HAL Id: hal-01945196
https://hal.archives-ouvertes.fr/hal-01945196
Submitted on 5 Dec 2018
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
A knowledge base with modularized ontologies for eco-labeling: application for laundry detergents
Da Xu, Mohamed Hedi Karray, Bernard Archimède
To cite this version:
Da Xu, Mohamed Hedi Karray, Bernard Archimède. A knowledge base with modularized ontologies
for eco-labeling: application for laundry detergents. Computers in Industry, Elsevier, 2018, 98, pp.118-
133. �10.1016/j.compind.2018.02.013�. �hal-01945196�
OATAO is an open access repository that collects the work of Toulouse researchers and makes it freely available over the web where possible
Any correspondence concerning this service should be sent
to the repository administrator: tech-oatao@listes-diff.inp-toulouse.fr This is an author’s version published in: http://oatao.univ-toulouse.fr/20083
To cite this version:
Xu, Da and Karray, Mohamed Hedi and Archimède, Bernard A knowledge base with modularized ontologies for eco-labeling:
application for laundry detergents. (2018) Computers in Industry, 98. 118-133. ISSN 0166-3615
Open Archive Toulouse Archive Ouverte
A knowledge base with modularized ontologies for eco-labeling:
Application for laundry detergents
Da Xu, Mohamed Hedi Karray*, Bernard Archimède
LaboratoireGéniedeProduction,LGP,UniversitédeToulouse,INP-ENIT,Tarbes,France
Keywords:
Ontologyengineering Ontologymodularization Knowledgebase OWLimports SWRL Eco-labeling
ABSTRACT
Alongwiththerisingconcernofenvironmentalperformance,eco-labelingisbecomingmoreandmore popular. However,the complexprocess ofeco-labeling is demotivatingmanufacturersand service providers to be certificated. The knowledge contained in eco-labeling criteria documents is not semanticallyexploitabletocomputers.Traditionalknowledgebaseinrelationaldatamodelisnotinter- operable, lacks inference support and is difficult to be reused. In our research, we propose a comprehensiveknowledgebasecomposedofinterconnectedOWL(OntologyWebLanguage)ontologies.
This ontology based knowledge base allows reasoning and semantic query. In this paper, a modularization scheme aboutontology developmentis introduced and ithasbeen applied toEU Eco-label(EuropeanUnionEco-label)laundrydetergentproductcriteria.Thisschemeseparatesentity knowledgeandruleknowledgesothattheontologymodulescanbereusedeasilyinotherdomains.
ReasoningandinferencebasedonSWRL(SemanticWebRuleLanguage)rulesinfavorofeco-labeling processisalsopresented.
1.Introduction
Since the late 1980s, there hasbeen a growing demand for products that do less harm to the environment. The public willingness to use buying power as a tool to protect the environment provides manufacturers with an opportunity to developnewproducts[1].Fromaglobalpointofview,promoteof environment-friendlyconsumptionandproductionwillcontribute notonlytothelifequalitybutalsotheeconomyitself.Buthow does a consumer judge and make good choices to reduce environmental impacts?Howshouldweassessthevalidityofa statement abouta productorservice'senvironmental impacts?
Theneedofevaluatingaproduct'senvironmentalperformancehas led to the establishment of eco-labels. Nowadays, most of the knowledgeandcriteriaabouteco-labeledproductsarepublished in official journals,webpages,and allkinds of documentation.
Usually,thisknowledgeispresentedinsuchcomplexregulation and specificationdocumentsthatitisdifficulttobeunderstood evenbyhumans.Theintegrationofthisknowledgeintosoftware requiresthatitmustbeexploitabletomachines.However,until now,there isstill a lackofcomputableformatof that.Besides,
traditional knowledge base in relational data model is not interoperable,lacksinferencesupportandisdifficulttobereused.
Inordertobetterunderstandthesecriteriaandrules,stakeholders need a common and machine accessible presentation of the knowledge.Toaddresssuchproblems,inourresearch,wepropose anontologicalknowledgebasecomposedofmodularizedontol- ogies.Thisschemehasbeenappliedtothecreationoftheontology knowledgebaseofEUEco-label'slaundrydetergentproducts.
DuetothefactthatEUEco-labelisalargeandcomplexlabeling system covering dozens of products and service groups, it is difficultandunrealistictocoverallitsproductsandservicesinthe research stage. Thus, we decide to choose laundry detergent productsgroupwhichhasamiddlesizeknowledgevolumetobe ourstudycase.TherestofthepaperwillfollowthisOutline:The first section presents a state of the art of eco-labeling and modularizedontology;inSection3,an overviewof thecriteria documentandrequirementanalysisispresented;Thethirdsection talksabouthowtheterminologyofontologyisretrieved;Section5 presents detailed design and construction of the ontology. In particular,anentity-ruleseparationpattern isintroduced.Basic ideaofthisseparationistoputdescriptiveentityknowledgeand subjectiveruleknowledgeintodifferentmodules.Thispatternis proventobeinfavorofmodularityandextendability,especiallyfor therulemodule.Itcanalsobeappliedtotheotherproductgroups’ ontologybuildingandevenothersimilarcriteria-likedocument's knowledgeextraction; thefifthsection is about how toutilize
* Correspondingauthor.
E-mailaddresses:da.xu@enit.fr(D.Xu),mkarray@enit.fr(M.H.Karray), bernard.archimede@enit.fr(B.Archimède).
reasonertodothereasoningupontheontologyknowledgebase andtheargumentation,which isveryimportanttoeco-labeling decisionsupportprocess;inSection7,wehaveabriefevaluation andanalysisfortheontology;Section8isaboutsomediscussionof experience feedback, and learned lessons. Finally, in the last section,wehaveconclusion,discussionandfuturework.
2. Stateofart
2.1.Eco-labelandEUEco-label
According to Global Eco-labelling Network1 (GEN), “eco- labelling” isa voluntary methodofenvironmental performance certificationandlabellingthatispracticedaroundtheworld.An
“eco-label”isalabelthatidentifiesoverallprovenenvironmental preferenceofaproductorservicewithinaspecificproduct/service category.Theyusuallyconcernthewholelifecycleoftheproduct andareissuedbyathirdparty[2].Eco-labelinghasanumberof benefitsfromvariouspointsofview.First,eco-labelingisagood waytoinformconsumersoftheenvironmentalimpactsofselected products.Inthepracticeofsomeexistenteco-labeling,thefitness of use and human health aspects are also included. All this informationwillhelpaconsumermakedecisionoutofdifferent willingness.Then,eco-labelingisgenerallycheaperthanregula- tory controls in terms of global economics. By empowering customersandmanufacturerstomakeenvironmentallysupport- ivedecisions,theneedforregulationiskepttoaminimum.Thisis beneficialtobothgovernmentandindustry[3].Eco-labelingwill also stimulate market development and encourage continuous improvementonproductsandservices.
