18 September 2017

intestazione repositorydell’ateneo

MOMIS: Exploiting agents to support information integration / G. Cabri; F. Guerra; M. Vincini; S. Bergamaschi; L.Leonardi; F. Zambonelli. - In: INTERNATIONAL JOURNAL OF COOPERATIVE INFORMATION SYSTEMS. - ISSN0218-8430. - STAMPA. - 11:3-4(2002), pp. 293-313.

Original

MOMIS: Exploiting agents to support information integration

Publisher:

PublishedDOI:10.1142/S0218843002000601

Terms of use:openAccess

Publisher copyright

(Article begins on next page)

Testo definito dall’ateneo relativo alle clausole di concessione d’uso

Availability:This version is available at: 11380/305413 since: 2016-11-19T11:54:05Z

This is the peer reviewd version of the followng article:

Exploiting Agentsto Support Inf ormation Integration

G. Cabri�, F. Guerra

�, M. Vincini

�, S.Bergamaschi

���, L. Leonardi

�, andF. Zambonelli

�

e-mail:�cabri.giacomo,guerra.francesco,vincini.maurizio,

bergamaschi.sonia,leonardi.letizia,zambonelli.franco� @unimo.it

�Dipartimentodi Ingegneriadell’Informazione

Universit̀adi ModenaeReggioEmilia

Via Vignolese905,41100Modena�

CSITE-CNRBologna

V.le Risorgimento2, 40136Bologna�

Dipartimentodi ScienzeeMetodidell’Ingegneria

Universit̀adi ModenaeReggioEmilia

Via Allegri 13,42100ReggioEmilia

Abstract. Informationoverloadingintroducedby thelargeamountof datathatis spreadover theInter-

netmustbefacedin anappropriateway. Thedynamismandtheuncertaintyof theInternet,alongwith

theheterogeneityof thesourcesof informationarethetwo mainchallengesfor thetoday’s technologies

relatedto informationmanagement.In theareaof informationintegration,this paperproposesanap-

proachbasedonmobilesoftwareagentsintegratedin theMOMIS (MediatorenvirOnmentfor Multiple

InformationSources)infrastructure,whichenablessemi-automaticinformationintegrationto dealwith

theintegrationandqueryof multiple,heterogeneousinformationsources(relational,object,XML and

semi-structuredsources).The exploitationof mobileagentsin MOMIS cansignificantlyincreasethe

flexibility of thesystem.In fact,their characteristicsof autonomyandadaptabilitywell suit distributed

andopenenvironments,suchasthe Internet.The aim of this paperis to show the advantagesof the

introductionin theMOMIS infrastructureof intelligentandmobilesoftwareagentsfor theautonomous

managementandcoordinationof integrationandqueryprocessingoverheterogeneousdatasources.

1 Intr oduction

Informationin the Internetworld is spreadandhighly heterogeneous,andproviding an integratedaccess

to multiplesourcesis achallengingissueconcerningcooperationandinteroperabilityin globalinformation

systems.In thepast,companieshave equippedthemselveswith datastoringsystemsbuilding up informa-

tivesystemscontainingdatathatarerelatedoneanother, but whichareoftenredundant,heterogeneousand

not alwayssubstantial.The problemsthat have to be facedin this field aremainly dueto both structural

andapplicationheterogeneityof thesources,aswell asto thelackof acommonontology, causingsemantic

differencesbetweeninformationsources.Moreover, thesesemanticdifferencescancausedifferentkindsof

conflicts,rangingfrom simplecontradictionsin nameuse,to structuralconflicts.With respectto conven-

tional view integrationtechniques[4], in thelargescaletherearecomplicatingfactorsdueto thesemantic

heterogeneityof thesources.Therefore,to meettherequirementsof global,Internet-basedinformationsys-

tems,it is importantthatthetoolsdevelopedfor supportingtheseactivitiesaresemi-automaticandscalable

asmuchaspossible.

To facethe issuesrelatedto scalability in the large-scale,in this paperwe proposethe exploitation

of the MOMIS (MediatorenvirOnmentfor Multiple InformationSources)infrastructure,which focuses

on capturingandreasoningaboutsemanticaspectsof schemadescriptionsof heterogeneousinformation

sources,enhancedby theintegrationof mobileagentsfor supportingintegrationandqueryoptimization.

MOMIS [7,11,8,10] is an infrastructurefor information integration that dealswith the integration

andqueryof multiple, heterogeneousinformationsources,containingstructuredandsemistructureddata.

MOMIS is asupportsystemfor semiautomaticintegrationof heterogeneoussourcesschema(relational,ob-

ject,XML andsemistructuredsources);it carriesout integrationfollowing asemanticapproachwhichuses

Descriptionlogics-basedtechniques,clusteringtechniquesandan ODM-ODMG [21] extendedmodelto

representextractedandintegratedinformation,ODM � . UsingtheODL � language,referredto theODM �model,it is possibleto describethesources(localschema)andMOMIS supportsthedesignerin thegenera-

tion of anintegratedview of all thesources(GlobalVirtual View), whichis expressedusingXML standard.

Theuseof XML in thedefinitionof theGlobalVirtual View lets to useMOMIS infrastructurewith other

openintegrationinformationsystemsby theinterchangeof XML datafiles.

In a wideandopenenvironment,suchastheInternet,MOMIS suffersfrom somelimitationsrelatedto

theuncertaintyandtheunreliability intrinsic to large-scaleenvironments.Moreover, a bettermanagement

of the connectionsandthe exchangesof informationhelp MOMIS in saving network resources.The ex-

ploitationof mobileagentscanovercomesuchlimitationsandincreaseflexibility . In fact,they well suit the

requirementsof Internetapplicationsthanksto their characteristics.Theagency feature[31] permitsthem

to exhibit ahighdegreeof autonomywith regardto theusers:they try to carryout their tasksin aproactive

way, reactingto the changesof the environmentthey arehosted.The mobility feature[32] takesseveral

advantagesin awideandunreliableenvironmentsuchastheInternet.First,mobileagentscansignificantly

save bandwidth,by moving locally to the resourcesthey needandby carryingthecodeto managethem.

Moreover, mobile agentscandealwith non-continuousnetwork connectionand,asa consequence,they

intrinsically suit mobilecomputingsystems.All thesefeaturesareparticularlysuitablein the information

retrieval area[15].

In this paperwe show theadvantagesof the introductionin theMOMIS architectureof intelligentand

mobile software agentsfor the autonomousmanagementand coordinationof the integrationand query

processesoverheterogeneousdatasources.In particular:

– agentscanbeexploitedto managethesourcesin anautonomousandintelligentway;

2

– agents,by moving on thesourcesites,cansave bandwidthanddealwith unstableconnections;

– dependingon thequeriesandon thesourcesite locations,agentsallow to choosethebestoneamong

differentquerystrategies;

– mobileagentswell fit a scenariowhereusers,sourcesor bothcanbemobile.

