A User-Aware and Semantic Approach for Enterprise Search

(1)

DOI: 10.4018/IJSWIS.2018100107

A User-Aware and Semantic

Approach for Enterprise Search

Giacomo Cabri, University of Modena and Reggio Emilia, Modena, Italy

Riccardo Martoglia, University of Modena and Reggio Emilia, Modena, Italy

ABSTRACT Thisarticledescribeshowinadditiontogeneralpurposessearchengines,specializedsearchengines haveappearedandhavegainedtheirpartofthemarket.Anenterprisesearchengineenablesthesearch insidetheenterpriseinformation,mainlywebpagesbutalsootherkindsofdocuments;thesearchis performedbypeopleinsidetheenterpriseorbycustomers.Thisarticleproposesanenterprisesearch enginecalledAMBIT1_{-SEthatreliesontwoenhancements:first,itisuser-awareinthesensethatit} takesintoconsiderationtheprofileoftheusersthatperformthequery;second,itexploitssemantic techniquestoconsidernotonlyexactmatchesbutalsosynonymsandrelatedterms.Itperformstwo mainactivities:(1)informationprocessingtoanalysethedocumentsandbuildtheuserprofileand(2) searchandretrievaltosearchforinformationthatmatchesuser’squeryandprofile.Anexperimental evaluationoftheproposedapproachisperformedondifferentrealwebsites,showingitsbenefits overotherwell-establishedapproaches. KEyWoRdS

Enterprise Search Engine, Information Retrieval, Semantic Knowledge and Similarity, Text Analysis, User-Awareness INTRodUCTIoN Enterprisesproduceandrelyonalargeamountofinformation.Asmallpartofinformationisavailable bypublicwebsites,whilethemostpartisexploitedbyemployeesoftheenterpriseitselfbymeansof intranet,andbycustomersoftheenterprisewhohaveaccesstosomeinformationforbusinesspurposes. Inthisscenario,thecapabilityofsearchingforneededinformationplaysafundamentalrole.On theonehand,enablinginternalemployeestofindtheneededinformationinashorttimeisnotonly usefultospeeduptheirwork,butalsotoavoidordecreasethefrustrationoflongandunsuccessful searches.Ontheotherhand,preciseandrelevantanswerstocustomersthatexploitthecompanyweb sitesforbothsearchingandinteractingcangrantahighdegreeofcustomersatisfaction. Inthisscenario,theauthorspointouttwoaspectsthatcanimprovetheuseofsearchenginesinan enterprisecontextbyprovidingmorerelevantsearchresults:user-awarenessandsemantics.Existing studiesandsurveysingeneralinformationmanagementcontextshavehighlightedthebenefitsthat canbebroughttosearchresultsbytheformer(Xiangetal.,2010)andlatter(Mangold,2007).User-awarenessmeanstoexploittheknowledgeoftheuserintermsofprofileandcontexttoeffectively tailorthesearchonthebaseoftheavailableinformation.Semanticscanbeusefultoovercomethe limitationsofasyntacticapproach,whichisoftenexploitedbutdoesnotconsidersimilarpiecesof informationexpressedindifferentways.Asfarastheauthorsknow,therearenoenterprisesearch enginesthatexploitbothaspectsinasingleapproach.

(2)

Startingfromtheaboveconsiderations,thegeneralsemanticfoundationsintroducedin(Martoglia, 2015)areexploitedtoproceedtowardsthegoalofachievingauser-awaresemanticenterprisesearch engine.ThesearchengineiscalledAMBIT-SE(AMBITSearchEngine).Itisnotagenericsearch engine,butasearchenginededicatedtosearchesinanenterprisescenario.TheAMBIT-SEapproach improvessearchresultsby: • Takingadvantageoftextualinformation,includinguserinformation.Indeed,textistheprimary componentofthedocumentsthatshouldbepresented/suggestedtousers,andalsooneofthe maininformationcharacterizinguserprofiles.Consider,forinstance,thecontentsofuserbrowsing history,thedescriptionofusers’interests,andsoon; • Exploitingsemantictechniques:insteadofapuresyntacticmatchingbetweenthequerykeywords andthewordsintheavailabledocuments,itreliesontheirmeaningandtakesintoaccount synonymsandrelatedterms. Theapproachpresentedinthispaperbringsthefollowingnovelcontributionswithregardto theinitialideaofanenterprisesearchenginesketchedin(Cabri,2016)andtothestateoftheart: • Differentlyfromthestateoftheartonavailableenterprisesearchengines,itisabletocombine semanticsanduser-awarenesswithoutrequiringanymanualwork(e.g.forannotatingdocuments, describinguserprofiles,etc.).Novelsemanticanduser-awaretechniquesallowtheenginetogo beyondstandardsyntacticsearchinacompletelyautomaticway; • Thesemantictextanalysistechniquesareadaptedandrefinedfrompreviousauthors’studies ontheeffectivenessofsemantictextmanagementinspecificsubjectareassuchassoftware engineering(Bergamaschietal.,2015;Martoglia,2011),agricultural(Beneventanoetal.,2016) anduser-centricculturalenhancementdata(Martoglia,2015).Inthepresentedapproach,the techniquesaregeneralizedtoworkinanon-specializedenterprisesearchsettingwithgeneral purposeontologiesandnewweightingschemes,allowingthemtobedirectlycomparedtothe newclasssimilaritycontribution; • Thestrengthofthesemantictextanalysiscontributionisaddedtothenewcontributiongiven bysemanticcategorization.Categorizationclassifiesdocumentsonthebasisofawell-known taxonomy(definedbyIPTC2_{)inordertoprovideimprovementsintheretrievaleffectiveness.} Tothisend,anovelclasssimilaritymetricisintroduced,withaweightingschemeexploiting thenovelconceptofinverse(document)classfrequency; • Anovelrankingselection/fusiontechniqueisemployed,producingafinaldocumentranking which:(1)reflectsboththequeryanduserprofileinaflexibleandcustomizableproportion;(b2 fusesbothclassandtextsimilaritycontributionsinthecasetheyarejudgedassignificant;(3) otherwise,itisabletoautomaticallyexcludeoneofthetwocontributionsfromthefinalresult. Thecombinedtextanalysis,user-awareandsemanticretrievaltechniquesultimatelyprovide enhancedsearchingeffectivenessoverstandardsearchtechniques,asalsoshownintheexperimental tests.Moreover,theapproachisdevisedforITSmallandMedium-sizedEnterprises(SMEs),providing themwitheasy-to-applymethodsthatallowthemtoqueryfortheinformationtheyneedintheway theyareusedto. Thispaperisorganizedasfollows.First,arelatedworkanalysisispresented(Section“Related Work”).Then,Section“SearchEngineArchitecture”describesthearchitectureoftheproposed searchengine.Section“InformationProcessing”explainshowthepresentedapproachprocessesthe documentsthatcanbe“searchable”bytheusers.Section“SearchandRetrieval”illustrateshowthe systemdefinesarankingofretrieveddocumentstosatisfytheuser’squery.Finally,theresultsofthe experimentscarriedoutonthesystemarereported,beforetheconclusions(Section“Conclusions”),