EUEco-labelisasuccessfulexampleamongalltheeco-labels.
Created in 1992, EU Eco-label is the only official European ecologicallabelauthorizedfor usein everymembercountryof theEuropeanUnion[4].Until2011,thereareover1300enterprises thathavebeenissuedEUEco-labellicenses.BySeptemberof2014, therearealreadyover43,000productsorservicesbeinglabelled [5].FranceisalwaysanimportantcontributortoEUEco-labeling.
ByMarchof2016,486enterprisesinFrancehaveobtainedEUEco- labellicensesinvariousproductgroupsandthatmakesFrancethe first place as for the enterprises’ possession of EU Eco-label licenses.As illustratedin Fig.1,theremovalof certainproduct group(e.g.IPV:Indoorpaintsandvarnishes,SSC:Soaps,shampoos, andhairconditioners,and OPV:Outdoorpaints andvarnishes.) whichhappenedin2016indicatesthatthealterationofEUEco- label criteria is continuous. It also implies that the change of knowledgeandrules.AlthoughthesizeofLD(Laundrydetergents) groupisnotthelargest,itkeepsincreasingintherecent4years.
EUEB (EuropeanUnionEco-labeling Board)is responsible to developandregularlyrevieweco-labelcriteria.EUEBwillsetupan advisory body including representatives on behalf of different stakeholders. Feasibility study will be carried out to draft the environmentalcriteria.Atlast,representativesfromeverymember statewill besummoned tovoteto approve the criteriaor the guideline [6]. The guideline developed by the advisory body, together with the possible amendment or annex will be the baselinesfortheknowledgebasethatwedevelopedinthiswork.
2.2.Ontologyandmodularizedontology
Derivedfromphilosophy,incomputerscience,werefertoan ontologyasaspecialkindofinformationobjectorcomputational artifact [7]. Studer et al. [8] gave definition stating that: “An
ontologyisaformal,explicitspecificationofasharedconceptuali- zation”. Today, somany ontologies and knowledge repositories havebeendevelopedandadaptedintoapplications,especiallyin biomedicaldomains [9]. Successfulexamplesand platformsare BioPortal,2UniProt,3LEO,4etc.
Despite quiteamountof ontologiesof differentdomainsare developed,a lot ofproblemsare encounteredwhen knowledge engineersaswellasgeneraluserswanttounderstandandreuse theontologiesintotheirowndevelopment.Asfortheapplication of ontology, there is definite need to gather knowledge from multiple remote ontological sources. It is known that, when knowledgeisdistributed,theideatocollectallknowledgeandput themintoasinglerepository(i.e.theintegrationapproach)isvery difficulttoimplement,becauseofsemanticheterogeneitycalling forhumanprocessing[10].Anotherveryimportantreasonisthe low reusable design of theseontologies. Good ontologydesign patternhasdrawntheattentionofmanyresearchers.In[11]and [12],amethodtodescribeontologydesignpatternispresented.A Semantic Web portal called OntologyDesignPatterns.org5 is also available.However,mostofthesubmittedpatternsarecataloged in Content Ontology Design Patterns which means that the patternsthemselvesmaycontaincertainsemanticsand domain knowledge,whichmaystillsetobstaclestoontologyreuse.Also, mostof thesepatterns’structureishardtobemodularizedand veryfewofthemcareaboutmodularityinaspecificway.Thus, better engineering principle and philosophy about ontology modularityisneeded.
Generally speaking, there are two important aspects of ontologymodularization:independentlydevelopingmodulesthat canbeintegratedcoherentlyand uniformly(ontologycomposi- tion)orextractingsuchmodulesfromanintegratedontologyfor supportinga particularusecases(ontology decomposition)[9].
Mostofourresearchfocusonthefirstaspectandweemphasize moreonreusing,inferenceandchangemanagementofontology knowledgebase.
To achieve ontology modularity in a distributed scenario, different methods and schemes have been proposed. For example, E-Connection is proposed as a set of “connected” ontologies. AnE-Connected ontology contains notonlyinfor- mationaboutclasses,propertiesandtheirindividuals,butalsoa newkindofproperties,calledLinkProperties,whichestablish theconnectionbetweentheontologies[13].Anotherinteresting approach is Distributed Description Logics (DDL) framework [14]andthedistributedreasonerDRAGO(DistributedReasoning Architecture for a Galaxy of Ontologies) [15] as formal and practicaltoolsforcomposingmodularontologies.Also,thereis Package-Based Description Logics as another formalism that supportscontextualreuseofknowledgefrommultipleontology modules[16].While,thesemethodsandformalismhavemore orlesslogiccompatibility problemswhenwetrytousethem together. For example, the underlying logic formalism of E- ConnectionisOWL-DL(i.e.SHOIN);logicformalismforDDLis SHIQ;whenitcomestoPackage-BasedDescription,itturnsinto SHOIQ.Veryfewofthesemethodshavefullcompatibilityand equal logic expressiveness as OWL standard. This couldlimit largescalereasoningandmodificationbetweenheterogeneous anddistributedmodularontologies.Frompracticalperspective, these methodshave notbeenapplied in such a considerable scale.Mostofthemethodsfocusonlow-levelmodularizationof syntaxandsemanticlevel,a higherlevelconsiderationwhich
1 http://www.globalecolabelling.net/.
2 http://bioportal.bioontology.org/.
3 http://www.uniprot.org/.
4 http://leo.informea.org/.
5 http://ontologydesignpatterns.org/wiki/Main_Page.
cares more about conceptualization itself and engineering efficiencyisstilllacking.
In [17],the authorshaveidentifieda high-levelviewof the frameworkformodularity.Thedimensionsoftheframeworkare related as follows. A module's use-caseresults in modules ofa certain type. A module of a certain type is created by a modularization technique. Modularization techniques result in moduleswithcertainannotationfeaturesorproperties.Thiswork provides feasibleengineering guidance for module partition or extraction.Itseemstoworkwellwithdescriptiveconceptualiza- tion,butitdoesnotaddresshowtodealwithrules.
As for OWL ontology, the current OWL imports syntax already provides the ability of modularization to a certain extent.Itisveryinterestingtoseethatin[18]theauthorspresent usecasesformodulardevelopmentofontologiesusingtheOWL imports mechanism. For cases (Ontology organization and factoring,interfacesbetweenontologiesandbetweenontologies andsoftware,ontologylocalisation,andontologyextension.)are presented toillustratehowtomakeuseOWLsyntaxaswell as importsconstructorstobuildmodularizedontologies.Theyhave chosen to implement all modules as separate files. In our research, we have also taken the same approach. However, in their research, wehave notseen how rules areaddressed ina modulardesign.
