si-ls 2008-09 - group 15 - sorbini, savioli, reale - presentation slides

8/14/2019 SI-LS 2008-09 - Group 15 - Sorbini, Savioli, Reale - Presentation Slides

1/88

Leveraging Data and Structure inOntology IntegrationO. Udrea L. Getoor R.J. Miller

Group 15Enrico Savioli Andrea Reale Andrea Sorbini

DEISUniversity of Bologna

Searching Information in Large Spaces Conference

March 11, 2009

Savioli, Reale, Sorbini (DEIS) UGM07 SI-LS 2009 1 / 39
http://find/http://goback/


2/88

Outline

1 Brief Introduction to Ontologies

What is an Ontology

2 Motivations and State of the Art

Bringing data together

Existing Solutions

3 ILIADS

Overview

The Algorithm

Experimental ResultsConclusions and Future Developments

4 Demo



3/88

Brief Introduction to Ontologies What is an Ontology

Ontologiesa quest for meaning

A very common problem in IT is data modeling Critical in the field of Information Systems A good data model enables good data usage

Tools used to describe concepts cannot expressthe implicit

interpretation of data

Semantic knowledge is lost after the modeling process

Example



4/88







Example



5/88







Example


B i f I d i O l i Wh i O l


6/88







Example

How can we find all of Queens members?


B i f I t d ti t O t l i Wh t i O t l


7/88


Ontologiesa formal model

Ontology

An ontology is a formal representationof

concepts within a domain (Song, Performer, Composer...)

relations between those concepts (plays-for, member-of, written-by...)individual instances (queen, f.mercury, b.may...)

Well-defined constructsto enrich descriptions with semantics

Formal models to enable automatic reasoning(DL, FOL)

(member(Y,Z) :- sings-for(Y,Z) ; plays-for(Y,Z))

OWL is the W3C standard for an ontology language




8/88



Ontology











9/88



Ontology











10/88


OWLA simple example

Whole lottalove

subClassOf subClassOfsubClassOf

Musician Writer

Iron MaidenVienna

PhilharmonicS. Harris

subClassOfsubClassOf

type typetype

type

ClassicalMusic

Ride of theValkyries

Requiem inD minor

Take the "A"Train

The Phantomof the Opera

typetype

type

JazzMusic

type

RockMusic

W.A. Mozart

composer-ofplays

Performer

performs

composer-of

Smoke onthe water

type

Music

type


Motivations and State of the Art Bringing data together


11/88


Motivating Scenario

The AAA principleAnyonecan sayAnythingaboutAny topic

The same thing might be described in different ways

There is a strong need to integrate heterogeneous schemas Not only in a web scenario

Example

DB schemas

XML schemas

Ontology schemas




12/88


Motivating Scenario




Example

DB schemas

XML schemas

Ontology schemas




13/88

g g g

Motivating Scenario




Example

DB schemas

XML schemas

Ontology schemas



14/88



15/88

Integration Techniques

Structure Based Matching

Uses schema meta-data(e.g.

informations about tables orconcepts models) to discover

mapping elements among

them, both on a structural and

element level

Instance Based Matching

Uses informations about data

instances(e.g. contents of atable or individuals of an

ontology) to discover mappings

among entities representing

them




16/88

Integration Techniques

Structure Based Matching

Uses schema meta-data(e.g.

informations about tables orconcepts models) to discover

mapping elements among

them, both on a structural and

element level

Instance Based Matching

Uses informations about data

instances(e.g. contents of atable or individuals of an

ontology) to discover mappings

among entities representing

them




17/88

Ontology Integration

Ontologies offer further ways to uncover aligning relations

between entities

Explicit theoretic model semantics can be leveraged to improve

integration quality

ExampleAssume that composed-by is a functional property

Requiem K626and Requiem in D minorare the same

If we have:

composed-by(Requiem K626, Mozart). composed-by(Requiem in D minor, W.A. Mozart).

Then we might say that aligning W.A. Mozartand Mozart is a

good choice




18/88



between entities


integration quality



If we have:



good choice




19/88



between entities


integration quality



If we have:



good choice




20/88



between entities


integration quality



If we have:



good choice




21/88



between entities


integration quality



If we have:



good choice


Motivations and State of the Art Existing Solutions


22/88

FCA-Merge and COMA++

FCA-Merge (human-aided tool to merge ontologies)

Collects domain related natural language documents

Searches those documents for ontologies concepts occurrences

Derives an alignment that has to bevalidated by an operator

COMA++ (framework with multiple match strategies)

Fragment based matching

Reuse of previous matching results

Comprehensive GUI for results evaluation




23/88

FCA-Merge and COMA++

FCA-Merge (human-aided tool to merge ontologies)

Collects domain related natural language documents

Searches those documents for ontologies concepts occurrences

Derives an alignment that has to bevalidated by an operator

COMA++ (framework with multiple match strategies)

Fragment based matching

Reuse of previous matching results

Comprehensive GUI for results evaluation




24/88

Whats missing

Both COMA++ and FCA-Merge use only structure level matching

Moreover none of them makes use of the semantic potential

offered by ontologies

There exist other solutions using reasoning support, however ... it is used only for an a-posteriori consistency checkof the result

An interesting improvement might be to leverage it for the actual

matching process




25/88

Whats missing






matching process




26/88

Whats missing






matching process




27/88

Whats missing






matching process

Thats where ILIADS kicks in!