(3)

RELATEd WoRK Thissectionpresentsareviewoftheexistingapproachesrelatedtouser-awaresemanticenterprise searchengines.Theauthorshaveconsideredbothacademicandcommercialapproaches,andhave classifiedthemonthebaseoftwoaspects:theuser-awarenessandthesemantics;theresulting taxonomyisreportedinFigure1. Mostoftheapproachesdonotconsidertogethertheseaspectsand/orcannotbeclassified asenterprisesearchengines.Forthesereasons,relatedworkwillbepresentedinthreeseparate subsections:semanticapproaches,user-awareapproachesandenterprisesearchengines. Semantic Approaches TheSemanticWebislikelytobethefieldwheremostoftheapproachesforsemanticdocument retrievalhasbeenproposed,asreportedin(Mangold,2007).Itisasurveywhichcoversapproaches thatexploitdomainknowledgetoprocesssearchrequests.Theauthorsdiscussalargevarietyofdomain knowledgeutilizationthatincludeautomaticqueryexpansionandontology-drivendocumentretrieval. Thepossibleineffectivenessofinformationretrievalsystemsismainlyduetotheinaccuracy withwhichaqueryformedbyafewkeywordsmodelstheactualuserinformationneed.Onewell knownmethodtosolvethisproblemisautomaticqueryexpansion,wherebytheuser’soriginalquery isaugmentedbynewfeatureswithasimilarmeaning(Carpineto&Romano,2012).Differentlyfrom theapproachpresentedinthispaper,complexqueryexpansiontechniques,suchastheonesdiscussed, usuallyrequiredifferentparameterstobespecified(asalsostatedin(Abdou&Savoy,2008)).For instance,themethodproposedin(Voorhees,1994)involvesasetofparametersspecifyingforeach runandforeachrelationtypeincludedintheontologythemaximumlengthofachainofthattypeof linkthatmaybefollowed.Generally,thereisnosingletheorycapableoffindingthemostappropriate values(Abdou&Savoy,2008)andthereforealongprocessofmanualtuningisneeded.

(4)

Moreandmoredocumentretrievalsystemsmakeuseofontologiestohelpusersbetterspecify theirinformationneedsandproducesemanticrepresentationsofdocuments.(Haslhoferetal.,2013) proposesaSimpleKnowledgeOrganizationSystem(SKOS)basedtermexpansionandscoring techniquethatleverageslabelsandsemanticrelationshipsofSKOSconceptdefinitions.TheHybrid spreadingactivationapproach(Rochaetal.,2004)requirestightcouplingbetweenthedocumentbase andtheontology,whichisagraphwhereconceptsandpropertiesarenodesandedges,respectively. Thesetofnodesthatmatchthegivenquerytermsisusedasthestartnodesofaspreadingactivation algorithm:documentswithhighestactivationarerankedhighestintheresultset.Differentlyfrom thepresentedapproach,thementionedsystemshavenonotionofusercontext. Giventheneedformanualintervention,typicalsemanticretrievaltechniquesobtainagood degreeofeffectivenessonlyonmanuallyannotatedcollectionsand/orwithexplicituserintervention. (Savoy,2005)comparetheretrievaleffectivenessofdifferentsearchmodelsinabibliographicdatabase context.Thesemodelsarefoundedonautomaticsyntactictext-wordindexingoronmanuallyassigned controlleddescriptors.(Thesprasith&Jaruskulchai,2014)proposesaqueryexpansiontechnique workingonMEDLINEdocumentsthathavebeenmanuallyassignedtocontrolledMeSH(Medical SubjectHeadings)vocabularies.TheindexingandretrievalapproachproposedbyAMBIT-SE,instead, exploitsthesemanticsofthetextwhileremainingcompletelyautomatic. (DeVochtetal.,2017)presentsasemanticsearchenginefocusinginparticularonintegration ofdifferentsourcesofdatainthescienceresearchfield.Toincreasetheprecisionoftheresults,the authorsannotatedandinterlinkedstructuredresearchdatawithontologiesfromvariousrepositories exploitingasemanticmodel.Thatapproachdoesnotconsideruser-awarenessandrequiresannotation. (Figueroa&Neumann,2016)focusesonthesearchinthecontextofCommunityquestion answering(cQA)platforms.Itinducesthesemanticclassesofquestion-likesearchqueriesbymeans ofthecontextualinformation.Contextissetuporrepresentedbyinferredviewsoftheirrespective searchsessions,namelyviewsmodellingpreviousqueriesenteredbythesameuser.Theideaof introducingsemanticclassesandofcombiningthemwithcontextualinformationisveryinteresting, butinthatworkislimitedtoaspecificfield(cQA). SINA(Shekarpouretal.,2015)isascalablekeywordsearchsystemthatcanansweruserqueries bytransforminguser-suppliedkeywordsornatural-languagesqueriesintoconjunctiveSPARQL queriesoverasetofinterlinkeddatasources.Inthiscase,differentlyfromthescenarioconsidered inthispaper,dataareexpressedingraphformat.Thesystemexploitssemanticstoimprovesearch overdifferentdatasources,butdoesnottakeintoconsiderationuserinformationtorefinequeries. User-Aware Approaches Theadvantagesoftakingintoconsiderationthecontexthasbeenpointoutseveraltimesinthe literature(Bolchinietal.,2011;Cabrietal.,2003;Falcarinetal.,2013;Xiangetal.,2010;Vuetal., 2017).Inparticular,afewworksconcerncontextmodelling,representation,andeffectivehandling. Forinstance,(Bolchinietal.,2011)proposestodesignacontextmanagementsystemwhichis notapplication-dependent,(Falcarinetal.,2013)proposesanarchitecturalframeworkforcontext datamanagement,while(Villegas&Müller2010)reportstheresultofastudyonvariouscontext modellingandmanagementapproaches.(Xiangetal.,2010)addressestheproblemofintegrating contextinformationintoarankingmodel.(Liuetal.,2004)proposesamethodtoderiveauserprofile basedonthesearchhistoryandonpre-determinedcategoryhierarchies.(Vuetal.,2017)proposesa personalisedquerysuggestionframeworkforIntranetsearch,relyingontwotemporaluserprofiles. Mostoftheseapproachesprimarilyfocusonspecificaspectssuchasexternalconditionsor location,theydonotconsiderthesemanticsofthecontextand/ortheyrelyonmanualworkinorder toclassifyandcategorizeusersanddocuments.Generalpurposesearchengines,suchasGoogle, typicallyprovideonlyverysimplelocalizationofsearchresultsonthebaseoftheIP-address.