Inthiswork,weapplyamethodusingimportssyntaxtobuild OWLontologyknowledgebasewithSWRLrules6inwhichsmaller ontologycomponentscanbemaintainedandreusedmoreeasily.
Weexpecttoexploreandfindoutsomeusefuldesignprinciples andengineeringexperienceregardingtooriginalOWLontology scheme.
Likesoftwareengineering,engineeringmethodologiesarealso requiredinontologydevelopment.Yet,inouropinion,ontology engineeringisnotasmatureassoftwareengineeringbecauseofits shorterhistoryandlimitedrelativescaleofpractice.Inspiteofthat, quiteseveralontologydevelopmentmethodshavebeenproposed, e.g.TOVE,METHONTOLOGY,DILIGENT,NeOnMethodology [19– 22].Mostofthesemethodsfollowa“waterfall”pattern.Common characteristicsthat canbegeneralized fromthese methodsare iterationandrefinement.Inourontologydevelopment,wedon’t rely on only one methodology exclusively, instead, we have adapted and customized those useful steps from all these methodologiestohaveadevelopmentmethodthatbestsuitthe Fig.1.TotalEUEco-labelproducts&servicesperproduct/servicegroup.
6 SWRLisanabbreviationforSemanticWebRuleLanguage,itextendsthesetof OWLaxiomstoincludeHorn-likerules.ItthusenablesHorn-likerulestobe combinedwithanOWLknowledgebase.Moredetailscanbefoundathttp://www.
w3.org/Submission/SWRL/.
task.Thekeystepsinourdevelopmentmethodare:requirement analysis,captureof motivatingscenariosand competencyques- tions, terminology collection, modeling, test reasoning and argumentation, evaluation and analysis. The rest part of this paper will describe these steps and present the modularized ontologiesindetail.
3. Requirementanalysis,motivatingscenariosandcompetency questions
Firstly,let'shaveabriefoverviewofthecurrenteco-labeling processforlaundrydetergentproducts.AsEUEco-labelhasbeen undergoingformorethantwentyyearsinEuropeanUnion,awell- defined coordination between the EU Commission and other membercountries’competentbodieshasbeenestablished.Onthe official web site of EU Eco-label,7 detailed documentation is providedtoenterprisestofacilitatetheapplicationprocess.Onthe same site, there is also a detailed product group catalog and correspondingcriteriaforeachproductorservicegroup.
Usually,whenanewproductorserviceisabouttobeaddedinto the product group catalog, various stakeholders and domain expertswillbeassembled.Afteracarefulsurveyanddiscussion,a technicalreportwillbedrafted.Accordingtothistechnicalreport, afeasiblecriteriawillbemadeandthenputintopracticeunderthe authorizationof EU commission. Fromtime totime, necessary reviseoramendmentstothecriteriamaybeapplied.Asaresult, the information implied in each product or service criteria becomesa complexknowledgesystemwhich involvesmultiple domains’ expertise, standards and best practice. Take laundry detergent for example, criteria is set for each of the following aspects:
1. Dosagerequirements.
2. Toxicitytoaquaticorganisms:CriticalDilutionVolume(CDV).
3. Biodegradabilityoforganics.
4. Excludedorlimitedsubstancesandmixtures.
5. Packagingrequirements.
6. Washingperformance(fitnessforuse).
7. Points.
8. Consumerinformation.
9. InformationappearingontheEUEco-label.
ThesecriteriahavebeenpublishedinCommissiondecisionof28 April2011onestablishingtheecologicalcriteriafortheawardofthe EUEco-labelforlaundrydetergent2011/264/EU.8Thiscommission decisioniscomposedofregulationarticles,annexwhereeachitem ofthecriteriaisexplained,andappendix.Theregulationarticles arenotveryinterestingasitgivesonlyadministrativedeclarations andreference.Mostoftheknowledgeaboutlaundrydetergentis elaboratedintheannexandappendix.Criterion“Dosagerequire- ments”specifiesthereferenceproductdosagerecommendedfor eachwash.Qualifieddetergentproductsshouldnotexceedcertain value.“Toxicitytoaquaticorganisms”specifiesthemaximumCDV value for qualified products. Similarly, in the next criterion
“Biodegradability of organics”, it indicates that the content of organic substances in the product that are aerobically non- biodegradable(notreadilybiodegradable)(aNBO)and/oranaero- bicallynon-biodegradable(anNBO)shallnotexceedcertainlimits.
Criterion“Excludedorlimitedsubstancesandmixtures”prohibits somesensitiveorhazardoussubstancesasingredients.“Packaging requirements”pointsoutacceptablethresholdweight/utilityratio (WUR)oftheproduct.“Washingperformance”ismoreaboutthe
product's performance test. The applicant shall provide a test reportindicatingthattheproductfulfillstheminimumrequire- mentsspecifiedinthetest.Criterion“Points”providesanindicator matrixofpoints.Eachoptionhas1or2points.Aminimumof3 points shall be achieved for a qualified product. Criterion
“Consumer information” examines if the dosage instruction, washing recommendations, or pretreatment information are properly printed on the product's package. The last criterion
“InformationappearingontheEUEco-label”isabouttheoptional textshowingontheEUEco-label.
Afterreadingandanalyzingthecriteriadocumentforlaundry detergentproducts,wehaveidentifiedtwoimportantmotivating scenariosorbasicrequirementsconcerningourontologyknowl- edge base. The first one is saving candidate product's detailed description.Forexample,someapplicantwantshisproducttobe eco-labeled, a description of the product should be provided.
Product's critical physical and chemical characteristics, param- eters, textual information or other specification should be instantiatedintheontologyandcanbequeriedafterwards.More technicallyspeaking,bothTBoxandABoxshouldbepreservedin theontologies.Theotherimportantmotivatingscenarioisjudging whethersomecandidateproductisqualifiedtobelabeledornot.
Thisscenariorequiresinferencesupportforontology.
Basedonthesetwoscenarios,somecompetencyquestionshave beendefined. Weexpectthattheontologytobedeveloped can answerquestionslike:
CQ1:Ifthisproductisqualifiedtobeeco-labeled?
CQ2:Whatisthequantitativevalueof thisproduct'scertain physicalorchemicalcharacteristics?(Criticaldilutionvolume, biodegradability,weight/utilityratio,etc.)
CQ3:Doesthisproductcontainsexcludedorlimitedsubstances andmixtures?
CQ4:Inwhichcountriesisthisproductbeingsold?
CQ5:Whatisthereferencedosageperwashforthisproduct?
CQ6:WhatisthecorrespondingEURiskPhraseforsomeGHS HazardStatement?
CQ7: What physical or chemical characteristics does some ingredienthave?Whataretheirvalues?
...