Theoutlineof this paperis the following: Section2 introducesthearchitectureof thesystemandex-

plainstherolesof theagents.Section3 showshow thesourcesaremanagedandintegrated,by meansof an

examplein theareaof carcatalogues.Section4 presentstherelatedwork. Section5 concludesthepaper

andsketchessomefuturework.

2 The SystemAr chitecture

MOMIS providesa setof techniquesfor thedesignerto facecommonproblemsthatarisewhenintegrating

pre-existinginformationsources,containingbothsemistructuredandstructureddata.Likeotherintegration

projects[1,38], MOMIS followsa“semanticapproach”to informationintegrationbasedontheconceptual

schema,or metadata,of theinformationsources,andonthe � � architecture[30]. In thefollowingwepresent

theextensionof MOMIS by integratingsoftwareagents(bothfixedandmobile)to whichdelegatespecific

tasks(seeFigure1). In therestof thepaperwereferto MOMIS asthenew infrastructureincludingagents.

Theresultingarchitectureis composedby thefollowing components:

1. acommondatamodel, ODM , andthecorrespondingODL language,todescribesourceschemasfor

integrationpurposes.ODM andODL havebeendefinedin MOMIS assubsetof thecorresponding

onesin ODMG, following the proposalfor a standardmediatorlanguagedevelopedby the � � /POB

workinggroup[14]. In addition,ODL introducesnew constructorstosupportthesemanticintegration

process;

2. Wrapperagents, placedovereachsources,translatemetadatadescriptionsof thesourcesinto thecom-

monODL representation,translate(reformulate)a globalqueryexpressedin theOQL 1 querylan-

guageinto queriesexpressedin thesourceslanguagesandexportqueryresultdataset;

3. a Mediator, which is composedof two componentsin its turn: theGlobal SchemaBuilder (GSB)and

the QueryManager (QM). The Global SchemaBuilder componentprocessesandintegratesODL descriptionsreceived from wrapperagentsto derive the integratedrepresentationof the information

sources.The QM componentperformsqueryprocessingandoptimization.Startingfrom eachquery

posedby the useron the Global Schema,the QM generatesOQL queriesthat aresentto wrapper

agentsby exploitingqueryagents, whicharemobileagents.QM automaticallygeneratesthetranslation

of the queryinto a correspondingsetof sub-queriesfor the sourcesandsynthesizesa unified global

answerfor theuser.

3

WordNet

Service level

ODB-Tools

Global SchemaMETADATA REPOSITORY

Data level

Wrapperagent

RelationalSource

QueryManager

MOMIS mediator

UserApplication

creates

IntegrationDesigner

legenda

CORBA Object

User

GUI

Software tools

CORBA interaction

User interaction

ARTEMIS

SI-Designer

Wrapperagent

XMLSource

Wrapperagent

ObjectSource

Wrapperagent

genericSource

Query agents

Fig.1. TheMOMIS systemarchitecture

MOMIS provides the capability of explicitly introducingmany kinds of knowledgefor integration,

suchasintegrity constraints,intra- andinter-sourceintensionalandextensionalrelationships,anddesigner

supplieddomainknowledge.A CommonThesaurus, which hastherole of a sharedontologyof thesource

is built in a semi-automaticway. The CommonThesaurusis a setof intra and inter-schemaintensional

andextensionalrelationships,describingtheknowledgeaboutclassesandattributesof sourcesschemas;it

providesa referenceon which to basetheidentificationof classescandidateto integrationandsubsequent

derivationof theirglobalrepresentation.

MOMIS supportsinformationintegrationin the creationof an integratedview of all sources(Global

Virtual View) in a way automatedasmuchaspossibleandperformsrevision andvalidationof thevarious

kindsof knowledgeusedfor the integration.To this end,MOMIS combinesreasoningcapabilitiesof De-

scriptionLogicswith affinity-basedclusteringtechniques,by exploiting acommonontologyfor thesources

built by usinglexical knowledgefrom WordNet[29,34].

The Global Virtual View is expressedby usingXML standard,to guaranteethe interoperabilitywith

otheropenintegrationsystemprototypes.

1 OQL is asubsetof OQL-ODMG.

4

2.1 The Rolesof the Agents

Theagentsexploitedby our systemaredevelopedin Jade[5], a Java-basedagentplatformwhich enables

alsomobility.

In our architecture,agentshave two mainroles.On theonehand,thewrappersareagentsthatconvert

thesourceinformationandreactto sourcechanges.On theotherhand,theQM exploits mobileagentsto

carryout theretrieval of information.

Whena new sourcehasto beintegratedin MOMIS, a mobileagentmovesto thesitewherethesource

resides.Suchagentchecksthe sourcetype and autonomouslyinstalls the neededdriver to convert the

sourceinformation.Moreover, a fixedagent,calledwrapperagent, is left at this site to presidethesource

(representedby roundedboxesin Figure1). Besidesinteractingwith queryagentsasdescribedlater, the

wrapperagentsmonitorthechangesthatmayoccurin thedatastructureof sources;in fact,theintegration

of a sourceis performedonceat thebeginning,while queriesoccurlater, evenaftera long period,during

which the sourcemay be modified.Whena changeoccursin a source,the correspondingwrapperagent

createsanappropriatemobileagentthatmovesto theMOMIS siteto inform aboutthenew structure,soas

to updatetheGlobalSchema.In thiscase,theautonomyfeatureof agentshelpsin decidinghow to dealwith

theoccurredchanges.This featuremakeoursystemextremelyflexible. In fact,anagent-basedarchitecture

canbeexploitedto addflexibility if themediatorsystemhasto managesemistructuredsources.As stated

previously, semistructureddatarepresentirregular, unknown or often changingstructuresetof data.An

architecturethatis ableto perceivechangesin sourceschemasandthatgivesmechanismsto autonomously

evaluatethetotal impactof a changeover thewholeschema,needslessinteractionswith thedesigner.