(5)

Enterprise Search Engines

Manyenterprisesearchengineshavebeenproposedintheliteratureoraresoldbycompanies.Most oftheproposalsdonotprovideanontology-basedsemanticanalysis,relyinginsteadonsyntactic andhand-codedrules.SomeexamplesareAlfresco3_,Autonomy4_,Solr5_.Google6_{isalsoanexample} ofasyntacticsearchenginethatcanbeexploitedinaenterprisesearchcontext. Thereare,ofcourse,exceptionstothisrule.TheSHOEproject(Heflin&Hendler,2000),which requiresadomain-ontologywheredocumenttypescorrespondtoontologyconcepts.Attivio7_,aproduct whichmanagesinformationintheRDFformat,providingsearchresultsandalerts.ExpertSystem’s Cogito8_{,whichprovidesautomateddisambiguation,classification,entityextraction,andmetadata.} Nevertheless,thesesystemsprovidenonotionofusercontext.Instead,Coveo9_{isatoolspecifically} orientedtoexploitcontextualknowledgefordealingwithinformationrelatedtocustomersandagents, butitdoesnotexploitsemanticinformation. Consideringthesemanticandcontextinformation,therearefewsystemsthatexploitit,evenifin asometimes-limitedway.TheOntogatorsystem(Hyvonenetal.,2003),partofanimagemanagement andretrievalsystem,providesaninteractiverecommendationsystemthatallowstheusertobrowse imagesbasedonontologicalproperties.Toexploitusercontexts,itintroducesviewstotheontology thatrelyondifferentconcepthierarchies,called“facets”.Eachviewrepresentsaspecificinformation-need.PrEmISES,proposedin(Ramona-Cristinaetal.,2016),isaframeworkthataimsataddressing informationmanagementneedsofSMEsrelyingonontologiestoaddsemanticallyenabledinformation integration.Theframeworkdefinitionisstillinaverypreliminaryphase;therefore,semanticand context-sensitivefeaturesarecurrentlyonlysketched.AnotherexampleisIBM’sContentAnalytics withEnterpriseSearch10_{,whichexploitsaframeworkcalledUnstructuredInformationManagement} Architecture(UIMA),inordertobuildanalyticapplicationsandtofindmeanings,relationships andrelevantfactshiddeninunstructuredtext.Contextinformationisprovidedbymeansofmanual annotations.Theseapproachesrequiremanualinterventiononthedocumentsand/oradoptastill limitednotionofcontext,i.e.theydonotexploitallofthedatapotentiallyavailablerelatedtothe user,suchasthecontentsofanywebpagevisited,attachmentdownloaded,andsimilardocuments. dISCUSSIoN Theauthorshavereportedseveralresearchesthatareconnectedtotheapproachpresentedinthis paper.Theypointoutthatthereisnoexistingapproachthataddressesallthefollowingaspectsat thesametime: • Semantics:Exploitationofsemantictechniquestoimprovetheresultsofthequery; • User-awareness:Exploitationofuserinformationtocustomizetheresultsofthequeries; • Generality:Capabilityofbeingappliedtogeneralsourcesofinformation,nottoaspecificfield; • Automation:Noneedformanualannotationoftheinformation. Startingfromtheaboveanalysis,theAMBIT-SEapproachaimsataddressingalltheabove-mentionedaspects.Inthenextsection,thearchitectureofAMBIT-SEispresented.

SEARCH ENGINE ARCHITECTURE

Figure2showsthearchitectureoftheproposedsearchengine,themainactivitiesandtherelated modules.In(Cabrietal.,2016)theauthorstackledindetailmanytextpre-processingissuessuch crawlingtechniques,paragraphidentificationandtexttagging.Inthispaper,thefocusisondefining thesemantictextanalysisandthenovelsemanticcategorizationtechniques(describedinSection “InformationProcessing”)Thegeneralarchitectureisalsosignificantlyextendedinordertomanage

(6)

theoutputofbothkindsoftechniquesandtoexploititintheactualsearch(Section“Searchand Retrieval”).Thearchitecturehighlightstwomainactivities:

1. Information processing (dashed line in figure):Thisactivityisappliedtoboththedocuments toberetrieved(e.g.webpagesforagivensite)andthedocumentsusefultodeterminethe user’sprofile(suchase-mails,webpagesviewed,profileinformation,pastsearchqueries,etc.). Theavailableinformationisextracted(crawled)andindexedbymeansofad-hoctechniques, alsoexploitingexternalknowledgesources.Thedatastructurescontainingalltheinformation processingresultswillbereferredtoas“semanticglossaries”,oneforthewebsite(s)andone fortheuserprofile.Bothglossariesarethencompared(“Semanticglossariescomputationand comparison”inthefigure)withad-hocdocumentsimilarityalgorithms,detailedinSection “SearchandRetrieval”.Thegenerated“Profilerankings”symbolizehowrelevanttheretrievable documentsareinrelationtotheuser’sprofile;

2. Search and retrieval (solid line in figure):Thisactivityprovidesusefulanswerstotheuserby retrievingthemostrelevantdocumentswithregardtotheuserquery,alsotakingintoaccountits pre-computedprofileinformation.Thisisachieved(“Semanticqueryprocessing”)bydetermining the“Queryrankings”ofthequerywithregardtotheavailabledocuments(semanticsimilarity discussedinSection“SearchandRetrieval”).Queryrankingsarefinallyfusedwithprofile rankingsinordertoprovidethefinalranking. Thesetwomainactivitiesaredetailedinthenexttwosections. INFoRMATIoN PRoCESSING Informationprocessingiscommontoboththedocumentsofthewebsite(s)tobeindexedandthe userprofiledocuments.Eachofthemwillcontributetoasemanticglossarystructurecontainingthe resultsoftheanalysis. Thefirststepoftheinformationprocessingiscrawling:web/filecrawlingisperformedinorder toretrievetherawdata(e.g.,Title,Content,URL,FileName,MetaDescription,MetaKeywords)of

(7)

thedocumentsthatwillbefurtheranalysedbythesubsequentsteps.Allofthecrawlingoperations arehandledbytheopen-sourceenterprise-classsearchenginesoftware,OpenSearchServer11_. Semantic Text Analysis

Assingleexistingtoolsdonotallowsufficientconfigurationandextensionoptions,theauthors designedacustomizedSemantictextanalysismodule,whichexploitsseveralopensourcelibraries toperformspecificactions.Theanalyserallowsustoextractthecontents(andmeanings)ofthe processedinformation,bymeansof: • Textextraction,anoperationthatcanvarygreatlydependingontheformatofeachdocument; thisistakencareofbycomponentsoftheopen-sourcesoftwareGATE12_; • Tokenization,thetermsareidentifiedandpunctuationisremoved; • StemmingandPartofSpeech(POS)Tagging,thetokensare“normalized”and“stemmed”,i.e., theyarereducedtotheirbaseform(managingpluralsandinflections)and“tagged”withPOS tags(i.e.,nouns,verbs,etc.);thisistakencareofbytheTreeTagger13_library; • Compositetermidentification,possiblecompositeterms(suchas“productionarea”or“wine tasting”)areidentifiedbymeansofasimplestatemachineandofPOStagsinformation; • WordSenseDisambiguation(WSD),themeaningofthekeywordsresultingfromtheprevious stepsismadeexplicitwithregardtoareferencethesaurus(typically,WordNet14_{,(Miller,1995));} inparticular,therelevantsynsetsareextractedandassociated; • Weightcomputation,keywordsarefinallyenrichedwithweightinformation(seebelow). Theextractedinformationenablestheretrievalofthemostrelevantinformationfortheuserin thesearchandretrievalphase.Inparticular,theproposedapproachexploitesthesemanticinformation fromthethesaurusandthesimilarityfunctionsdescribedinSection“SearchandRetrieval”.Thanksto them,AMBIT-SEwillbeabletoretrievedocumentsonthebasisofsynonymsandrelatedkeywords; forinstance,documentsabout“dogs”areconsideredasrelevanttoaqueryabouta“terrier”.Moreover, bymeansofWSD,documentsabouta“pet”inthesenseof“domesticatedanimal”willnotbemixed upwiththosewhere“pet”means“aspeciallovedone”. Fortheweightinformation,textanalysisexploitsavariantofthestandard“tf-idf”weighting scheme.Thisisintroducedtoconveytheinformationoftheclassicschemewhilealsokeepingthe weightsnormalizedintherangeof[0,1]:thisenablesaneffectiverankingcomparisonandfusion (moreonthisinSection“SearchandRetrieval”).GivenadocumentDx_{,eachkeywordk} _D i x _∈ x_is assignedakeyword weight kwi x_definedas: kw_ix kf idf i x i = ⋅ (1) where: kf f f i x i x l l x = max (2) idf log N n max log N n i i l l =                      /    (3)