One thing that draws our attentionis that, among those 9 criteria,somearenotsuitabletobemodeledinontologies.Inour research, wehadexpectedourknowledgebasetocover allthe criteria,butwefoundthatsomecomplexcriterionisdifficulttobe translated in ontology. Because both the syntax and semantic complexity of this criterion exceed what is allowed by OWL language.Forexample,thespecificationofconsumerinformation (Criterion8)hasalmostnoquantitativeparameter'srequirement, instead,whethertheinformationshowingonthepackageisgood ornotismostlysubjecttothejudgmentofhumanexperts.Asfor thewashingperformance(Criterion6),atestreportisneeded.The production of this report must becarried out by a certificated laboratory and then reviewed by human experts too. Another exampleisthecriterionofpoints(Criterion7).Inthiscriterion,itis required to calculate the points that a candidate product accumulates. With regard to OWL 2 and SWRL which are monotonous in terms of logic, it is hard to modifyan already builtmodelordoaccumulativecalculationbyitself.Ifwetranslate such kind of criterion into OWL ontology forcefully, we may encounterverybulkyontologystructure.Becauseforeverysingle pointsitem,wemayhavetouseapropertytosavethepoints,then asetofcorrespondingruleshastobetranslatedandestablishedto calculatethepointsofthisitem.Sucheffortswillgreatlyincrease modelingcomplexityandaffectthereasoningperformance.Thus, for the sake of a betterinference performance of the decision
7 http://ec.europa.eu/environment/ecolabel/how-to-apply-for-eu-ecolabel.html.
8 http://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32011D0264.
support process, we decide to take a trade-off strategy that criterion1,2,3,4,5arechosentobetranslatedintoontology.The rest of thecriteriawill beimplemented byexternal traditional programlogic,buttheverificationresultofthesecriteriawillbe storedintheontologyknowledgebaseaswell.
4. Terminologycollection
Atfirst,wetriedtoutilizesomeOntologyLearningtechniques.
Aftersomesurveywork,Text2Onto[23]waschosentobethetool thatextractsontologyfromthecriteriadocument.Unexpectedly, the resultof Text2Onto9 was notsatisfactory. Afterparsingthe criteria document in text, only about a dozen classes were identified, two object properties wereidentified.For theother ontologylearningtools,eithernodownloadlinksareprovided,or thetoolisnotrunnable.Sinceautomaticextractionofontologydid notworkverywell.Evenweagreethatautomaticextractionmay helpinsomecases[24],wedecidedtodoitmanually.
The first critical task before modeling is to identify the terminology of the ontology. Because we have the experience that once a terminology is acquired, class definition and class hierarchywillbeeasilyretrievedfromtheterminology.Then,the definitionofobjectpropertyanddatapropertywillcorrespond- inglybecomeeasier.Inthisstep,wehaveutilizedcardsortingand ladderingtechniquesthataredescribedin[25].Usefultermswere identifiedandrecordedwhenweroughlybrowsedthedocument.
In thisfirststep,bothnounsandverbswererecorded.Multiple iterationswerecarriedouttomakesurewedon’tmissimportant terms.Then,wetriedtogroupthesetermsintodifferentcatalogs.
For example, “preservative”, “fragrance”, “stabilizer”, “coloring agent”,“substance”,and“solvent”describethingsinthesamefield, sotheyshouldbecatalogedintoasamegroup.Nextstep,weput thesegroupedtermsinto“ladders”.Inotherwords,termswere organized by“is-a” relationshipin hierarchy structureand this structurebecametheprototypemodelingofourontology.Inthe previousexample,“substance”hasamoregenericmeaning,thenit wasladderedinahigherlevelthantheothersinthehierarchy;the othertermsassociateditthrough“is-a”relationinthelowerlevel.
Atlast, areviewtoalltheselectedterms wereconductedwith domainexpertmakingsurethemodelingiscomplete.
5. Amodularizedmodeling
Since wealreadyhaveaprototypemodelingof theontology composedoftheselectedterminology.Hereinthisstep,weshould translatethemodelingintospecificontologysyntax.10Theaxioms ofclass,properties,andindividualsshouldbeinserted.Putitmore vividly,theoutputofterminologycollectionismorelikebuildinga skeleton of theontology;themodeling inthis stepis closerto enrichtheontologywithfleshandblood.Aswehavestatedinthe beginningofthispaper,averyimportantissueofourresearchis
“reuse”.Inpursuitofbetterre-usability,weproposeamodularized methodologytoseparatetheentitymodel(staticconceptualiza- tion)andrulemodel(dynamicconceptualization).Inotherwords, we shouldidentifyinwhich parttheknowledgeaboutlaundry detergent is relatively constant, and in which part frequent changes may takeplace. Asa result of this, in Fig.2,we have twokindsofmodules:oneistheentitymodulewithsolidborder line, which represents therelative staticconceptualization; the
otheristherulemodulewithdottedborderline,whichrepresents moredynamiccriterionrulesthatrelayonentitymodule.
In our design, still in Fig. 2, the main module named laundry_detergentcontainsgenericconcepts,rolesandindividuals of the domain. For the other more generic entities, module laundry_detergent reaches tothemvia dependencies.In OWL2 scheme, we can implement this dependency by using import syntax,whichmeansanontologywilluseallthoseconceptsand relationships from the imported ontology. For our laundry detergent product group, we haveentity moduleiso_standards, whichcontainsalltheISOstandardsreferences;ghs_hazard_state- ment,inwhichstoresallthehazardstatementsandcodesofGHS (Globally Harmonized System of Classification and Labeling of Chemicals);regulation_european_commission,wherestoresallthe European Commission regulation reference; european_risk_- phrases, where all relevant Europeanrisk phrases of chemicals are listed; commission_decision, which refers to all relevant European Commission decision documents; didlist, which is a database for detergent ingredients. As we have put them into independentmodules,theyareeasiertobeimportedandreused byotherdomainontologies.Pleasenotethat,althoughthemain modulelaundry_detergentimportsthesesub-modules,itdoesnot meanthatlaundry_detergentneedallthecontentinthem.Maybe onlyapartorevenaverysmallpartofcontentisusefulforthe upper-levelmodules.