Wrapperagent

RelationalSource

QueryManager

Wrapperagent

XMLSource

Wrapperagent

ObjectSource

Query agent

Fig.2. A trip query

5

Wrapperagent

RelationalSource

QueryManager

Wrapperagent

XMLSource

Wrapperagent

ObjectSource

Query agents

Fig.3. A parallelquery

TheQM worksasfollows.Whenauserqueriestheglobalschema,it generatesasetof sub-queriesto be

sentto thesinglesources.To thispurpose,it canexploit oneor morequeryagents, whichmoveto thesource

siteswherethey interactwith thewrapperagents(seethe“moving smiles” in Figure1). TheQM chooses

thenumberof queryagentsto useby a componentthatanalyzeseachquery[3]. In somecases,it is better

to delegatethesearchto a singlequeryagent,which performsa “trip” visiting eachsourcesite:it canstart

from thesourcethatis supposedto reducethefurthersearchesin themostsignificantway, thencontinuesto

visit sourcesites,performingquerieson thebasisof thealready-foundinformation(seeFigure2). In other

cases,sub-queriesarelikely to bequiteindependent,soit is betterto delegateseveralqueryagents,onefor

eachsourcesite:in thisway thesearchesareperformedconcurrentlywith ahighdegreeof parallelism(see

Figure3). In any case,thepeculiarfeaturesof mobileagentsareexploitedandmake theMOMIS searches

suitableto the Internetenvironment.First, by moving locally to the sourcesite,a queryagentpermitsto

significantlysave bandwidth,becauseit is not necessaryto transfera largeamountof data,but thesearch

computationis performedlocally wherethedataresides.Second,MOMIS canqueryalsosourcesthatdo

nothavecontinuousconnections:thequeryagentmovesto thesourcesitewhentheconnectionis available,

performslocally thesearchevenif theconnectionis unstableor unavailable,andthenreturnsto theQM site

assoonastheconnectionis availableagain.Finally, thisfits well mobilecomputing,wheremobiledevices

(which canhostusers,MOMIS, or sources)do not have permanentconnectionswith the fixed network

infrastructure.

As statedabove,agentsmustinteractandcoordinatewith eachotherto accomplishtheir tasks.In par-

ticular, aninteractionthatworthbeingoutlinedis theonebetweenthequeryagentsandthewrapperagents.

6

Whena queryagentarrivesata sitewherea sourceresides,it hasto submitits queryto thewrapperagent,

and,then,to retrieve the results.This interactioncanoccurby usingdifferentprotocols.MOMIS agents

exploit FIPA speechactsto exchangeinformation,andwethink it is thesimplestwaybecauseit enablesto

sendthewrapperagentamessagecontainingthequeryitself. However, it is notobviousthatsuchcommu-

nicationmechanismis allowedon everysite,becauseof securityandlocalizationissues.Anotherpossible

protocolthatwe areevaluatingis basedon a shareddataspaceswhereagentsput andretrieve messages,

uncouplingin this way the interactionsandachieving moreflexibility . See[17] for a comparisonamong

differentkindsof coordinationfor Internetapplicationsbasedonmobileagents;anapproachthatis gaining

groundmoreandmorein theInternet-agentareais theonebasedonprogrammabletuplespaces[16], which

enablecontext-dependentcoordination[19].

2.2 Wrapping semistructuredsources

During the information integration process,many problemsarisedue to structuraland implementation

heterogeneity(includingfor exampledifferencesin hardwareplatforms,DBMS, datalanguages)andfrom

thesemanticheterogeneity, whendifferentnamesareemployedto representthesameinformation(naming

conflicts)or whendifferentmodellingconstructsareusedin differentsourcesto representthesamekind of

information.In the following we explain theapproachadoptedby a wrapperagentto manageinteraction

with thesources.

To managetheimplementationheterogeneity, a mediatorsystemtypically encapsulateseachsourceby

a wrapper, which logically convertsthe underlyingdatastructureto the commondatamodel.Thus, the

wrapperarchitectureandinterfacesarecrucial,for managingthediversityof datasources[38].

In particular, themaintasksof awrappercomponentare:

– during the integrationprocess,the wrappertranslatesthe schemaof a sourceinto the commondata

modelof themediator(in ourapproachby usingthelanguageODL );– during the queryingprocess,the wrapperconvertsqueriesposedover the commondatamodel into

requestssuitablefor the sources,and it convertsdatareturnedby the sourceinto the commondata

model.

For aconventionalstructuredinformationsource(e.g.relationaldatabasesor objectorienteddatabases),

the schemadescriptionis alwaysavailableandcanbe directly translatedinto the selectedcommondata

modelby theavailablewrappers.For example,for objectorienteddatabasesandflat files, MOMIS wrap-

persperformasyntactictranslation,while for relationaldatabasesthetranslationis basedontransformation

rule-sets,asdescribedin [24], to performrelationalto ODMG schemaconversion.For semistructuredin-

formationsources(e.g.,Webdatasources),a schemadescriptionis in generalnot directly availableat the

7

sources;in fact,a basiccharacteristicof semistructureddatais that they are“self-describing”,hencethe

informationassociatedwith theschemais specifiedwithin data [12].

Thus, to integratea semistructuredsource,a specificwrapperhasto implementa (semi)-automatic

methodologyto extract andexplicitly representthe schemaof the source.In [35,13] methodologiesto

extractstructuresfromsemistructureddatafilesareproposed.Moreover, in orderto integratesemistructured

informationsources,wehave to take into accountthatseveraldatamodelsrepresentingthiskind of source

have beenproposedin literature.In particular, theOEM model(ObjectExchangeModel) maybethought

asthe de factomodelto representsemistructureddata.It wasoriginally introducedfor the Tsimmisdata

integrationproject[22].

Following theOEM model,MOMIS representssemistructuredinformationsourcesasrooted,labeled

graphswith the semistructureddata(e.g.,an imageor free-formtext) as nodesand labelson edges.A

semistructuredobject(object,for short)canbeviewedasa triple of theform � id, label, value � ,

whereid is theobjectidentifier, label is a stringdescribingwhat theobjectrepresents,andvalue is

the value,that canbe atomicor complex. An atomicvaluecanbe integer, real, string, image,while the

complex valueis a setof semistructuredobjects,that is, a setof pairs(id,label).A complex objectcanbe

thoughtastheparentof all theobjectsthat form its value(childrenobjects).A givenobjectcanhave one

or moreparents.We denotethe fact that an objectso’ is a child objectof anotherobjectso by so �so’ andusenotationlabel(so) to denotethe label of so. In semistructureddatamodels,labelsare

descriptive asmuchaspossible.Generally, the samelabel is assignedto all objectsdescribingthe same

conceptin a givensource.To representtheschemaof a semistructuredsourceS, we introducethenotion

of objectpattern. All objectsso of S arepartitionedinto disjoint sets,denotedset,suchthat all objects

belongingto thesamesethavethesamelabel.An objectpatternis thenextractedfrom eachsetto represent

all theobjectsin theset.Then,anobjectpatternis representative of all differentobjectsthatdescribethe

sameconceptin agivensemistructuredsource.An objectpatterndescriptionfollowsopenworld semantics

typicalof theDescriptionLogicsapproach[44]. Objectsconformingto a patternsharea commonminimal

structurerepresentedby non-optionalproperties,but canhave furtheradditional(i.e., optional)properties.

In this way, objectsin a semistructureddatasourcecanevolve andaddnew properties,but they will be

retrievedasvalid instancesof thecorrespondingobjectpatternwhenprocessinga query. In thefollowing,

wedescribeanexampleof wrapper, in particular, in chargeof dealingwith XML sources.