(8)

kf andidf arenormalizedkeywordfrequencyandinversedocumentfrequency: fi

x_istheraw frequencyofkeywordkiindocument D

x_{,Nisthetotalnumberofindexeddocumentsandn} ithe number of documents where keyword ki appears. Note that, by definition, 0≤kfi ≤1

x and 0≤idfi ≤1. Semantic Categorization Besidessemantictextanalysis,asemanticcategorizationprocessisalsoappliedtothedocumentsin ordertotageachdocumentwithappropriatesubjectclasses.TheMediaTopicNewsCodestaxonomies andvocabulariesprovidedbyIPTCareadopted.Thesearewell-knowntaxonomiesofferingavery goodlevelofdetailandcoverageofawiderangeoftopics.Inthiscase,thestartingpointistheoutput ofad-hoccategorizationtoolsfromExpertSystem15_{,whereeachclasstagisgivenascore s c} i

( )

, 0≤s c

( )

_i ≤1;thehighertheweightthemorerelevanttheclassisforthedocument.Forinstance, adocumentaboutthetypical“Terrierfoodproducts”willpresumablyhave“Petproductandservice” amongitshighestscoringassociatedtags.However,theauthorsdeemthatifmostofthedocuments ofthecollectionaretaggedwiththesameclass,thisclasswillnotbeparticularlydistinctiveand usefulintheretrievalphase.Therefore,thisfactiscapturedbygoingbeyondtheplainscoreandby defininganewweightingschemefortheclassesinspiredbythetextanalysisscheme(Equation1): givenadocumentDx_{,eachclasstagc} _D i x _∈ x_{isassignedaclassweightcw} i x_definedas: cw_ix s c icf i x i =

( )

⋅ (4) where: icf log N t max N t i i l l =                        log _ (5)

icf _{standsforinverse (document) class frequency;Nisthetotalnumberofdocumentsandt}_i isthenumberofdocumentstaggedwithclassci.

Website(s) and User Semantic Glossaries

Byapplyingbatchinformationprocessingtothedocumentcollection,thewebsite(s) semantic glossary isautomaticallygenerated.Conceptually,itconsistsofaglobal view(allkeywords/classestogether withtheiroccurrencesandadditionalextracteddata),andaper-document view(keywords/classes occurrencesineachdocumentwiththeirstatistics).Asimplifiedsampleofthelatterviewisshown inTable1-leftandTable1-right).Inparticular,“Document”isthelistofthedocumentsIDsinwhich eachkeyword/classoccurs,“Synsets(s)”arethesynsetsselectedastheoutputofWSD,while“KF”, “KWeight”and“CWeight”aretheweightsillustratedinEquations(1)and(2). EachtimeauserlogsinAMBIT-SE,abatchanalysisisalsoscheduledonthedataofherprofile Uinordertogenerate/updateitsuser semantic glossary(whichsharesthesamestructureofthe websitesemanticglossarydiscussedabove).Inparticular,theprofilecontainsaction history data includingalistofpastaccesseddocumentsandpastsearchesperformedonthewebsite.Theideais thatsuchdatacanbeanalysedinthesamewayasthewebsitedocuments,thereforeexploitingallthe powerofthedocumenttextanalysisandcategorizationinordertoassociatemeaningfulkeywords

(9)

Duetotheircomplexityandsotheneedforcomputationalpower,alltheanalysesareperformed whennousersareloggedin;theywillbeavailableinordertoprocessfuturerequestsfromthesame userinamoreaccurateway.

SEARCH ANd RETRIEVAL

WhenauserUsubmitsaqueryQ,AMBIT-SEhastoansweritaseffectivelyasitcan.Thetechnique illustratedinthissectionaimstoproducethemosteffectiveranking τ oftheindexeddocumentsD ∈DwithregardtobothUandQ.Thisisdonebytakingintoaccountalltheinformationavailable inthesemanticglossariesproducedbytheanalysisprocess(textanalysisandclassification)andby meansofthefollowingad-hocsimilaritymetricsbetweentwogenericdocuments(i.e.keywordsets) Dx_andDy_: • ThesimilarityTextSim D D

(

x_, y

)

_{,consideringkeywordinformation;} • ThesimilarityClassSim D D

(

x_, y

)

_{,consideringclassinformation.} Inparticular,TextSimandClassSimwillbeappliedtoboth: • TheuserprofileU(i.e.,)withregardtoeachindexeddocument(asperformedonpastnavigated data):thisproducestheprofilerankings τtext U _{and τ} class U _; • ThequeryQ(i.e.,)withregardtoeachdocument;thisproducesthequeryrankings τtext Q _and τ_classQ _. Inthefinalpartofthissection,theauthorswillshowhowthisrankinginformationisexploited inranking selection/fusioninordertoproducethefinalranking τ .Next,theTextSimandClassSim similaritymetricswillbedefined.

Text and Class Similarity Metrics

Buildingonpreviousresearchontextretrievalforspecificsubjectareasassoftwareengineering (Bergamaschietal.,2015;Martoglia,2011),agricultural(Beneventanoetal.,2016)anduser-centric culturalenhancementdata(Martoglia,2015),thetextsimilarityTextSim D D

(

x_, y

)

_{formulabetween} documentsDx_andDy_isdefinedas:

TextSim D D max KSim k k kw kw x y k D k D i x j y i x j y ix x jy y , ,

(

)

=

∑

∈ ∈

(

)

⋅ ⋅ D Dx (6)

Table 1. Sample portions of the extracted Document Semantic Glossary: per-doc view for keywords (left) and classes (right)

Document Keyword Synset(s) KF KWeight Document Class CWeight

P02001 Terrier 00919240-n 0.445 0.277 P02001 Petproductandservice 0.645

P02005 Veal 00414222-n 0.210 0.131 P02005 Foodindustry 0.442

(10)

wheretheweightingschemeillustratedinEquation(1)isexploitedandKSimisatermsimilarity formula(seelater)takingintoaccountthesemanticinformationextractedfromthesemanticglossary. Simplyput,thesimilarityTextSim D D

(

x_, y

)

_{betweentwodocumentsD}x_andDy_{isdeterminedby} summingthemaximumkeywordsimilarityscoresKSim k k

(

i, j

)

betweeneachpairofkeywordski andkjbelongingtodifferentdocuments,multipliedbytheweightsofboth.AstoKSim k k

(

i, j

)

,in AMBIT-SEsemanticframework,kiandkjcanbe:

1. Equal or synonyms:KSim ,

( )

=1;

2. Related,i.e.thethesaurushypernymypathsimilaritybetweenthekeywords’synsetsexceedsa giventhresholdTh(Bergamaschietal.,2015):KSim ,

( )

=r,whererisanarbitraryvaluebetween 0and1,default0.7;

3. Unrelatedotherwise:KSim ,

( )

=0.