5.1.ModuleLaundry_detergent
Thismoduleistheskeletonofthelaundrydetergentdomain ontology.Almostalltheimportantdomainconceptsandrelation- shipsaredefinedinthisontologymodule.Fig.3,aclassdiagramin UMLillustratesthemainclasses defined inthismodule.Onthe rightsideofthediagram,wecanfindahierarchyofthecandidate laundry detergent product and there are five kinds of laundry detergentsthatareconcernedinthiscriterion:colorsafedetergent, heavydutydetergent,lowdutydetergent,fabricsoftener,andstain remover.Thecorecandidatelaundrydetergentclassisassociated withseveralotherparameterclassesviaobjectproperties.These properties or relations are developed from the verbs that are identifiedintheterminologycollectionprocess.Objectproperties are important part for a complete laundry detergent product profile.Theseobjectpropertieslinktheotherparameterclassto thecorecandidatelaundrydetergentproductclass.Asillustrated inFig.3,eachinstanceofcandidatelaundrydetergentmusthaveat leastonekindofchemicalasingredient.Itisrequiredtospecifythe manufactureroftheproduct,thecountrieswhereitwillbesold, andtheproducttype. Eachcandidatelaundry detergentshould alsobeassociatedwithoneandonlyoneparameterinstancefor thecriticaldilutionvolume,referencedosage,weightutilityratio, aerobically non-biodegradability, and anaerobically non-biode- gradability.Foreachparameterclass,adatapropertyhasValuehas been defined in order to assign concrete value to different parameters.hasFunctionalUnitisdefinedtospecifyvariouskinds offunctionalunits(e.g.g/kgwash,ml/kgwash)forthisconcrete value.Ifsomeparametervalueofacandidatelaundrydetergent doesn’tcomplytothecriteria,itwillbecatalogedintotherejected detergentclass.
5.2.ModuleDidlist
Thismoduleistheconceptualizationofthedetergentingredi- entdatabase.InEUEco-labellaundrydetergentproductcriteria, this database is recorded in an excel file, which is not very convenienttobeusedinapplicationsorothersoftwaresystems.
This module is interesting because it will be reused in other productgroupcriteria.Wehavedevelopedanexcelscannertoread
9The version we used is here http://storage.googleapis.com/google-code- archive-downloads/v2/code.google.com/text2onto/text2onto-071109.zip.
10ThelaundrydetergentcriteriaontologycanbeaccessedonGithub:http://
github.com/xudaddd/EU-Ecolabel-laundry-detergent-product-criteria-ontology.
thisexcelfile,thengeneratedthismoduleasOWLfiles.Fig.4isthe representation of this module in UML class diagram. In this module, all the detergent ingredients are sub-classified into groups: amphoteric surfactants, anionic surfactant, cationic surfactant, non-ionic surfactant, preservative, and other ingre- dients. Various functional units are identified by scanning the wholeexcelfile.Fullnamelabelandannotationareattachedto eachofthemaccordingly.Eachingredienthasoneandonlyone anaerobicdegradationcharacteristice.g.“N”meansanaerobically notbiodegradable;Eachingredienthasoneandonlyonekindof aerobic degradation characteristic e.g. “I” means aerobically inherentlybiodegradable,butnotreadilybiodegradable.
5.3.ModuleEuropean_risk_phrases
ThismodulecoversalltheEuropeanRiskPhrasesspecification.
Since European Risk Phrases is an external standardization reference that appears in criterion 4, it's better to keep these specificationstobeanindependentmodule.Mostofthismoduleis theriskphraseindividuals. Eachriskphraseindividualhastwo datapropertyassertions,e.g.individual“R49”hasRiskCode“R49”; hasPhraseStatement“maycausecancerbyinhalation”.Thismodule isreusableinotherEUEco-labelproductgroup.
5.4.ModuleGhs_hazard_statement
SimilartopreviousmoduleEuropean_risk_phrases,GHS(Glob- ally Harmonized System of Classification and Labelling of Chemicals)isalsoanexternalreferenceincriterion4.Amapping betweenGHSstatementandEuropeanRiskPhrasesispresentedin this criterion. A module following almost the same pattern as
moduleEuropean_risk_phrasesismodularized.Mostofthismodule ishazardstatementindividuals.Eachhazardstatementindividual hastwodatapropertyassertions,e.g. individual“H261”hasHa- zardCode “H261”; hasHazardStatement “In contact with water releasesflammablegases”.Thismodulecanbereusedintheother EUEco-labelproductgroups,likeall-purposecleaners,cosmetic products.
5.5.ModuleIso_standards,Regulation_european_commission,and Commission_decision
These modules store the external documentation reference.
They record relevant EU documents, standard, commission decisionorregulationsthatarereferredinthisdetergentlaundry criteria.Thesedependencyandreferencescontributetoabetter understandingofthecriteriainabiggerpicture.Fig.5presentsthe structure of these three modules. Each of these individuals is equippedwithURLsthatlinktoexternalresources.Thesethree modulescanalsobereusedandsupplementedbyotherdomains andotherEUEco-labelproductgroups.
Several advantages exist in this modularized design. As more coherent concepts and relationships are gathered togethertoformmodules,it’llbeeasiertomanageknowledge and data in large scale. Complex conceptualization can be achievedbyintegratingmultiplesmallmodules.Also,it'seasier toconfigureand replacemodules ratherthan tomake slight changesdirectlyinalargestructure.Takethesameexamplein Fig. 2. We have a general conceptualization of laundry detergent product which is stored in domain module laun- dry_detergent.This majorontologymodulecan bereplaced by othermodulesdescribingotherproductgroupswhilestillmaking Fig.2.OntologymodularizationschemaforEUEco-labellaundrydetergentproductgroup.
use ofsub-modules likedidlist,ghs_hazard_statement and euro- pean_risk_phrases,etc.Thisactuallyhappensinatleasttwoother product groups “rinse-off cosmetic products” and “all-purpose cleaners and sanitary cleaners” which use the same detergent ingredientdatabase(Fig.6).Re-usabilityisachievedbyextracting
the common knowledge module and have it shared between domainontologies.
Modularizationimpliesseparationofconceptualization.Inour case, we can seethat it will bepractical to extractrules from ontologymodules.In otherwords,it'sbettertokeepsubjective Fig.4. StructureofModuleDidlistpresentedinUML.
Fig.3.StructureofModuleLaundry_detergentpresentedinUML.
constraints and world description separated. We call this the separation of rules and entities. For example, in the detergent ontologyshowninFig.2,ontologiesrepresentedinellipseswith solid borders are concept-centered, which means the main function of these ontology is to describe the concrete world.
These ontologies contains concepts and relationships that are
meanttodescribeorrecordthefactsabouttherealworld.Onthe otherhand,as fora productgroup's guidelineor criteria,quite muchofthisinformationisinvolvedwithhumanobjectives.They aretherules andwillingnessthat human beingsimposetothe world.Generallyspeaking,thedescriptionoftheconcreteworld doesnotchangeasmuchashuman'ssubjectivewillingnessand Fig.5. StructureofmoduleIso_standards,Regulation_european_commissionandCommission_decisionillustratedinOntoGrafProtégéplug-in.
Fig.6. Basicreusepatternwhichhappensbetweenlaundrydetergents,rinse-offcosmeticproductsandall-purposecleanersandsanitarycleaners.
rules. In ourresearch, we implement such separation between rulesandentitiesinordertoloosethecouplingbetweenthesetwo aspects, andthen realizeabetterreusability.Thisseparation of rules and entities is a significant difference between our modularizationmethodandpreviousones.