A wrapper for XML-based files Accordingto theadopteddatamodel(ODM ), wedevelopedawrapper

to manageXML files. This wrapperaimsto mapthe datamodelof an XML file into the corresponding

objectpatternmodel.Thiswrappercouldbethoughtasthecoreof furtherextensionsthataimatmanaging

otherXML basedfiles,suchasXHTML files.

8

TheExtensibleMarkupLanguage(XML) [43] is a W3C recommendationandit arisesasa language

to describeinformationsourcesby usinga universalformat.Oneof themain goalsof this standardis to

exchangefilesacrosstheInternet.In comparisonwith HTML, XML canbesynthesizedasfollows [40]:

– Theusermaydefinepersonaltagnamesatwill;

– Thestructureof theXML file mayincludetagsnestedto any level;

– A XML file may includea descriptionof its structurefor useby applicationsthat needto perform

structuralvalidation.Thisdefinitionis calledDTD (DocumentTypeDefinition).

In this way, a XML file may be thoughtasself-describinglike a semistructureddatasource.The main

analogieswith ourmodelof a semistructureddatasourcemaybesummarizedasfollows:

– objectpatternattribute ��� XML tag

– objectpattern��� DTD element

– atomicvalueof anobjectpatternattribute ��� PCDATA value

By using this mapping,our wrapperparsesthe DTD associatedto eachwell-formedXML file and

generatesatranslationfromanXML statementintoanODL statement.Thismappingimpliessomecritical

aspectsdueto thelackof semanticsof XML w.r.t. ODL . In particular, themostrelevantare:theorderin

which attributesaredescribedin theDTD, thetranslationof theconceptof attribute from XML language

into ODL language,the poor type systemprovided by XML and the weak semanticsof intra-schema

references.In thelattercase,to avoid lossof informationduringthetranslationprocess,thedesignermaybe

askedto supplyfurtherinformationby a graphicalinterface.XML Schema,a recentW3Cstandard,allows

to expressmoresemanticson structuresanddatatypes,by usinga XML-basedlanguage.Our wrapperis

ableto integrateXML-Schemasourcesgeneratingamoresignificanttranslationin ODL , furthermorethe

wrappercanbeextendedto manageotherXML-basedlanguages(i.e.RDF, XHTML).

3 Integration Process

TheMOMIS approachto performGlobalVirtual View is articulatedin thefollowing phases:

1. Generationof a CommonThesaurus.

TheCommonThesaurusis asetof terminologicalintensionalandextensionalrelationships,describing

intra andinter-schemaknowledgeaboutclassesandattributesof sourcesschemas.We expressintra

andinter-schemaknowledgein form of terminologicalandextensionalrelationships(synonymy, hyper-

nymyandrelationship) betweenclassesand/orattributenames.In thisphase,to extractlexiconderived

relationshipstheWordNetdatabaseis used.

9

<!ELEMENT fiat(car*)>

<!ELEMENT car(name,engine,dimensions,tires,

performance,price)>

<!ELEMENT engine(name,cylinders?,layout?,

capacity_cc?,compression_ratio?,

power_kw, fuel_system)>

<!ELEMENT dimensions(length,width,height,

luggage_capacity)>

<!ELEMENT performance (urban_consumption,

combined_consumption,speed)>

<!ELEMENT name (#pcdata)> ...

Fig.4. Fiatdatabase(fiat)

2. Affinity analysisof classes.

Relationshipsin theCommonThesaurusareusedto evaluatethelevel of affinity betweenclasses,both

intra andinter sources.Theconceptof affinity is introducedto formalizethekind of relationshipsthat

canoccurbetweenclassesfrom theintegrationpointof view. Theaffinity of two classesis established

by meansof affinity coefficientsbasedon classnames,classstructuresandrelationshipsin Common

Thesaurus.

3. Clusteringclasses.

Classeswith affinity in differentsourcesaregroupedtogetherin clustersusinghierarchicalclustering

techniques.Thegoal is to identify theclassesthathave to be integratedsincedescribingthesameor

semanticallyrelatedinformation.

4. Generationof themediatedschema.

Unificationof affinity clustersleadsto theconstructionof thepredictedschema.A classis definedfor

eachcluster, which is representative of all cluster’s classesandis characterizedby the unionof their

attributes.Theglobalschemafor theanalyzedsourcesis composedof all classesderivedfrom clusters,

andis thebasisfor posingqueriesagainstthesources.

3.1 Running Example

In orderto illustratehow theMOMIS approachworks,wewill usethefollowing exampleof integrationin

theCarmanufacturingcatalogs,involving two differentdata-sourcesthatcollectinformationaboutvehicle.

The first data-sourceis theFIAT catalog,containingsemistructuredXML informationaboutcarsof the

Italiancarfactory(seeFigure4).

10

Vehicle(name, length, width, height)

Motor(cod m, type, compression ratio,

KW, lubrification, emission)

Fuel Consumption(name, cod m, drive trains,

city km l, highway km l )

Model(name, cod m, tires, steering, price)

Fig.5. Volkswagendatabase(vw)

Theseconddata-sourceis theVolkswagen database(vw) reportedin Figure5, a relationaldatabase

containinginformationaboutthis kind of car. Both databaseschemataarebuilt by analyzingthewebsite

of this factory.

In the following we introducethe generationof the CommonThesaurusassociatedwith the example

domain,startingfrom thelexiconrelationshipsdefinitionby usingWordnet.

3.2 Generationof a Common Thesaurus

The CommonThesaurusis a set of terminologicalintensionaland extensionalrelationships,describing

intra andinter-schemaknowledgeaboutclassesandattributesof sourcesschemas;it providesa reference

to definethe identificationof classescandidateto integrationand subsequentderivation of their global

representation.In theCommonThesaurus,weexpressknowledgein form of intensionalrelationships(SYN,

BT, NT, andRT) andextensionalrelationships(SYN ����� , BT ����� , andNT ����� betweenclassesand/orattribute

names.

TheCommonThesaurusis constructedthroughanincrementalprocessduringwhich relationshipsare

addedin thefollowing order:

1. schema-derivedrelationships: Intensionalandextensionalrelationshipsholdingat intra-schemalevel.

TheserelationshipsareextractedanalyzingeachODL schemaseparately. In particular, intra-schema

RT relationshipsareextractedfrom the specificationof foreign keys in relationalsourceschemasor

complex attributes(relationships)in objectorienteddatabase.Whenaforeignkey is alsoaprimarykey

bothin theoriginalandin thereferencedrelation,a BT/NT relationshipis extracted.We show themost

significantintra-schemarelationshipautomaticallygeneratedfrom MOMIS:�vw.Model RT vw.vehicle ��vw.Model RT vw.motor ��fiat.engine RT fiat.car �

11

2. lexical-derivedrelationships: Terminologicalrelationshipsholdingat inter-schemalevel areextracted

by by analyzingdifferentsourcesODL schemastogetheraccordingto theWordnetsuppliedontology.