Pleasenotethat,sinceboth0≤KSim ,

( )

≤1and0≤kw≤1,then0≤TextSim ,

( )

≤1.This willbeusefulintherankingselection/fusionphase(Section“RankingSelection/Fusion”).

Inadditiontodocumentkeywords,theclassesassociatedbythesemanticclassifiercanalso significantlyhelpinretrievingusefuldocuments.Thisisobviouslytrueifbothdocumentsarestrongly characterizedbyacommonIPTCclass(e.g.“processindustry”);however,alsodocumentsabout colafactoriestaggedwithasimilarclass“foodindustry”wouldbeofinterest.Thisisachievedthrough ClassSim D D

(

x_, y

)

_{,whichquantifiesthesimilarityof D}x_{and D}y_{onthebasisoftheirassociated} IPTCclassesc_ix _∈Dx_andc _D j y _∈ y_: ClassSim D D max CSim c c cw cw x y c D c D i x j y i x j ix x yj y , ,

(

)

=

∑

∈ ∈

(

)

⋅ ⋅ yy x x c ∈D

{

}

(7)

AsforEquation(6),CSim c c

( )

i, j betweenclassesciandcjrangesbetween1(equalclasses),

r(similar,i.e.“near”ontheIPTCtaxonomy,classes)and0(otherwise).Again,since0≤CSim ,

( )

≤1, then 0≤ClassSim ,

( )

≤1. Ranking Selection/Fusion Aspreviouslyseen,givenaprofileU,Equations(3)and(4)inducerankings τtext U _{and τ} class U _oneach documentD∈DwithregardtoU,respectively.Similarly,whenthetwoequationsarecomputedwith regardtoaqueryQ,rankings τtext Q _{and τ} class Q _{areinducedonthesamedocuments.Theaimofthe} ranking selection/fusionprocesspresentedinthispaperistoexploittheinformationofthoserankings inordertoobtainthefinalranking τ .Theaimsarethefollowing: 1. Thefinalrankingτ shouldreflectbothQandU,anditshouldbepossibletodefinetherelevant importanceofthequeryQwithregardtotheprofileU(forinstance,toprivilegetheinformation containedinQ); 2. Thefinalranking τ shouldbeflexiblydefinedsotoreflectbothtextandclassinformation

(11)

shouldbeabletoautomaticallydecideifoneofthetwokindsofrankings(text,class)isnot significant,forinstanceduetoverylowsimilarities,whichwouldonlynegativelyaffectthefinal ranking,thereforeexcludingitfromthefusion; 3. ThescoreofeachdocumentD,andnotonlyitspositionintherankings,shouldbetakeninto accountinordertodirectlyreflectitsrelevanceinthefinalranking. LetusseehowAMBIT-SEachievesthesepoints.Bymeansofalinearcombinationscorefusion method,thescoresinducingthefusedrankings τtextand τclass,whichtakeintoaccountbothUand

Q,aredefinedas:

s text D s textQ D sUtext D

Q Q

τ

( )

₌_α _⋅ τ

( )

_{+ −}

(

₁ _α

)

_⋅ τ

( )

₍₈₎

s class D sclassQ D s classU D

Q Q τ

( )

₌_α _⋅ τ

( )

_{+ −}

(

₁ _α

)

_⋅ τ

( )

where 0≤α_Q ≤1isapreferenceweightdeterminingtherelevantimportanceofQwithregardto U.Thedefaultvalueis αQ=0.7,meaningthattheinformationinQistypicallymoresignificant thanthatinU;thiscanbevaried,forinstance,bythesystemadministratordependingontheactual usagescenarios.

Theabovedefinedτtextandτclassrankingsareeventuallyfusedinafinalrankingτ ,whichtakes intoaccountbothtextandclasscontributionsasofEquation(8):

s Dˆτ

( )

=β_text⋅sτtext

( )

D + −

(

1 β_text

)

⋅sτclass

( )

D (9)

where 0≤β_text ≤1isapreferenceweightdeterminingtherelativeimportanceof τtextwithregard to τclass.Thedefaultvalueis0.5.Notethatrankingselectionisimplementedinthefollowingway: if

D D

s text D s class D

∑

τ

( )

∑

τ

( )

_thenβ

textisautomaticallysetto1inordertoexcludethecontribution of τclass(andvice-versa),thereforeavoidingpossiblydetrimentalnoise.Onlydocumentsthatare partofallrankingswillappearinthefinalranking. EXPERIMENTAL EVALUATIoN Thissectionpresentstheresultsofseveraltestsperformedondifferentkindsofwebsites.Thispaperis focusedoneffectivenessevaluation;forreadersinterestedintimeperformances,thecurrentprototype hasaresponsetimeof40msonaverageonastandardsingle-nodeconfiguration. Fivewebsiteswereselectedforevaluationpurposes;foreachoneofthem,appropriateinformation needswereestablishedbyexaminingcommonsearchesperformedinthepast. Thechoiceofthewebsitewasdrivenbytheaimsoftheproposedapproach.First,thesitesare allrepresentativeofrealandtypicalbusiness-relevantsites.Moreover,thesiteswereselectedsoasto coveranumberofdifferenttopics(e.g.health,cars,recycling,etc.)andthusdifferentterminologies andtextcontent.Finally,theyhavedifferentfeaturesintermsofstructure.Thisisinlinewiththemain goalofthissection,i.e.toevaluatetheeffectivenessandtheflexibilityofanenterprisesearchengine. Astopossibleexperimentalcomparisons,asanticipatedintherelated,therearecurrentlyno approachesthatcanbedirectlycomparedwiththeproposedone.Inparticular,amongtheapproaches