Fordetergentproducts,theconcentrationofdifferentchemical ingredientshastocomplywithcertainlimitandstandard.Wecan hardlysaythatsuchgoal-orientedspecificationisplaindescription oftheworld.Moreover,suchrulesmaychangetimeaftertime.This actually happens, because the product guideline keeps being updatedasEUCommissionkeepsgeneratingnewamendmentsor revise. In our approach, we have each criterion item be an independentmodule(notcompletelyindependentactually,asthese rule modules may also havedependencies to other external or internalontologymodules).Forexample,eachofthe5criterionof thelaundrydetergentproductgroupismadeintoanindependent OWLfile.IntheOWLfile,firstly,thefundamentalentitymodulesare imported(inFig.2,module hierarchywhoseroot islaundry_de- tergentisimportedbyallthefivecriterion),thenSWRLruleaxioms areinserted.Aseachcriterionisdistributedinitscorresponding module alone, wecan easilyreplace themwithnew rules and managetheminaconfigurablewaywithoutimpactingtheothers.
Atlastbutnottheleast,forthecriteriaontologyasawhole,an entry module is introducedto include all the criteria, e.g. the laundry_detergent_criteriamoduleontherightsideofFig.2.For applications, once the ontology entry is provided, the whole ontology composed of all the entityand rule modules will be retrieved.Withthisconfigurabledesign,expansionandalteration totheontologywillbeeasier.Forexample,whenanewcriterionis abouttobeapprovedbythecommission,inFig.7,wecanupdate theproductcriteriatoanewversionbyaddinganewrulemodule andnewentrycalledLaundry_detergent_criteria_2.0withoutlosing traceofthepreviousone.Thenewlyaddedrulemodulecouldbe
aboutanothernewcriterionorjustanupdateversionofexistent criteria.Theremovalofcertainmoduleissimilar,allweneedisto introduceanotherentrymodule.For example,if thenewentry moduleimportscriterion2,3,4,5,thuscriterion1willberemoved fromthisversionofcriteriaontology.
Inthissubsection,we’veintroducedamodularizedmodelingof EUEco-labellaundrydetergentproductcriteria.Theseparationof entitymodulesandrulesisoneofthemajorcontributionsofour work.Themainreferencesourcesusedinthisworkaretheofficial criteriadocumentationforlaundrydetergentproductgroup.This documentationconsistsofmultiplePDFfiles(aboutfortypagesin all).Twodevelopersandanexpertineco-labeling(Certification engineerforDetergentproductsinaprivatecompany)areinvolved inthemodelingand developmentprocess.For therequirement analysisprocess,wehaveconducteda carefulreadingofallthe documentation which took about one week of time. The terminology collection process took about three weeks. The modulepartitionandentitymodulemodelingtookusaboutfour weeks.Thenwealmostspend double timei.e. 8weeks for the translation and modeling of the SWRL rules. For each step we adoptedanagilemethodologyinwhichweusedDevOPSpractice (Developmentand Operations)[26]. Weinitiatewhatwecalled DevExp(Developmentand expert) loop throughfeedback from experttodevelopers.Thegoalistoamplifythefeedbackloopso thattheprocessisswiftandseamless.Thefeedbackloopledtoan increaseinefficiencyoftheontologyconstruction.Theroleofthe expert was to check the output, identify problems if any and validateeachstep.
6. Reasoningandargumentation
Aconsiderableadvantageofusing OWLontologyis that the underlying DL (Description Logic) formalism allows reasoning.
Fig.7.Ontologyexpansionbyaddingnewrulemoduleofcriteriafordetergentproductgroup.
Actually,theDLcomputationcomplexityandthedevelopmentof reasoners are very important research issues for ontology and ontology engineering. Investigating the trade-off between the expressivityofDLsandthecomplexityoftheirinferenceproblems hasbeenoneofthemostimportantissuesinDLresearch[27].As fortheexpressivenessofOWL,it'smostlyrelatedtotheunderlying DLexpressiveness.ConcerningthelatestW3Cdiscussion,thereare threeOWLschemesondifferentlevelsofexpressiveness:OWL- Lite, OWL-DL, OWL-Full [28]. OWL-DL is best supported by reasoners becauseit is decidablewhich meansa propertrade- off betweenreasoning performanceand expressivity. A famous algorithmimplementedbymanymodernDLreasonersiscalled TableauAlgorithm[27].
In Protégé editor, several third-party reasoners have been developedas plug-ins. In fact,today'sreasoners canalso stand alongasAPIsorevenindependenttools.SinceProtégéisanopen sourceproject,foralmostallitsreasonerplug-ins,wecanfindAPIs thatcanbeintegratedintoprogramminglanguagelikeJavaorC++.
Some common reasoners for Protégé (the version we used is Protégé5.0.0beta24):FaCT++isasoundandcompletereasonerfor SHOIQ(thesamedescriptionlogicunderlyingOWL-DL)[29].Pellet isalsoasoundandcompletereasonerthatissaidtosupportE- Connectionsandthatwouldbeveryinterestingforourresearch [30].Hermit[31]worksbestwithourontologyknowledgebaseas fortheSWRLrules,allthereasoningtasksinvolvedinthispaperis completedbyHermit(Theversionweusedis1.3.8.413).
Now,let'shavealookattheSWRLrules.Asstatedinprevious sections,aftercheckingallthecriteriainthelaundry detergent product,wefoundthatonlythefirstfivecriteriaarepropertobe translatedintoSWRLrules.ThemainfunctionorobjectiveofSWRL rulesisfordeterminingwhetheracandidateshouldberejectedor accepted.Theyaremanuallytranslatedbyontologydevelopers.For thecheckandvalidationoftheserules,reasonerwillbeusedtosee ifthereissyntaxorvariableerrors.Iferrorsorinconsistencyexist in the rules, reasoning process will be blocked. For the final validation of the rules, we will applya reasoning comparison.
Besidestherulesandreasoningprocess,wewillconductmanual evaluationinwhichhumanbeingsreadthecriteriadocumentation andchecktheproduct'sprofile,thencomparethereasoningresult withthemanualevaluationresult.Iftheyhavethesameresults,it provesthattheruleshavebeencorrectlytranslatedandmodeled.
Morespecifically,takethefirstcriterionforexample,itisabout therecommendeddosageofdetergentforeachwash.Thedetailsof thiscriterionisshowninFig.8.Foreachtypeofproduct,sincethe valuefor“powder/tablet”and“liquid/gel”isthesame,wemerge thetwo requirementsinto one.In Protégé, theSWRL rules are editedinatabasinFig.9.It'swritteninthepopularManchester Syntax[32].Pleasenotethatherewehaveanotheradvantageof dividingrulesintomodules.Forthatreasoningisaprettycostly computationtask,itwillbeinterestingtomakereasoningseparate anddistributed.In ourmodularizationof ontologies,byputting SWRL rules in different modules, unnecessary interference
between rules is avoided. For example, some domain experts finisheditingcriterionNo.1andhewantssometest,allheneedsto do istochoosetherule moduleofcriterion No.1and start the reasoner. The reasoning will be based only on criterion No.1 becausetheothercriteriarulesarestoredintheotherrulemodules andareexemptedfromcurrentontologycompositionandtest.
