddefaults.WedefineD =∅and,fori≥0, 0 α : β D =D ∪{d= d∈DandW ∪Con(D )|=α} i+1 i γ i S ThenDiscalledgroundedinW iffD = D i≥0 i Themainreasoningserviceprovidedistheprocedurethatcomputesalloftheexten- sionsofaterminologicaldefaulttheoryD.Theprocedureworksinatop-downfashion: itstartswiththesetofallgrounddefaultsD ,andtheninspectsifthereisaconflicting 0 defaultdinthatset(i.e.,thejustificationofdconflictswithW ∪Con(D )).Thereare 0 twowaystoeliminatetheconflictandtheproceduredoesboth.First,itsimplyremoves thedefaultdfromD andappliesitselfagaintothesmallersetD \d.Theotherway 0 0 istokeepdinD ,andpruneotherdefaultsthatcausetheconflict.Inotherwords,fora 0 conflictingdefaultd,theprocedureisalsoinvokedwiththemaximalsubsetsD0 ofD 0 s.t.d∈D0andthejustificationofddoesnotconflictwithW∪Con(D0).Itisshownin [1]howfindingthemaximalsubsetsofD abovereducestotheproblemofcomputing 0 maximally consistent subsets: given two sets W and D , find the maximal subsets of 0 D0 ∈D s.t.W ∪D0isconsistent. 0 3 Optimizations TothebestofourknowledgetherehasbeennoadaptationofBaaderandHollunder’s algorithmtoOWL-DL.Recently,however,therehasbeenworkinDescriptionLogics thathasallowedustoadaptthealgorithmandoptimizeittoproduceaworkingimple- mentation. First, there is a practical tableau tracing algorithm [11] with low overhead which can be used solve the problem of computing maximally consistent subsets for 4 more expressive logics (up to SHOIN). Second, there has been recent work on op- timizingreasoningfordynamicAboxes[8].Thisworkallowsustooptimizethecon- sistency check every time there is an update to the working set of defaults D . Third, 0 weproposeatechniqueforgeneratingpossible(candidate)extensionsinabottom-up fashion.Wediscussallofthesetechniquesinthissection. 3.1 FindingMaximallyConsistentSubsets Inthissection,weexplainhowweapplyrecentworkintableautracinganddebugging OWL ontologies to solve the problem of finding the maximal subsets D0 ∈ D such thatCon(D0)∪W isconsistent.Theapproachin[1]introducesclashformulaeforthis problem-insteadweusejustifications.AjustificationforanABoxinconsistencyisa minimalfragmentoftheKBinwhichtheABoxisinconsistent.Justificationsarecom- putedbyatracingfunctionthatrecordsthechangesinthetableauandtheaxiomsand non-deterministicchoicesthatareresponsibleforthosechanges.Wereferthereaderto [11]formoredetails. Every justification contains DL axioms which are responsible for the particular clash.Inourcase,theseaxiomscomeeitherfromW orCon(D).Ajustificationwhere alloftheaxiomscomefromW indicatesthatW byitselfisinconsistent.Inthispaper, withoutlossofgenerality,wewillassumethatW isconsistent5.Thus,everyjustifica- tionwillcontainatleastoneaxiomfromCon(D). In order to find the maximally consistent subsets of Con(D), the first step is to findallofthejustificationsfortheinconsistencyofW∪Con(D).Usingtheinformation fromthosejustifications,thenwecandeterminetheminimumnumberofaxiomscom- ingfromCon(D)thatshouldberemovedtorenderW∪Con(D)consistent.Weusethe algorithmfrom[9]tofindalljustifications:itdoestableautracingtofindonejustifica- tionandReiter’sHittingSetTree(HST)algorithm[14]tofindallotherjustifications. TheintuitionbehindtheusageoftheHittingSetreliesonthefactthatinordertomake the KB consistent, one needs to remove from the KB at least one axiom from each justification. Aftercomputingalljustifications,thenextstepistodeterminewhatisthesmallest setofaxiomsthatweneedtoremovetomakeitconsistent.Becauseofthenatureofthe justifications, we only need to remove one axiom from each justification to make the KBconsistent.Itisimportanttonoteherethateventhoughtheaxiomsinajustification