(12)

whoseimplementationwaspubliclyavailable,thoseofferingsemanticfeaturesrequireunavailable manualannotationsand/orwerestrictlydesignedtoworkonspecificontologies(Haslhoferetal.,2013; Thesprasith&Jaruskulchai,2014;DeVochtetal.,2017).Thisisalsotrueforsemanticenterprise searchengines(Cogito,Attivio,ContentAnalytics).Theconsidereduser-awaresearchenginesdonot exploitthesemantics(Bolchinietal.,2011;Cabrietal.,2003;Falcarinetal.,2013;Xiangetal.,2010; Vuetal.,2017).Also,thepubliclyavailableapproachesworkingoncontextinformationarerestricted toaspectssuchastimeandlocation(Falcarinetal.,2013;Xiangetal.,2010;Villegas&Müller2010) orveryspecificscenarios(e.g.agentmanagementforCoveo,imagesearchforOntogator,cQAfor (Figueroa&Neumann,2016))whichmakesthemnotapplicabletotheconsideredcase.Summarizing, thereasonthatpreventsdirectcomparisonsare:first,onlyafewnumberofapproachesexploitboth semanticsanduserawareness;second,mostoftheapproachesfocusonaspecificapplicationfieldand cannotbeappliedtogeneralwebsites;third,severalapproachesarenotfreelyavailabletobetested. Anyway,forreference,thefinalpartofthissectionisdevotedtoadirectcomparisonwithGoogle searchengine;evenifitdoesnotexhibitallthefeaturesofAMBIT-SE,theauthorsconsidereditan importantbenchmarkinthefield. Table2reports,foreachwebsite,theaddress,abriefdescriptionofitscontentsandanexample ofaconsideredinformationneed.Inthecontextoftheinformationneedsofeachwebsite,50plausible queriesweresubmittedtothesystem.Moreover,twooptionsforuserprofilesareconsidered:asimpler setting,wheretheprofilecontainsonlydocumentsrelevanttotheinformationneed(“homogeneous profile”),andamorecomplexone,wheretheprofilecontainsalsoanequalnumberofirrelevant documents(“heterogeneousprofile”). Ineachsituation,theoutputofAMBIT-SEiscomparedwitha“goldstandard”,i.e.relevant answersmanuallyselectedfromthewebsites,andprecisionandrecallareassessed;inparticular, thetestscomputeinterpolatedprecisionat11recalllevelsof0.0,0.1,0.2,...,1.0(0,10,20,...,100 percent)(Baeza-Yates&Ribeiro-Neto,1999),thustakingintoaccountnotonlytherelevanceofthe resultsbutalsotheirpositioninthereturnedranking.Moreover,a“stable”situationisassumed, whereusersanddocumentshavebeenalreadyautomaticallyprocessedandtheirrelevantkeywords andclassesstoredintheSemanticGlossaries.Alltheparametersaresetatdefault. Figures3through5depicttheresults(averagedonallthequeries)obtainedbyAMBIT-SE, comparedwiththebaselineofasyntacticretrievalmethodignoringsynonyms,relatedtermsand classinformation.Inparticular,thisbaselineisrepresentativeofthedocumentretrievaltechniques commonlyexploitedbymostcommercialsystemsandstandardenterprisesearchengines(e.g. Alfresco,Autonomy,Solr).Furthercomparisonswithavailablesystems(e.g.Google)willbe discussedinthefinalpartoftheexperimentalanalysis.Inordertobettervisualizethedifference betweenAMBIT-SE’sresultsandthebaseline,thefiguresdetailthetheprecisionachievedbythe differentfeaturesofAMBIT-SE.Inparticular,thewhitebars(atthebottom)representtheprecision ofthesyntacticmethod(baseline);thecontributionsontoparetheimprovementsachievedbymeans ofthesemantic(black)anduser-awarefeatures(gray). Asreaderscansee,theimprovementsinprecisionofferedbythesemanticanduser-awarefeatures ofAMBIT-SEaresignificantineachscenario,rangingfrom20%toeven80%.Letusstartbyanalysing thehttp://www.cobat.it/results(Figure3(left)).Here,thesemanticfeaturesofferanadvantageat allrecalllevels;theresultsalsobenefitedfromtheuser-awarefeatures,becauseofthelargenumber ofkeywordscontainedwithinthedocumentsassociatedtotheuserprofiles.Forinstance,theuseof semanticsenablesthematchingofdifferentbutveryrelatedtermslike“battery”and“accumulator”. Thosecontributionsgounnoticedinstandardsyntacticsearch. Consideringthesecondwebsite(Figure3(right)),theuseofsemanticsprovidesevengreater benefits:forinstance,theuseofsynonymsandrelatedtermsareabletoexploitanumberofterms correlationssuchasbetween“money”and“bookkeeping.”Thiswasespeciallyevidentinthelonger andlessdirectqueriessubmitted,whicharehardertosatisfybyasearchenginewithoutadditional

(13)

informationintheformofsynonymsandrelatedterms.Summingupthecontributionoftheprofile analysis,AMBIT-SEreachesinsomecasesanimprovementofnearly80%inprecision. InthetestofFigure4(left),theauthorsfoundthatinsomecasesthequeriesdidnotdirectly benefitfromsynonymsandrelatedtermsmanagement,maybeduetoveryspecificterminologythatis usedinthepages.Anyway,theuseofsemanticsandtheprofileanalysisproviderelevantadvantages eveninthisscenario.MovingtoFigure4(right),theuseofwordstems,synonyms,relatedterms andprofilecanturnevendifficultqueriesintomanageableones.Forinstance,insomecases,the syntacticbaselinecouldnotretrieveanyrecordssincethequeriesdon’tincludetermsthatmatch exactlytheonesfoundintherelevantdocuments:amongtheexploitedtermscorrelations,thevery frequentonebetween“pet”and“animal”. Thefinalwebsiteconsidered(Figure5)ischaracterizedbypagescontainingverylittletext,thus providingadifferenttaskwithregardtotheothers.Alsointhiscase,alotofcomplexqueriesdid notyieldsatisfactoryresultsforthesyntacticbaseline,thusprovidingveryconsistentadvantages fromthesemanticfeatures(precisionimprovementof30%orgreater).Duetothepeculiarnatureof thewebsite,somequerieswereapparentlytoodifficultevenwiththeadditionalinputprovidedby thesemanticsandprofiles.Anyway,onaverage,theeffectoftheadvancedAMBIT-SEfeaturesis evidenteveninthissituation. Afurthertestispresentedwhichisaimedatevaluatingtheimpactoftheclassificationfeatures ofthepresentedengine,includingthenovelrankingselectionandfusioncapabilities.Figure6(left) showstheprecisionfigures(averagedoverallthequeriesandallrecalllevels)thatareachievedby

Table 2. Experimental setting: details of the five business-relevant websites selected for evaluation purposes

Website Description Examples of Information Needs

http://www.cobat.it/ Arelativelysmallwebsitethatprovidesinformationand servicesfordisposingandrecyclingfourproblematic wastecategories:batteriesandaccumulators,tires, electricandelectronicdevices,andphotovoltaicpanels. Retrievedocumentspertainingtothedisposalof batteriesandaccumulators. http://evergreensmallbusiness.com/ ABlogthatpublishesdifferentkindsofinformationand adviceforsmallbusinesses,allclassifiedincategories suchasbusinesstaxes,management,personalfinance, etc. Retrievearticlespertainingtobookkeeping. http://www.bagnolottanta.it/ Ajournalisticwebsitewithseveralthematiccolumns

suchashistory,culture,theatre,etc. Retrievearticlesinspecificcolumnssuchasbooks. http://truegoods.com/ AnIndieonlineshopthatspecializesonhealthyand

naturalproducts. Retrieveinformationonproductsbelongingtothepet-carecategory. http://www.gruppozatti.it/ Anauthorizedcardealerwhichsellsseveralbrandsof

bothnewandusedcars. Retrievedifferentinformationaboutcarsbelongingtotheusedcategory.