Weassumethereadershavebasicideasaboutthesyntaxand semanticsofSWRL.(AgoodreferenceofSWRLspecificationcanbe foundontheW3Cwebsite11)Thebasicideaforthecriteriarulesis introducingtwoconceptscalledRejectedDetergentandCandidate- LaundryDetergent.Aslongastheprofileofsomedetergentproduct doesn’t comply with the criteria rules, this product should be classifiedasanindividualorinstanceofRejectedDetergent.Inother words,thisclasscanbetreatedasthe“objective”ofthereasoning task.In thebeginningoftheeco-labelingreasoningprocess,we inputproductprofileasindividualoftheCandidateLaundryDeter- gentclass,oncethereasoningprocessisstarted,criteriarulesare appliedupon it.Afterthereasoning, ifanproductindividualis classifiedundertheclassofRejectedDetergent,thenweassertthat thisproductdoesn’tcomplywithEUEco-labelcriteria.
In practice, thecriteria ontology will worksomehow like a templateforreallaundrydetergentproductprofile.Anewproduct profile imports the criteria ontologyentry(module Laundry_de- tergent_criteriain Fig.2 forexample), thena detergent product profile ontology is constructed according to the pre-defined specificationinthecriteriaontology.Afterthereasoning,allthe reasoning and inference result of this profile ontology will be storedinourknowledgebaseasreferencecasesforfurtherreuseor review.Thus,theknowledgebasewillbecomposedmainlyoftwo parts:acriteriaontologyrepositorythatstoresallkindsofEUEco- labeling products’ criteria in modularized ontologies; and a historical case repository that reserves allthe product profiles’ reasoningresults.
Intherestpartofthissection,asimpleproductprofileexample willbepresentedtoshowhowexplanationisgeneratedattheend ofreasoning.Typically,wesavea candidatedetergentprofilein ManchesterSyntax12whichisillustratedinFig.10.Oneproduct couldbemarketedacrossseveraldifferentEuropeancountriesat thesametimeandthisfactisexpressedintheaxiomexpression isMarketedInmin1Country.isMarketedInisanobjectpropertythat we’vedefined and minis therestriction typewhich meansthe cardinality of this property is at least one. Table 1 shows the product'sparametersindetail.
Comparedwiththecriteriavalue,thetwoknowningredientsin Table1don’thaveanyhazardcode,neitheraretheyinthelistof excludedorlimitedsubstances.Thismeanstheyaregoodtobe added into laundry detergent products. However, some of this product's parameter value exceeds the criteria value, e.g.
Fig.8.Criterionofdosagerequirements.Therecommendeddosageforeachwashshallnotexceedtheamountsabove.
11http://www.w3.org/Submission/SWRL/.
12 http://www.w3.org/TR/owl2-manchester-syntax/.
recommended dosageand weight utility ratio,so it shouldbe consideredasaRejectedDetergent.AfterlaunchingHermitreason- er,wecangetaninferredclasshierarchyshownontheleftsideof Fig.11.Wefindthat ourexampleindividualheavy-dutylaundry detergent example NO.0 has been classified under the concept RejectedDetergent.All theaxiomswithyellowbackgroundcolor indicatesthattheyareinferredbythereasoner.Ifweclickonthe smallquestionmarksuitedontherightsideofeachnewlyinferred axiom,wecanchecktheexplanationsofhowthereasonerreaches tothisnewinference.Fig.12showsseveralexplanationitemsfor thisnewinferenceresultandwhyourexampleproductprofileis classifiedintotheRejectedDetergent.FromexplanationNo.1,wecan seethatourexamplebreakstheruleofrecommendeddosageand explanationNo.2isabouttheweightutilityratiorule.Inourcase, thereare 15explanationitemsfound. If wescroll downin the windowshowninFig.12,wecanfindalltheothers.
7. Evaluationandanalysis
Thelaundrydetergentontologyisthefirstcriteriaontologythat wehavedevelopedfor EUEco-labeling.Another twoimportant criteria ontologies about rinse-off cosmetic product and all- purposecleanerareunderdevelopment.Alltheseontologieswill be included in a knowledge base framework. Adjustment and improvementinfavorofglobalperformancearebeingtakeninto account.Evaluationofsingleontologyandthewholeknowledge base is also undergoing. The advantage of the design of modularizationandseparationhasbeenobservedbyresearchers as module Didlist, module European_risk_phrases, etc. can be directlyreusedbynewlydevelopedontologies.
Aswehavepresentedinrequirementanalysissection,avery importantmotivation ofthis ontologydevelopment is tojudge Fig.9.SWRLruleseditedinProtégéeditor.
Fig.10.CoreconceptHeavyDutyDetergentdefinedinManchesterSyntaxaxiomsin Protégéeditor.
Table1
Detailedparametersofproductprofileexample:heavy-dutylaundrydetergentexampleNO.0.
Propertyparameter Value Criteriavalue
Producttype Liquid
Recommendeddosage(referencedosage) 20.0ml/kgwash 17.0ml/kgwash
Salescountry France
Weightutilityratio(WUR) 2.0g/kgwash 1.5g/kgwash
Criticaldilutionvolume(CDV) 30,000.0l/kgwash 35,000.0l/kgwash
Aerobicallynon-biodegradability(aNBO) 0.5g/kgwash 0.55g/kgwash
Anaerobicallynon-biodegradability(anNBO) 0.5g/kgwash 0.7g/kgwash
Knowningredient Aceticacid;C8-18-Amphoacetates
whetheracandidateproductisqualifiedtobelabeled.According toaclassificationofontologyevaluationapproachesin[33],our evaluationapproachisclosertoanapplication-basedevaluation method,i.e.usingtheontologyinanapplicationandevaluatingthe results.Wehaveseenthatthislaundrydetergentcriteriaontology is successfully applied in a decision-support system in [34].
Synthetically, taking into account the criteria and aspects introducedin[35]and[36],wehaveevaluationresultasfollowing:
Syntax The criteria ontology is described in standard OWL syntax.
Semantics.SinceSWRLrulesaredefinedintheontologyand inferencesupportisabasicrequirementofourontology,multiple reasonerse.g.Fact++,Hermit,andPellethavebeenappliedtocheck and verifythesemanticconsistency. So,the ontologyis always logicallyconsistent.
Vocabulary.Almostalltheclasses,propertiesandindividualsin theontologyhavea meaningfulidentifierwhichfollowsCamel case naming pattern. For those entities that have abbreviation namesandvaguemeaningnamese.g.CDVandH400,ardfs:label axiomisaddedascomplement.