cancomefromeitherW orCon(D),weareonlyinterestedinremovingaxiomsfrom Con(D). Thus, we pre-process each justification, discarding the axioms coming from W andonlyleavingtheaxiomsfromCon(D). Atthispoint,weapplyReiter’sHSTalgorithmagainontheprunedjustificationsto obtaintheminimumsetcover.Forexample,forthesetofjustificationsS ={{a},{a,b},{c},{c,d}} theminimalhittingsetis{a,c}.Thisminimalhittingsetgivesusthesmallestsetofax- iomsfromCon(D)thatweneedtoremovetomakeCon(D)∪W consistent.Notethat atthispoint,therearenoconsistencychecksperformedwhilecomputingtheminimal hittingsets(unlikethepreviousapplicationofReiter’salgorithm)andthejustifications 5InthecaseWisinconsistent,thenthereisonlyoneextensionconsistingofalloftheformulae 5 containless.Nonetheless,wehaveobservedthatthistaskisthebottleneckofthesystem duetotheexponentialnatureofthehittingsettreealgorithm. 3.2 ABoxUpdates In the procedure described in Section 2.2 often there are incremental changes to the workingsetofdefaultsD :eithertheconflictingdefaultsisremoved,oraminimalset 0 ofotherdefaultsthatcausetheconflict.Thesemodificationsareequivalentofremov- ing (and adding) type assertions to W∪ Con(D ), thus reasoning time overall would 0 improveifweusedaDLreasoneroptimizedforABoxupdated. IncrementalConsistencyChecking Therecentworkthatweusedwas[8],whichisan approach for incrementally updating tableau completion graphs under ABox updates. Theirideaisnottothrowawaytheoldcompletiongraphafteranupdate.Instead,the update algorithm adds new (resp. removes existing for deletions) components (edge and/or nodes) induced by the update to the old completion graph; after this, standard tableau completion rules are re-fired to ensure that the model is complete. Therefore thepreviouscompletionisupdatedsuchthatifamodelexists(i.e,theKBisconsistent after the update) a new completion graph will be found. In [8], it was observed that updatesdidnothavealargeeffectontheexistingcompletiongraph,thereforeordersof magnitudeperformanceimprovementsareachieved. Applying Incremental Reasoning The incremental reasoning techniques can be ap- pliedinanumberofplacesinthealgorithmfrom[1]toimproveoverallperformance. First,weusethemtocomputethelargestgroundsubsetofrules-thepseudo-codefor thisprocedureisgiveninFigure3.2. functionFind-Largest-Grounded-Subset(W,D) begin (1) D =∅,changed=true l (2) whilechangeddo (3) changed=false (4) foreachd =α:β/γ ∈D\D i l (5) ifW∪Con(D)|=αthen l (6) incupdate(W,γ,add) (7) D =D ∪{d } l l i (7) changed=true (8) returnD l end Fig.1.Pseudo-codetogetthelargestgroundedsubset.Itisimportanttonoteline(6)whichin- crementally updates the world description W with a new type assertion. In line (5), to check whetherW ∪Con(D )|=αwesimplyincupdate(W,¬α,add)andinspecttheupdatedcom- i pletiongraph. 6 Another application of the ABox updates is in the procedure of finding all justifi- cations(Section3.1).Whilebuildingthehittingsettree,insteadofrebuildingthecom- pletion graph from scratch (when removing or adding axioms) we update only those portionsaffectedbythechange. 3.3 GeneratingCandidateExtensions Considering that the bottleneck of the algorithm that computes the extensions is in the subtask that computes the maximally consistent subsets (empirical results are in Section 5), we introduce here another optimization aimed at reducing the number of calls to the afore-mentioned subtask. The purpose of this optimization is to find the obvious conflicts and generate approximate extensions before the algorithm in [1] is used. The optimization is based on the idea of generating extensions bottom-up, as