(14)

Figure 4. Test results for http://www.bagnolottanta.it/ (left) and http://truegoods.com/ (right)

Figure 5. Test results for http://www.gruppozatti.it/

(15)

exploitingonlythetextrankingandsimilarities(leftbars)withregardtotheonesachievedbythe completesetuppresentedinthispaper.Thebenefitsareusuallyintheorderofmorethan10%of improvement.Lookingatspecificcases,thefusedrankingwasabletotakethebestfromtheclass andtextrankings,togethercapturingtheuserinterestsmorecompletely. Inordertocompletetheevaluation,theauthorsperformedaquantitativecomparisonwitha state-of-the-artsearchengineandadiscussionaboutresultsisreportedinthefollowing.Thechosen searchenginewasGoogle,asmentionedbefore. Figure6(right)showstheresultsofthesystemcomparedwiththoseobtainablethroughtheGoogle searchengine.Indeed,Googleisoneofthemostwidelyusedsearchenginesalsointheenterprise searcharea.SimilarlytoAMBIT-SE,Googleiscompletelyautomaticandrequiresnomanualwork (e.g.annotation)onthedocumentstobeprocessed.Inthiscase,theconsidereddocumentsethas beenrestrictedtothespecificconsideredwebsitepages.DifferentlyfromAMBIT-SE,Googledoes notperformsemanticanalysisofthetexts,howeveritemploystextprocessingtechniquessuchas stemming.MultipleusersweresimulatedinGooglethroughdifferentsearchsessions.Theresults showthattheaverageprecisionachievedbyGoogle(intherangeof10-30%)isquitelowerthanthe oneachievedbyAMBIT-SE(alwaysabove55%).Inparticular,Googleoffersalevelofperformance thatisonlymarginallybetterthanthesyntacticbaselinediscussedintheprevioustests. CoNCLUSIoN Searchenginesrepresentameansessentialforalotofactivities.Thisisespeciallytrueinanenterprise context,wherethesuccessoftheenterpriseisoftenstrictlydependentontheabilityofitsemployees andcustomerstofindtheneededinformation. InthiscontexttheauthorshaveproposedAMBIT-SE,anenterprisesearchengineapproachthat reliesontwoaspects,user-awarenessandsemantics,jointlyexploited.First,itbuildsaprofileof theuser,whichisexploitedtosearchfortheinformationthatbestmatcheswithherneeds.Second, semantictechniquesenabletheretrievalofinterestinginformationthatcouldnotbeconsidered exploitingastandardsyntacticsearch.Theexperimentalevaluationhasshownthatthepresented approachperformsbetterthantraditionalsearchmethods. Inthefuture,furtherwaysofexploitingsemanticsanduserinformationinthesearchwillbe considered,forinstancebyadaptingsomeoftheideascomingfrompastworksindifferentcontexts (e.g.,semanticsearchinheterogeneousanddynamicgraphdata(Cataniaetal.,2013)andmultimedia data(Granaetal.,2013)). ACKNoWLEdGMENT Thisworkwassupportedbytheproject“AlgorithmsandModelsforBuildingcontext-dependent InformationdeliveryTools”(AMBIT)co-fundedbyFondazioneCassadiRisparmiodiModena(SIME 2013.0660),bythe“SocialGQ”FAR2016FIMDepartmentProject,UniversityofModenaandReggio Emilia,andbytheEUH2020FIRSTproject(GAnumber:734599—H2020-MSCA-RISE-2016).

(16)

REFERENCES

Abdou,S.,&Savoy,J.(2008).Searchinginmedline:Queryexpansionandmanualindexingevaluation.

Information Processing & Management,44(2),781–789.doi:10.1016/j.ipm.2007.03.013

Baeza-Yates,R.A.,&Ribeiro-Neto,B.(1999).Modern Information Retrieval.Boston,MA,USA:Addison-WesleyLongmanPublishingCo.,Inc.

Beneventano,D.,Bergamaschi,S.,&Martoglia,R.(2016).Exploitingsemanticsforsearchingagricultural bibliographicdata.Journal of Information Science,42(6),748–762.doi:10.1177/0165551515606579 Bergamaschi,S.,Martoglia,R.,&Sorrentino,S.(2015).Exploitingsemanticsforfilteringandsearching knowledgeinasoftwaredevelopmentcontext.Knowledge and Information Systems,45(2),295–318.doi:10.1007/ s10115-014-0796-1

Bolchini,C.,Orsi,G.,Quintarelli,E.,Schreiber,F.A.,&Tanca,L.(2011).Contextmodelingandcontext awareness:Stepsforwardinthecontext-addictproject.A Quarterly Bulletin of the Computer Society of the IEEE

Technical Committee on Data Engineering,34,47–54.

Cabri,G.,Gaddi,S.,&Martoglia,R.(2016).AMBIT-SE:TowardsaUser-awareSemanticEnterpriseSearch Engine.InProceedings of the 12th International Conference on Web Information Systems and Technologies

(WEBIST 2016)(Vol.2,pp.98–108).Springer.doi:10.5220/0005788800980108

Cabri,G.,Leonardi,L.,Mamei,M.,&Zambonelli,F.(2003).Location-dependentServicesforMobileUsers.

IEEE Transactions on Systems, Man, and Cybernetics. Part A, Systems and Humans,33(6),667–681.doi:10.1109/

TSMCA.2003.819496

Carpineto,C.andRomano,G.(2012).Asurveyofautomaticqueryexpansionininformationretrieval.ACM

Comput. Surv.,44(1):1:1–1:50.

Catania,B.,Guerrini,G.,Belussi,A.,Mandreoli,F.,Martoglia,R.,&Penzo,W.(2013).WearableQueries: AdaptingCommonRetrievalNeedstoDataandUsers(VisionPaper).InProceedings of the 7th International

Workshop on Ranking in Databases (DBRank).doi:10.1145/2524828.2524835

DeVocht,L.,Softic,S.,Verborgh,R.,Mannens,E.,&Ebner,M.(2017).SocialSemanticSearch:ACaseStudy onWeb2.0forScience.International Journal on Semantic Web and Information Systems,13(4),155–180. doi:10.4018/IJSWIS.2017100108

Falcarin,P.,Valla,M.,Yu,J.,Licciardi,C.A.,Frà,C.,&Lamorte,L.(2013).Contextdatamanagement:An architecturalframeworkforcontext-awareservices.Service Oriented Computing and Applications,7(2),151–168. doi:10.1007/s11761-012-0115-1

Figueroa,A.,&Neumann,G.(2016).Context-awaresemanticclassificationofsearchqueriesforbrowsing communityquestion–answeringarchives.Knowledge-Based Systems,96,1–13.doi:10.1016/j.knosys.2016.01.008 Grana,C.,Serra,G.,Manfredi,M.,Cucchiara,R.,Martoglia,R.,&Mandreoli,F.(2013).UNIMOREat ImageCLEF2013:ScalableConceptImageAnnotation.InProceedings of the Image Retrieval in Conference

and Labs of the Evaluation Forum (ImageClef).