Structure.Thestructureofourontologyisrelativelysimple,the depthofbothclasshierarchyandpropertyhierarchyisnotmore thantwo.Themostoutdegreeforanindividualthatreaches to otherindividualsviapropertiesis14.Takingallthemodulesinto account,68classes,46objectproperties,21datapropertiesand 460individualsaredefinedandstoredinourontology.Thenumber oftotalaxiomsis5786.DLexpressivityisALCHQ(D).Ourontologies canbeeasilyunderstoodand manipulatedbyotherknowledge engineers.
Documentation. Each module of the ontology has a textual annotation.Forthosekeytermsthatcomefromspecificdomain glossaries,textualannotationandexternallinksareprovided.For everySWRLrule,annotationaswellasthecorrespondinganchor positioninthedocumentisindicated.
As regards to more specific validation, the competency questionsthataredefined inrequirement analysissectionhave been translatedinto SPARQL queries. They work fine with our
ontologyandcorrectresultcanbequeried.Herearetwoexamples aslistedbelow(Figs.13and14).
CQ1:Ifthisproductisqualifiedtobeeco-labeled?
CQ2:What isthe valueof this product's certainphysicalor chemicalcharacteristics?(criticaldilutionvolume,biodegradabil- ity,weight/utilityratio,etc.)
Besidesthisintuitiveevaluation,wehavealsoappliedamore systematic evaluation method that is presented in [37]. Three typesofevaluationmeasureshavebeenidentifiedandconsidered:
structuralmeasures,functionalmeasures,andusability-profiling measures.Inpractice,moreprinciplesandparametersareusedto reflectthequalityofontology:
a. Cognitiveergonomics:
Depth:Maximum3.
Breadth:Maximum12forclasses;maximum115forindividua- ls.Tangledness:Low.
Class/propertyratio:1.01(68/67).
Annotations:49.
Anonymousclasses:None.
b. Transparency:
Modularitydesign:12modules(7entitymodulesand5 rule modules).
Axiom/classratio:85.09(5786/68).
Patterns:No.
Specificdifferences:No.
Accuracy:Good.
Complexity:Medium.
c. Computationalintegrityandefficiency:
Logicalconsistency:Good.
Disjointnessratio:0.97(66/68).
Restrictions:Welldefinedandannotated.
Cycles:None.
d. Meta-levelintegrity:
Meta-levelconsistency:Good.
Tangledness:Low.
e. Flexibility:
Fig.11.ReasoningresultshowninProtégéeditor.
Modularity:Good.
Partitioning:Functionalpartitionandentity&ruleseparation.
Context-boundedness:Unknown.
f. Compliancetoexpertise:
Precision:Medium.
Recall:Good.
Accuracy:Good.
g. Compliancetoproceduresformapping,extension,integration, adaptation:
Accuracy:Good.
Recognitionannotations(esp.lexical):16.
Modularity:Excellent.
Tangledness:Low.
h. Organizationalfitness:
Organizationaldesignannotations:13.
Commercial/legalannotations:None.
Usersatisfaction:Good.
From the evaluation results, we know that our ontology is competentforthelaundrydetergentproductevaluationtask.In spiteofthat,therearestillsomeaspectsthatneedimprovement, e.g.thereisnoexistentpatternsreusedinourontologydesignand moreannotationsarestillneeded.Ifreusedinothercontexts,how wouldourontologyandmodulesreactisstillunknown.Wewill keepworkonthesedrawbacksinthefuture.
Fig.12.Reasoner'sexplanationtowhyheavy-dutylaundrydetergentexampleNO.0isnotgood.
Fig.13.QueryanswerforCQ1,theresultshowstheIDofproduct.
8. Experienceandlessonslearned
Bydevelopingthismodularizedontologyknowledgebase,we have acquired some interesting experience and lessons about ontologydesignandapplication.Asfaraswecansee,peoplehave been trying to build more and more complex knowledge representation. If we take documents, which are written in whateverlanguage,asamodelorrepresentationofknowledge.To someextent,developingontologyislikeatranslationprocessthat translates models of human language to formal knowledge representation which can be accessible by machines. As the expressiveness of human language is very high, a computable modelingandtranslatingschemethathascompetentexpressive- nessisneeded.Theexpressivenessand modelingcomplexityof ontologylanguagehasbeenincreasing.Wecanseethisfromthe evolutionofOWLtoOWL2.Itisalsoobservedthat,intheearly daysofontologyresearch,simpleknowledgecontente.g.medical terminologyoftenusedtobetheobjectofstudy.Today,complex documentse.g. specifications, legal terms, executive ordersare expected to be made into ontology. In order to handle more complex knowledge representation or modeling in human language, more comprehensive consideration should be taken into account. The entity-rule separation pattern as well as modularization is such kinds of consideration and exploration thattrytohandlesuch more andmore sophisticatedmodeling tasks. As we have discussed in the beginning of Section 5, descriptiveentity-relatedknowledgeisrelativelyconstantwhich
meanstheydon’tchangeverymuch.While,thesubjectiverule- relatedknowledgepartcouldbealteredfrequently.Weputthemin separationinordertobettermanageandcontrolthechange.The philosophy generalized from this entity-rule separation and modularization pattern can be applied into other modeling or applicationdomain.Whendealingwithcriteriaalikeknowledge representation,wecanapplythisentity-ruleseparationpatternto modeldescriptiveentity-relatedknowledgeandsubjectiverule- relatedknowledgeintodifferentmodels,whichwillfacilitatereuse andmaintenance.
Fig. 15 is a more detailed mind map specification for the applicationofthisentity-ruleseparationpattern.Thepointofour learnedlessonisthat beforedivingintotheconcretemodeling, higher level abstraction and conceptualization should take precedence. In our case, the target documentation is the Eco- labelingcriteria.Accordingtothecharacteristicsofthedocument andthedomainknowledge,modularizationschemebasedonthe entity-ruleseparationpatternisproposed.Then,ineachmodule, theconcretemodelingandpotentialreuseproceed.However,in reality, the boundary of each taskcould not bevery clear. For example,reusabilityisaveryimportantfactorwhenwedecideto setupDidlist,Ghs_hazard_statementandEuropean_risk_phrases.In evenmoregeneralizedcasesand otherdomains,othermodula- rizationschemesarealsopossible.Itdependsontheobjectiveand applicationscenarioofthemodeling.However,inourresearch,we haveseenthat,insteadofdirectandprematuremodeling,extra work before that is in favor of a good ontology quality and Fig.14.QueryanswerforCQ2.
Fig.15.Beforemodelingtheontologyorevenreusing,ahighlevelabstractione.g.entity-ruleseparationandmoduledivisionisneededsometimes.