exploredin[12,3].Defaultsthatconflictareplacedintoseparateextensions.However, whendeterminingwhetherdefaultsd andd conflict,wesimplifytheconflictdetection 1 2 bytakingonlytheworlddescriptionWintoaccount,andignoringthecurrentextension. Followingweprovideadefinitionofobviouslyconflictingdefaults. Definition4(ObvioislyConflictingRules) ForadefaulttheoryhW,Diwedefinetwo closeddefaultsd ,d ∈ D tobeobviouslyconflictingiffatleastoneofthefollowing 1 2 holds: W ∪Con(d )|=¬Jus(d ), W ∪Con(d )|=¬Con(d ), 1 2 1 2 W ∪Con(d )|=¬Jus(d )orW ∪Con(d )|=¬Con(d ) 2 1 2 1 To find all of the obvious conflicts, we use a naive O(N2) algorithm where N is thenumberofcloseddefaults.Notethatinthiscasewecanagainexploitincremental reasoningtechniques.Usingtheinformationabouttheconflicts,weproceedtogenerate possible extensions. Since the extensions are built only on the basis of the simplified idea of obvious conflicts, and are different from extensions in terminological default theories,wecallthemcandidateextensions. Definition5(CandidateExtension) LetT = hW,Dibeacloseddefaulttheory.Let D ={α:β/γ | α:β/γ ∈DandW |=α}. g LetS =∅and∀i≥0define: 0 S =S ∪{d|d∈D andthereisnohs.t.h∈S andconflict(d,h)} i+1 i g i Then S ∪ {D \ D } is a set of generating defaults of a candidate extension iff g S S = S . i≥0 i Tocomputethecandidateextensions,wefollowtheabovedefinition:workingonly withthesetofdefaultsthatisgroundw.r.t.theworlddescriptionW,webuildcandidate extensionsusingbacktrackingsearchmakingsurethatnotwo”obviously”conflicting defaults are in the same extension. Then to each generated extension we append the setofnon-grounddefaults(defaultsnotusedinthesearch).Thesecandidateextensions are passed to the main (top-down) procedure where the rest of the conflicts are found andpruned.Wehaveevaluatedtheimpactofcandidateextensions(resultsavailablein Section5)andtheinitialresultsareencouraging. 7 4 Implementation Wehaveapreliminaryimplementationofouroptimizations.Itisprovidedasanexten- siontoanopensourceDLreasoner[13]anditprovidesrealizationofindividualswith terminologicaldefaultrules,witheithercredulousorskepticalreasoning.Wehavealso providedaUIfordefaultsbyextendingtheopensourceOWLOntologyeditorSwoop [10].Morespecifically,weaddedsupportfordefaultrulesediting(showninFigure2) andinformingtheuseroftheinferencesmadebytherules.Also,weupdatethecurrent ontologywiththesetofinferredfactsfromthedefaults. Fig.2.Screenshotofdefaultseditor 5 Evaluation Toevaluateourimplementation,weusedanIBMThinkPadT42withPentiumCentrino 1.6GHzprocessorand1.5GBofmemory.Firstweshowtheperformancebenefitsofin- crementalreasoningsupportandgeneratingcandidateextensions.Theontologiesused forthesetestsweremodifiedversionsoftheLehighUniversityBenchmark(LUBM)[7], withtwoopendefaultrulesoftheNixondiamondshape.Thenumberofextensionsis2i (Columnext.),whereiisthenumberof’Nixon’individuals.Performanceresults(times inseconds)areshowninTable1.Cand ext isthetimespentgeneratingthecandidate 8 extensions,groundedisthetimespentcomputinglargestgroundedsubsets,andHST is thetimespentcomputingmaximallyconsistentsubsets.Thebenefitofcandidateexten- sionsbecomesobviousincaseswithmoreextensionsforadefaulttheory. Size fully-optimized updates-only not-optimized indsext.cand ext HST totalgrounded HST totalgrounded HST total 50 8 .291 .300 .631 .803 .770 1.513 .952 1.671 2.553 100 8 .731 .3111.122 1.192 .861 2.003 2.585 1.902 4.447 200 8 2.504 .4303.084 1.9831.101 3.165 11.446 2.93414.190 50 64 .2911.2411.933 3.5474.364 8.462 3.46610.36514.141 100 64 .7611.4522.763 4.4085.08610.515 5.98714.05221.601 200 64 2.6131.9935.597 6.6117.61016.614 20.29220.47643.072 