Haslhofer,B.,Martins,F.,&Magalhães,J.a.(2013).Usingskosvocabulariesforimprovingwebsearch.In

Proceedings of the 22nd International Conference on World Wide Web(pp.1253-1258).

Heflin,J.,&Hendler,J.(2000).Searchingthewebwithshoe.InArtificial Intelligence for Web Search.Papers

from the AAAI Workshop.

Hyvonen,E.,Saarela,S.,&Viljanen,K.(2003).Ontogator:combiningview-andontology-basedsearchwith semanticbrowsing.InProceedings of XML Finland.

Liu,F.,Yu,C.,&Meng,W.(2004).Personalizedwebsearchforimprovingretrievaleffectiveness.IEEE

Transactions on Knowledge and Data Engineering,16(1),28–40.doi:10.1109/TKDE.2004.1264820

Mangold,C.(2007).Asurveyandclassificationofsemanticsearchapproaches.InSemanticsandOntology. doi:10.1504/IJMSO.2007.015073

(17)

Martoglia,R.(2011).FacilitateIT-ProvidingSMEsinSoftwareDevelopment:aSemanticHelperforFiltering andSearchingKnowledge.InProceedings of the 23rd International Conference on Software Engineering and

Knowledge Engineering(pp.130-136).

Martoglia,R.(2015).Ambit:Semanticenginefoundationsforknowledgemanagementincontext-dependent applications.InProceedings of the 27th International Conference on Software Engineering and Knowledge

Engineering(pp.146–151).doi:10.18293/SEKE2015-027

Miller,G.A.(1995).WordNet:ALexicalDatabaseforEnglish.Communications of the ACM,38(11),39–41. doi:10.1145/219717.219748

Ramona-Cristina,P.,Vasilateanu,A.,&Goga,N.(2016).Ontologybasedmulti-systemforSMEknowledge workers.InProceedings of the 2016 IEEE International Symposium on Systems Engineering (ISSE),Edinburgh, UK,October3-5.doi:10.1109/SysEng.2016.7753132

Rocha,C.,Schwabe,D.,&deAragao,M.P.(2004).Ahybridapproachforsearchinginthesemanticweb.InWWW

’04: Proceedings of the Thirteenth International Conference on World Wide Web.doi:10.1145/988672.988723

Savoy,J.(2005).Bibliographicdatabaseaccessusingfree-textandcontrolledvocabulary:Anevaluation.

Information Processing & Management,41(4),873–890.doi:10.1016/j.ipm.2004.01.004

Shekarpour,S.,Marx,E.,NgongaNgomo,A.-C.,&Aue,S.(2015).SINA:Semanticinterpretationofuser queriesforquestionansweringoninterlinkeddata.Journal of Web Semantics,30,39–51.doi:10.1016/j. websem.2014.06.002

Thesprasith,O.,&Jaruskulchai,C.(2014).Queryexpansionusingmedicalsubjectheadingstermsinthe biomedicaldocuments.InIntelligent Information and Database Systems - 6th Asian Conference, ACIIDS 2014, Bangkok,Thailand,April7-9(pp.93-102).

Villegas,N.M.,&Müller,H.A.(2010).Managingdynamiccontexttooptimizesmartinteractionsandservices. InM.Chignell,J.Cordy,J.Ngetal.(Eds.),TheSmartInternet,LNCS(Vol.6400,pp.289-318).SpringerBerlin Heidelberg.doi:10.1007/978-3-642-16599-3_18

Voorhees,E.M.(1994).Queryexpansionusinglexical-semanticrelations.InProceedings of the 17th Annual

International ACM-SIGIR Conference on Research and Development in Information Retrieval. Dublin, Ireland, 3-6 July (pp.61-69).

Vu,T.,Willis,A.,Kruschwitz,U.,&Song,D.(2017).PersonalisedQuerySuggestionforIntranetSearchwith TemporalUserProfiling.InProceedings of the 2017 Conference on Conference Human Information Interaction

and Retrieval(pp.265-268).doi:10.1145/3020165.3022129

Xiang,B.,Jiang,D.,Pei,J.,Sun,X.,Chen,E.,&Li,H.(2010).Context-awarerankinginwebsearch.In

Proceedings of the 33rd International ACM SIGIR Conference on Research and Development in Information RetrievalSIGIR ’10(pp.451-458). ENdNoTES 1 _{AMBITstandsfor“AlgorithmsandModelsforBuildingcontext-dependentInformationdeliveryTools”} 2 _{IPTCstandsfor“InternationalPressTelecommunicationsCouncil”,http://www.iptc.org/site/Home/} 3 _{http://www.alfresco.com/} 4 _{http://www.autonomy.com/} 5 _{http://lucene.apache.org/solr/} 6 _{http://www.google.com/} 7 _{http://www.attivio.com/} 8 _{http://www.expertsystem.com/it/cogito/} 9 _{http://www.coveo.com/} 10 _{https://www.ibm.com/} 11 _{http://www.opensearchserver.com/} 12 _{https://gate.ac.uk/} 13 _{http://www.cis.uni-muenchen.de/.17ex~schmid/tools/TreeTagger/} 14 _{http://wordnet.princeton.edu/or,incaseofspecificcontexts,otherspecializedresources}

(18)

Giacomo Cabri is a full professor in Computer Engineering at the University of Modena and Reggio Emilia, Italy. He received the Master degree in Computer Engineering from the University of Bologna, and the PhD in Computer Engineering from the University of Modena and Reggio Emilia. His main research interests are: engineering of software agents, autonomic computing, distributed applications, mobile computing. He has published 12 chapters of scientific books, 34 international journal papers, 119 international conference/workshop papers. He received 7 best paper awards. He was the general chair of IEEE WETICE 2004, ACM PPPJ 2008, IEEE WETICE 2014. He is member of the editorial board of the International Journal of Computational Intelligence Theory and Practice (IJCITP), Serials Publications, of the journal Scalable Computing: Practice and Experience, and of the journal EAI Endorsed Transactions on Context-Aware Systems and Applications, ICST, Gent, Belgium. He was guest editor of 6 special issues.

Riccardo Martoglia is an assistant professor, habilitated as associate professor, at the University of Modena and Reggio Emilia. His research activity is in the field of Information Systems, Information Retrieval and Semantic Web, in particular the study of new methods for effectively and efficiently searching and managing non-conventional data. He published in more than 90 publications, including monographic books, book chapters, international refereed journal articles and conference papers (VLDB Journal, Information Systems, EDBT, etc.). He has participated in several national and European research projects on issues closely related to the representation, querying and access to data. He has been member of the organizing committee of several conferences and workshops, including GraphQ@EDBT, WETICE, and served on the program committees of a number of conferences, such as SEKE, ICEIS, MOBIWIS. He has also served on the editorial board of recognized international journals, including JUCS