Considerthefiat andthevw sources,themostsignificantlexical relationshipsderivedusingWordNet

arethefollowing:�fiat.car SYN vw.vehicle ��fiat.engine.compression ratio SYN

vw.motor.compression ratio ��fiat.dimension BT vw.vehicle.width �

3. designer-suppliedrelationships: Intensionalandextensionalrelationshipssupplieddirectly by thede-

signer, to capturespecificdomainknowledgeaboutthe sourceschemata.Considerthevw source,in

which themodelentitycanbeconsideredasaspecializationof thevehicleentity. Thisrelationshipcan

not beautomaticallyextractedusingboththelexical andthestructuralapproaches,hencewe supplied

thefollowing relationship:�vw.Model NT fiat.car �

This is a crucial operation,becausethe new relationshipsareforcedto belongto the CommonThe-

saurusandthususedto generatetheglobalintegratedschema.Thismeansthat,if anonsenseor wrong

relationshipis inserted,the subsequentintegrationprocesscanproducea wrong global schema.Our

systemhelpsthe designerin detectingwrong relationshipsby performinga Relationshipsvalidation

stepwith ODB-Tools.Validationis basedon thecompatibilityof domainsassociatedwith attributes.

Referringto theCommonThesaurusresultingfrom our example,we show somesignificantrelation-

ships(for eachrelationship,controlflag [1] denotesa valid relationship,while [0] aninvalid one):�fiat.performance.combined consumption RT�vw.fuel consumption.highway km l � [0]�fiat.dimensions BT vw.vehicle.height � [1]�vw.Model.name RT vw.vehicle.name � [1]

4. inferredrelationships: in thisstepMOMIS performsreasoningabouttheCommonThesaurusrelation-

shipsby exploiting thesubsumptionandinheritancecomputationperformedby ODB-Tools[9]. In the

examineddomainODB-Toolssysteminfersthefollowing relationships:�vw.Model RT fiat.dimensions ��vw.Model NT fiat.engine ��vw.motor NT fiat.car �

All theserelationshipsareaddedto theCommonThesaurusandthusconsideredin thesubsequentphaseof

constructionof GlobalSchema.For a moredetaileddescriptionof theabovedescribedprocesssee[10].

12

Terminologicalrelationshipsdefinedin eachstephold at the intensionallevel by definition.Further-

more,in eachof theabove stepthedesignermay“strengthen”a terminologicalrelationshipsSYN, BT and

NT betweentwo classes� � and � � by establishingthat they hold alsoat theextensionallevel, thusdefin-

ing alsoanextensionalrelationship.Thespecificationof anextensionalrelationship,on onehand,implies

theinsertionof a correspondingintensionalrelationshipin theCommonThesaurusand,on theotherhand,

enablesubsumptioncomputation(i.e., inferredrelationships)andconsistency checkingbetweenthe two

classes� � and � � .

Global Classand Mapping Tables

Startingfrom theoutputof theclustergeneration,wedefine,for eachcluster, aGlobalClassthatrepresents

the mediatedview of all the classesof the cluster. For eachglobal classa set of global attributesand,

for eachof them,the intensionalmappingswith the local attributes(i.e. theattributesof the local classes

belongingto thecluster)aregiven2.

Shortly, we cansaythat the global attributesareobtainedin two steps:(1) Union of the attributesof

all theclassesbelongingto thecluster;(2) Fusionof the“similar” attributes;in this stepredundanciesare

eliminatedin asemi–automaticwaytakinginto accounttherelationshipsstoredin theCommonThesaurus.

For eachglobal classa persistentmapping-tablestoringall the intensionalmappingsis generated;it is a

tablewhosecolumnsrepresentthe set of the local classeswhich belongto the clusterandwhoserows

representtheglobalattributes.

The final stepof the integrationprocessprovidesthe export of the Global Virtual View into a XML

DTD, by addingtheappropriateXML TAGsto representthemappingtablerelationships.Theuseof XML

in thedefinitionof theGlobalVirtual View lets to useMOMIS infrastructurewith otheropenintegration

informationsystemsby theinterchangeof anXML datafile. In addition,theCommonThesaurusis trans-

latedinto XML file, sothatMOMIS mayprovidea sharedontologythatcanbeusedby differentsemantic

ontologylanguages[25,23].

In thereferringexamplethefollowing GlobalClassesaredefined:

– Vehicle:containsthevehicle,model,car sourceclassesandaglobalattributesindicatesthesource

name;

– Engine:containstheengine, motor sourceclasses;

– Performance:containstheperformance, fuel consumption sourceclasses;

– Dimensions:containsthedimensions sourceclass.

2 For a detaileddescriptionof the mappingselectionandof the tool SI-Designerwhich assistthe designerin this

integrationphasesee[6].

13

Dynamic local schemamodification

Now, let supposethata modificationoccursin a localsource,for examplein theFIAT DTD anew element

truck is added:

<!ELEMENT truck(name, engine, dimensions, price, capacity)>

In this casethe local wrapperagent thatresidesat theFIAT sourcecreatesa mobileagentthatgoesto

theMOMIS siteandchecksthepermissionto modify theGlobalSchema:if theagentis enabled,it directly

performstheintegrationphases(i.e. CommonThesaurus, Clustering, MappingGeneration) causedby the

schemamodificationandnotifiesto the Integration Designerits actions.Otherwise,if the agenthasnot

enoughrights,it delegatesthewholeintegrationre-processto theIntegrationDesigner.

In our examplethenew truck elementis insertedby the local wrapperagent in theVehicleGlobal

Classandthemappingis performed.

3.3 The Query Manager Agents

Theuserapplicationinteractswith MOMIS to querytheGlobalVirtual View by usingtheOQL language.

This phaseis performedby the QM componentthat generatesthe OQL queriesfor wrapperagents,

startingfrom aglobalOQL queryformulatedby theuserontheglobalschema.UsingDescriptionLogics

techniques,the QM componentcan generatein an automaticway the translationof the genericOQL queryinto differentsub-query, onefor eachinvolvedlocal source.Thequeryagentsaremobileagentsin

chargeof sendingsuchsub-queriesto thedatasourcesitesandtherethey interactwith the local wrapper

agentsto carry out the queries;thenthey reportthe dataresultto the QM. To achieve the global answer,

theQM hasto mergeeachlocal sub-queryresultreportedby anagentinto a unifieddataset.This process

involvesthe solutionof redundancy andreconciliationproblems,dueto the incompleteandoverlapping

informationavailableon thelocalsources,i.e.ObjectFusion[36].