Table1.Comparisonofthreeversionsofpellet-defaults.Not-optimizeddoesnotgenerateany candidateextensionsorperformreasoningwithupdates.Updates-only,includestheincremental reasoning support but not the candidate extensions. Fully-optimized includes all of the tech- niquespresentedinthispaper. Next,wepresentacomparisonofourtoolwithXRay[5],whichisapopularProlog- basedimplementationforquery-answeringinpropositionaldefaultlogics,andsmodels [12],animplementationofthestablemodelsemanticsforlogicprogramswithnegation as failure. The first set of defaults rules that we used was a set of sample defaults on XRay’shomepage[6]: Student:Student Student:Student Student:Adult d = d = d = 1 Adult 2 ¬Employed 3 ¬Married Adult:¬Student Adult:¬Student d = d = 4 Employed 5 Married This theory has only one extension (where only {d ,d ,d } are fired for each 1 2 3 namedindividual).Unfortunately,since(tothebestofourknowledge)smodelsdoesnot supportclassicalnegation,wecouldnotevaluateitsperformanceforthesetestcases. Thesecondtestcaseisadefaulttheorywithonlytwoopendefaults,oftheNixon diamondshape.ResultsareshowninTable2.Inthecasewhentherearefewextensions, ourapproachoutperformsXRay,eventhoughouralgorithmisnotspecificallytunedfor propositional default logics. In the second example, in cases with a lot of extensions, wedonotfareaswell. 6 RelatedWork The majority of reasoners for default logic focus on propositional default theories, where quantifiers are not allowed. Two capable systems in this group that are com- 9 TypeandSizeofKB Pellet-defaults XRaysmodels Type Size Ext. grounded HST total total total 50 1 .181 .370 .615 1.472 N/A xrayex. 100 1 .5 .541 1.090 4.637 N/A 500 1 15.1421.71215.252timeout N/A 3 8 .03 .15 .2 .03 .02 Nixon 6 64 .130 .280 .63 .11 .04 11 2048 2.1833.3751.0745 .180 1.444 Table2.EvaluationResultsforcomputingalloftheextensions. parable in functionality include XRay [5] and DeReS [3]. However, the restricted ex- pressivity of the languages these reasoners handle make them of limited interest for OWLreasoning.ItwouldbeunrealistictoignoremanyofOWL’skeyconstructs−e.g. existential,universal,andcardinalityrestrictions−forthebenefitofaddingdefaultrules. Smodels [12] and dlv [4] are both efficient implementations of the answer set se- mantics. Since the answer set (or stable models) semantics for logic programs can be embedded into Reiter’s default logic (an answer set is analogous to a default theory extension),thesesystemscanalsobeseenasdefaultlogicreasoners.Themaindiffer- ence with our approach is that we are description logic-based, whereas they use logic programmingtechniques. 7 DiscussionandFutureWork Itisobviousfromtheevaluationthatourapproachdoesnotworkwellwithdefaultthe- orieswithlotsofextensions.Thereasonisthatthenumberofextensionsisdetermined bythenumberofconflictingrules.Everytimeaconflictingdefaultisdetected,thesub- task to compute the maximally consistent subsets procedure is executed (which is the bottleneckofoursystem).Agoalforfutureoptimizationscouldbetoavoidcallingthis subtaskasmuchaspossible.Wehavetakenthefirststepwiththebottom-upgeneration ofcandidateextensions.However,furtherworkisneeded.Webelieveitispossibleto achievemuchmoreaccurateandefficientgenerationofcandidateextensions,basedon thetechniquesin[12,3,2]. Intermsofexpressivity,animportantnextstepwouldbetohandleoneofthemany prioritized default logics, where default rules are partially ordered to determine pre- ferredextensions. 8 Acknowledgements ThankstoChristianHalaschek-Wiener,AdityaKalyanpurandEvrenSirinforhelpful comments and discussions. This work was supported in part by grants from Fujitsu,