For example,thequery“retrieve thecarnameandpricefor power levelspresentin every sources”,is

expressedasfollows:

Q: select V1.name, V1.power, V1.price,

from vehicle V1

where 1 <= (select count(*)

from vehicle V2

where V2.power = V1.power

and V2.source <> V1.source)

14

Processingthe above global query would individuateall the local classesinvolved: vw.Vehicle,

vw.Model, vw.Motor, fiat.car, fiat.engine.

TheQM, by thequeryreformulationprocess[10], definesthefollowing local queries(QL1 expressed

by SQLandQL2 by XQuery[39]):

QL1: select M1.name, Motor.KW, M1.price

from vw.Model M1, vw.Motor

where M1.cod_m = Motor.cod_m

QL2: FOR $C in document("fiat.xml")//car

RETURN

<car>

<name>$C/name</name>

<price>$C/price</price>

<kw>$C/engine/power_kw</kw>

</car>

Wrapperagent

FIATSource

QueryManager

Wrapperagent

VWSource

Query agents

Fig.6. Parallelqueriesto two sourceswith thesizeof exchangeddata

Sincethetwo queriesareindependent,theQM couldassigneachoneto aqueryagent,soastheretriev-

ing processis executedin parallelby two agentsthatmove to the two local sources(seeFigure6). After

having performedthequery, eachqueryagentcomesbackto the MOMIS siteandreturnsthe dataresult

15

to theQM, which fusesthetwo data-setobtainingthefinal result(in theexamplethefusionis obtainedby

joining thedata-seton thepowerattribute).Thefinal queryis thefollowing:

QF: select DISTINCT QL1.name, QL1.KW, QL1.price

from QL1, QL2

where QL1.KW = QL2.power_kw

UNION

select DISTINCT QL2.name, QL2.power_kw, QL2.price

from QL1, QL2

where QL1.KW = QL2.power_kw

Let ussupposethateachquery(QL1, QL2) reportsanamountof N Kbytesof dataresultsandthatthe

join queryQF mantainsonly the50%of data(N/2). Thetotal amountof transferreddatais 2 * N Kbytes,

while the”useful” datais N Kbytes.

Wrapperagent

FIATSource

QueryManager

Wrapperagent

VWSource

Query agents

Fig.7. Queryagentinteractionscandecreasethesizeof theexchangeddata

Following amoreautonomousapproach,for thisclassof queries(i.e.,whereafusionof dataderivedby

differentlocal sourcesis necessary)thedataprocessshouldbemovedto thesystemswheresourcesreside

to minimizethedatatransfer. In fact,queryagentscanobtainintermediateresults,which aresentto other

queryagentsto performfusionat thelocal sites(seeFigure7).

For instance,thefollowing servicequeriesusinglocalquerylanguageareadded:

QS1: select DISTINCT Motor.KW

16

from vw.Motor

QS2: FOR $KW in distinct(document("fiat.xml")//car)

RETURN

<kw>$KW</kw>

Theseservicequeriesareexecutedby eachqueryagentsto the local sourcesandthedataresults(con-

tainingonly theavailableenginepower) areexchangedwith theotheragent.Let ussupposethat thedata

amountof theseservicequeriesarenegligible. In eachsinglesourcetheintermediatedatasetis joinedto the

local queriesto obtainthe fusedresult(the join attributesareextractedfrom themappingtablesof which

eachqueryagenthasacopy):

Q1: select DISTINCT QL1.name, QL1.KW, QL1.price

from QL1, QS2

where QL1.KW = QS2.power_kw

Q2: select DISTINCT QL2.name, QL2.power_kw, QL2.price

from QL2, QS1

where QL2.power_kw = QS1.KW

In this way, theunionof Q1 andQ2 resultsarethesameobtainedby thefirst shown approach,while

the datatransferamountwith the MOMIS centralsite is drasticallyreduced,sinceonly the requestdata

aremovedby theagent.Supposingthatashortnegligible message(theservicequeries)betweentheagents

impliesa transferof only the”useful” datato MOMIS, thetotal amountof transferreddatais now aboutN

Kbytes,lessof the2 * N Kbytesof thepreviousapproach.

Moreover, if the sitesof the involved sourcesare“near” (in Internetmetrics)oneeachotherbut are

“distant” from theMOMIS site,onequeryagentthatperformsa trip query(seeFigure8 sidea) requires

only two connectiontransfersbetweendistantsites(oneto go andoneto comeback),while a traditional

client-serverapproachwould requiretwo of suchconnectionsfor each remotesource(seeFigure8 sideb),

thuswastingmorebandwidth.For instance,if anAmericanuserwantto get informationaboutItalian car,

all sourcesitesarelikely to belocatedin Europe,while theMOMIS sitemanagingtherequestis locatedat

theothersideof theAtlantic Ocean.

17

Wrapperagent

FiatSource

QueryManager

Wrapperagent

VWSource

Wrapperagent

RenaultSource

Query agent

Wrapperagent

FiatSource

QueryManager

Wrapperagent

VWSource

Wrapperagent

RenaultSource

Slow connections

Fast connectionsa) b)

Requestsand

replies

Fig.8. Agentapproach(a) vs. traditionalapproach(b)

4 RelatedWork

Although thereareseveral mobile-agentapproachesconcerninginformationretrieval [15,45,3], demon-

stratingthat mobility canbe effectively exploited in distributed informationmanagement,thereare few

agent-basedapproachesin theareaof informationintegration.

A significative work is the MCC InfoSleuth(tm)[37,27]. It is an agent-basedsystemfor information

gatheringandanalysistasksperformedover networksof autonomousinformationsources.A key motiva-

tion of theInfoSleuthsystemis thatrealinformationgatheringapplicationsrequirelong-runningmonitoring

andintegrationof informationatvariouslevelsof abstraction.To thisend,InfoSleuthagentsenablea loose

integrationof technologiesallowing: (1) extractionof semanticconceptsfrom autonomousinformation

sources;(2) registrationandintegrationof semanticallyannotatedinformationfrom differentsources;and

(3) temporalmonitoring,informationrouting,andidentificationof trendsappearingacrosssourcesin the

informationnetwork. Anotherexperienceis theRETSINA multi-agentinfrastructurefor in-context infor-

mationretrieval [41]. In particular, the LARKS descriptionlanguage[42] is definedto realizethe agent

matchmakingprocess(at bothsyntacticandsemanticlevel) by usingseveraldifferentfilters:Context, Pro-

file, Similarity, Signatureand Constraintmatching.Differently from our approach,both InfoSleuthand

18

RETSINA doesnot takeadvantagefrom themobility featureof theagents,whichweconsiderfundamental

in a dynamicanduncertainenvironmentsuchastheInternet.

In the areaof heterogeneousinformationintegration,many projectsbasedon a mediatorarchitecture

have beendeveloped.Themediator-basedTSIMMIS project[22] follows a ‘structural’ approachanduses

aself-describingmodel(OEM) to representheterogeneousdatasources,rulesexpressedin MSL (Mediator

SpecificationLanguage)to enforcesourceintegrationandpatternmatchingtechniquesto performaprede-

finedsetof queriesbasedon a querytemplate.Differently from MOMIS proposal,in TSIMMIS only the

predefinedqueriesmaybeexecutedandfor eachsourcemodificationa manuallymediatorrulesrewriting

mustbeperformed.

In thefollowing wepresentothersystems,whosemaindifferencewith regardto MOMIS is thelackof

a tool aid-supportfor thedesignerin theintegrationprocess.

The GARLIC project[20] builds up on a complex wrapperarchitectureto describethe local sources

with anOO language(GDL), andon thedefinitionof Garlic Complex Objectsto manuallyunify thelocal

sourcesto defineaglobalschema.

The SIMS project[2] proposesto createa global schemadefinition by exploiting the useof Description

Logics (i.e., theLOOM language)for describinginformationsources.Theuseof a globalschemaallows

bothGARLIC andSIMSprojectsto supportall possibleuserquerieson theschemainsteadof apredefined

subsetof them.

TheInformationManifold system[33] providesa sourceindependentandqueryindependentmediator.

Theinputschemaof InformationManifold is asetof descriptionsof thesources.Givenaquery, thesystem

will createaplanfor answeringthequeryusingtheunderlyingsourcedescriptions.Algorithmsto decidethe

usefulinformationsourcesandto generatethequeryplanhave beenimplemented.Theintegratedschema

is definedmainlymanuallyby thedesigner, while in ourapproachit is tool-supported.

Infomaster[28] provides integratedaccessto multiple distributed heterogenousinformationsources

giving the illusion of a centralized,homogeneousinformation system.It is basedon a global schema,

completelymodelledby theuser, andacoresystemthatdynamicallydeterminesanefficientplanto answer

theuser’squeriesby usingtranslationrulesto harmonizepossibleheterogeneitiesacrossthesources.

5 Conclusionsand Future Work

This paperhaspresentedhow a systemfor information integrationcanbe improved by the exploitation

of mobile agents.In particular, agentsare useful in the managementof the sources,which, in an open

anddynamicscenariolike theInternet,canbespreadandcanchangein a uncontrolledway. Themobility

featurepermitsto overcomethelimitationsof thetraditionalapproachesof distributedsystems.

With regardto futurework, wesketchsomeresearchdirections.

19

Thefirst onerelatesto thedynamicintegrationof informationsources.Currently, whena new sourceis

added(or whenis deleted),theMOMIS systemhasto berestarted.Ouraimis to allow sourceintegrationat

runtime,possiblyby exploiting mobileagentsthatsearchfor interestingnew sourcesor checktherequest

of anadministratorto integratehisnew source.Thiswill permitto facethehighdegreeof dynamismof the

Internet.

Then,we areevaluatingwhich further componentsof the systemcanbe modelledasagents,and,in

particular, which onescantake advantagefrom themobility feature.The fact thatsomecomponentsmay

bemobilepermitsto deploy thewholesystemin a moreflexible andadaptableway.

Finally, an areathat is worth to be explored is the interactionbetweenagents,which can occur in

differentwaysandwith differentlanguages.Appropriatelanguagesor interactionmeansshouldbechosen

to grantinteroperabilityandflexibility to agent-basedsystems.Ontheonehand,complex languagesfor the

knowledgeexchangehavebeenproposed[26]; on theotherhand,mobility of agentspromotestheadoption

of simpleanduncoupledcoordinationprotocols[18].

Acknowledgements

Work supportedby ItalianMinistry for UniversityandScienceandTechnologyResearch(MURST)within

theprojectframeworksof “MUSIQUE: Infrastructurefor QoSin WebMultimediaServiceswith Hetero-

geneousAccess”and“D2I: Integration,Warehousing,andMining of HeterogeneousDataSources”and

AgentLink II - Europe’sNetwork of Excellencefor Agent-basedComputing.

References

1. Y. Arens,C.Y. Chee,C. Hsu, andC. A. Knoblock. Retrieving and integratingdatafrom multiple information

sources.InternationalJournalof IntelligentandCooperativeInformationSystems, 2(2):127–158,1993.

2. Y. Arens,C. A. Knoblock,andC. Hsu. Queryprocessingin thesimsinformationmediator. AdvancedPlanning

Technology, 1996.

3. J.Baek,J.Yeo,G.Kim, andH. Yeom.Costeffectivemobileagentplanningfor distributedinformationretrieval. In

Proceedingsof the21��� InternationalConferenceon DistributedComputingSystemsi(ICDCS), Phoenix,Arizona,

Apr 2001.

4. C. Batini, M. Lenzerini,andS. B. Navathe. A comprehensive analysisof methodologiesfor databaseschema

integration.ACM ComputingSurveys, 18(4):322–364,1986.

5. F. Bellifemine,G. Caire,T. Trucco,andG. Rimassa.Jadeprogrammer’s guide,jade2.4,sep2001.

6. I. Benetti,D.Beneventano,S.Bergamaschi,A. Corni,F. Guerra,andG. Malvezzi.Si-designer:a tool for intelligent

integrationof information. InternationalConferenceonSystemSciences(HICSS2001), January2001.

7. I. Benetti,D.Beneventano,S.Bergamaschi,andM. Vincini. An informationintegrationframework for e-commerce.

IEEEIntelligentSystemsMagazine, 2002.

20

8. D. Beneventano,S. Bergamaschi,S. Castano,A. Corni, R. Guidetti,G. Malvezzi,M. Melchiori, andM. Vincini.

Information integration:The momisprojectdemonstration.In VLDB 2000,Proceedingsof 26�! International

Conferenceon VeryLarge DataBases,September, 2000,Cairo, Egypt, pages611–614.MorganKaufmann,2000.

9. D. Beneventano,S. Bergamaschi,C. Sartori,andM. Vincini. ODB-QOPTIMIZER:A tool for semanticquery

optimizationin oodb. In Int. ConferenceonDataEngineering- ICDE97, 1997.http://sparc20.dsi.unimo.it.

10. S. Bergamaschi,S. Castano,D. Beneventano,andM. Vincini. Semanticintegrationof heterogenousinformation

sources.Journalof DataandKnowledge Engineering, 36(3):215–249,2001.

11. S. Bergamaschi,S. Castano,andM. Vincini. Semanticintegrationof semistructuredandstructureddatasources.

SIGMODRecords, 28(1),March1999.

12. P. Buneman.Semistructureddata. In Proc. of 1997Symposiumon Principlesof DatabaseSystems(PODS97),

pages117–121,Tucson,Arizona,1997.

13. P. Buneman,S.Davidson,M. Fernandez,andD. Suciu. Adding structureto unstructureddata. In Proc.of ICDT

1997, pages336–350,Delphi,Greece,1997.

14. P. Buneman,L. Raschid,andJ.Ullman. Mediatorlanguages- a proposalfor a standard,April 1996. Availableat

ftp://ftp.umiacs.umd.edu/pub/ONRrept/medmodel96.ps.

15. G. Cabri,L. Leonardi,andF. Zambonelli.Agentsfor InformationRetrieval: Issuesof Mobility andCoordination.

Journalof SystemsArchitecture, 46(15):1419–1433,December2000.

16. G. Cabri,L. Leonardi,andF. Zambonelli.MARS: A ProgrammableCoordinationArchitecturefor Mobile Agents.

IEEEInternetComputing, 4(4):26–35,July/August2000.

17. G. Cabri,L. Leonardi,andF. Zambonelli. Mobile-AgentCoordinationModelsfor InternetApplications. IEEE

Computer, 33(2):82–89,February2000.

18. G. Cabri, L. Leonardi,andF. Zambonelli. XML Dataspacesfor the Coordinationof InternetAgents. Applied

Artificial Intelligence, 15(1):35–58,January2001.

19. G. Cabri,L. Leonardi,andF. Zambonelli.Engineeringmobileagentapplicationsvia context-dependentcoordina-

tion. IEEETransactionsonSoftware Engineering, 2002.to appear.

20. M.J. Carey, L.M. Haas,P.M. Schwarz,M. Arya, W.F. Cody, R. Fagin,M. Flickner, A.W. Luniewski, W. Niblack,

D. Petkovic, J. Thomas,J.H. Williams, andE.L. Wimmers. Objectexchangeacrossheterogeneousinformation

sources.Technicalreport,StanfordUniv., 1994.

21. R. G. G. Cattell,editor. TheObjectDatabaseStandard: ODMG 3.0. MorganKaufmannPublishers,SanMateo,

CA, 2000.

22. S.Chawathe,GarciaMolina, H., J.Hammer, K.Ireland,Y. Papakostantinou,J.Ullman,andJ.Widom. TheTSIM-

MIS project:Integrationof heterogeneousinformationsources.In IPSJConference, Tokyo,Japan, 1994.

23. DAML JointCommittee.DamlProject.Availableat http://www.daml.org.

24. C. FahrnerandG. Vossen.Transformingrelationaldatabaseschemasinto objectorientedschemasaccordingto

odmg-93.In Proceedingsof InternationalConferenceonDeductiveandObjectDatabases, pages429–446,LNCS

Vol 1013,1995.

25. D. Fensel,I. Horrocks,F. vanHarmelen,S.Decker, M. Erdmann,andM. Klein. OIL in anutshell.In Proceedingsof

theEuropeanKnowledgeAcquisitionConference(EKAW-2000), LectureNotesIn Artificial Intelligence.Springer-

Verlag,2000.To appear.

21

26. Tim Finin, YannisLabrou,and JamesMayfield. KQML as an agentcommunicationlanguage. In Jeffrey M.

Bradshaw, editor, SoftwareAgents, chapter14,pages291–316.AAAI Press/ TheMIT Press,1997.

27. J. Fowler, B. Perry, M. H. Nodine,andB. Bargmeyer. Agent-basedsemanticinteroperabilityin infosleuth. SIG-

MOD Record, 28(1):60–67,1999.

28. M. R. Genesereth,A. M. Keller, andO. Duschka.Infomaster:An informationintegrationsystem.In Proceedings

of 1997ACM SIGMODConference, 1997.

29. J. Gilarranz,J. Gonzalo,and F. Verdejo. Using the eurowordnetmultilingual semanticdatabase.In Proc. of

AAAI-96SpringSymposiumCross-Language Text andSpeech Retrieval, 1996.

30. R. Hull and R. King et al. Arpa i�

reference architecture, 1995. Available at

http://www.isse.gmu.edu/I3Arch/index.html.

31. NicholasR. Jenningsand Michael J. Wooldridge,editors. Agent Technology: Foundations,Applications,and

Markets. Springer-Verlag,Berlin, 1998.

32. NeeranM. Karnik andAnandR. Tripathi. Designissuesin mobile-agentprogrammingsystems.IEEE Concur-

rency, 6(3):52–61,July/September1998.

33. A. Y. Levy, A. Rajaraman,andJ.J.Ordille. Queryingheterogeneousinformationsourcesusingsourcedescriptions.

In Proc.of VLDB1996, pages251–262,1996.

34. A.G. Miller. Wordnet:A lexical databasefor english.Communicationsof theACM, 38(11):39–41,1995.

35. S.Nestorov, S.Abiteboul,andR.Motwani. Inferringstructurein semistructureddata.SIGMODRecord, 26(4):39–

43,1997.

36. Y. Papakonstantinou,S.Abiteboul,andH. Garcia-Molina.Objectfusionin mediatorsystems.In VLDBInt. Conf.,

Bombay, India,September1996.

37. B. Perry, M. Taylor, andA. Unruh. Informationaggregationandagentinteractionpatternsin infosleuth(tm). In

Proceedingsof theFourth IECISInternationalConferenceonCooperativeInformationSystems, 1998.

38. M.T. RothandP. Scharz.Don’t scrapit, wrap it! a wrapperarchitecturefor legacy datasources.In Proc. of the

23rd Int. Conf. onVeryLargeDatabases, Athens,Greece,1997.

39. S.Boag,D. Chamberlin,D. Florescu,J.Robie,J.Simeon,andM. Stefanescu.Xquery1.0:An xml querylanguage.

In W3CWorkingDraft 15February2001, Dec2001.

40. D. Suciu.Semistructureddataandxml, 1998.

41. K. Sykara.In-context informationmanagementtruoughadaptativecollaborationof intelligentagents.In Intelligent

InformationAgents. M. KluschEd.- Springer, 1999.

42. K. Sykara,M. Klusch,S. Widoff, andJ. Lu. Dynamicservicematchmakingamongagentsin openinformation

environments.SIGMODRecords, 28(1),March1999.

43. The World Wide Web Consortium (W3C). Extensible markup language (xml). Available at

http://www.w3.org/XML/.

44. W. A. WoodsandJ.G. Schmolze.TheKL-ONE family. In F. W. Lehman,editor, SemanticNetworksin Artificial

Intelligence, pages133–178.PergamonPress,1992. Pubblishedasa Specialissueof Computers & Mathematics

with Applications, Volume23,Number2-9.

22

45. J. Yang,V. Honavar, L. Miller, andJ. Wong. Intelligent mobile agentsfor informationretrieval andknowledge

discovery from distributed dataand knowledgesources. In Proceedingsof the IEEE Information Technology

Conference, pages99–102,Syracuse,New York, 1998.

23