ILIADS Overview


28/88

ILIADSIntegrated Learning In Alignment of Data and Schema

Takes two OWL Lite ontologies as input

Combines"traditional" schema matching approaches with alogical inference algorithm Inference results are used to influence confidence in a presumed

mapping

Makes use of both schema (structure) and data (individuals)

Outputsa set of axioms (the alignment) that tights the input

ontologies together


ILIADS Overview


29/88




mapping



ontologies together


ILIADS Overview


30/88




mapping



ontologies together


ILIADS Overview

ILIADS


31/88




mapping



ontologies together


ILIADS The Algorithm

Al ith O i


32/88

Algorithm Overview

INPUT: Consistent Ontologies O1 and O2OUTPUT: Alignment A

01: Initialize algorithms structures (O is O1 O2)02: repeat:

03: Compute similarity scores between clusters

04: Heuristically select a type of clusters05: for each couple (c, c) of that type do06: Determine a candidate relationship a(c,c)07: Perform incremental inference

08: Update similarity score

09: Select the best couple, update O and A

10: until there are clusters with similarity > t11: return A



Th Al ith


33/88

The AlgorithmStep by step



04: Heuristically select a type of clusters

05: for each couple (c, c) of that type do06: Determine a candidate relationship a(c,c)07: Perform incremental inference






Th Al ith


34/88

The AlgorithmStructures initialization

The algorithms groups equivalent entities in clusters

Clusters are classified by the type of their entities Clusters of Classes

Clusters of Properties Clusters of Individuals

A new alignment can result in Merging of clusters A new subsumption relationship between clusters

At the beginning a cluster is created for each entity



The Algorithm


35/88









The Algorithm


36/88









37/88


The Algorithm


38/88

The AlgorithmSimilarity Computation

Entities similarity score

sim(e, e) = x simlex(e, e) + s simstruct(e, e

) + e simext(e, e)

lexical: Jaro-Winkler (similar to edit distance) and thesauri

structural: Jaccard for neighborhoods (Jacd(S1, S2) =|S1S2||S1S2|

)

extensional: Jaccard on extensions

The set of parameters {x, s, e} is different for each entity type

Similarity between clusters is computed combining the similarities

of their entities



The Algorithm


39/88

The AlgorithmSimilarity Computation

Entities similarity score

sim(e, e) = x simlex(e, e) + s simstruct(e, e

) + e simext(e, e)

lexical: Jaro-Winkler (similar to edit distance) and thesauri

structural: Jaccard for neighborhoods (Jacd(S1, S2) =|S1S2||S1S2|

)

extensional: Jaccard on extensions

The set of parameters {x, s, e} is different for each entity type

Similarity between clusters is computed combining the similarities

of their entities



40/88


The Algorithm


41/88





05: for each couple (c, c

) of that type do06: Determine a candidate relationship a(c,c)07: Perform incremental inference






The Algorithm


42/88

The AlgorithmSelecting a relationship

Iterates over each couple (c, c) such that sim(c, c) is above athreshold t

For each of those a candidate relationship is chosen between: Equivalence Subsumption

The selection is done by looking at the intersection of entitiesextensions:

The set of its instance individuals, for a class The couples of individuals involved, for a property



The Algorithm


43/88








The Algorithm


44/88








The Algorithm


45/88

The AlgorithmSelecting a relationship - Example

MusicArtist

The BlackSabbath

W.A. Mozart

Music Piece

Requiem K626

Smoke onthe water

subClassOf

type

type

ClassicRock

type

subClassOfMetal

composer-of

subClassOf

Fusion

Quadrant Four

Whole lottalove

Paranoid

Orchestral

subClassOf

type

plays

type

type

type

type

subClassOf subClassOf

subClassOf

MusicianWriter

Iron Maiden

ViennaPhilharmonic

S. Harris

subClassOf

subClassOf

type

type

type

type

ClassicalMusic


Requiem in

D minor

Take the "A"Train


type

type

type

Jazz

Music

type

RockMusic

W.A. Mozart

composer-of

plays

Performer

performs

composer-of

Smoke onthe water

typeMusic

Iron Maiden

type

Billy Cobham

typecomposer-of

plays

plays




The Algorithm


46/88

e go tSelecting a relationship - Example

MusicArtist

The BlackSabbath

W.A. Mozart

Music Piece

Requiem K626

Smoke onthe water

subClassOf

type

type

ClassicRock

type

subClassOfMetal

composer-of

subClassOf

Fusion

Quadrant Four

Whole lottalove

Paranoid

Orchestral

subClassOf

type

plays

type

type

type

type


subClassOf

MusicianWriter

Iron Maiden

ViennaPhilharmonic

S. Harris

subClassOf

subClassOf

type

type

type

type

ClassicalMusic


Requiem in

D minor

Take the "A"Train


type

type

type

Jazz

Music

type

RockMusic

W.A. Mozart

composer-of

plays

Performer

performs

composer-of

Smoke onthe water

typeMusic

Iron Maiden

type

Billy Cobham

typecomposer-of

plays

plays


sameAs



The Algorithm


47/88

gSelecting a relationship - Example

MusicArtist

The BlackSabbath

W.A. Mozart

Music Piece

Requiem K626

Smoke onthe water

subClassOf

type

type

ClassicRock

type

subClassOfMetal

composer-of

subClassOf

Fusion

Quadrant Four

Whole lottalove

Paranoid

Orchestral

subClassOf

type

plays

type

type

type

type


subClassOf

MusicianWriter

Iron Maiden

ViennaPhilharmonic

S. Harris

subClassOf

subClassOf

type

type

type

type

ClassicalMusic


Requiem in

D minor

Take the "A"Train


type

type

type

Jazz

Music

type

RockMusic

W.A. Mozart

composer-of

plays

Performer

performs

composer-of

Smoke onthe water

typeMusic

Iron Maiden

type

Billy Cobham

typecomposer-of

plays

plays


sameAs

Number of instances belongingonly to "Rock Music" over the

total number of instancesconsidered:

2/3



The Algorithm


48/88




The Algorithm


49/88




The Algorithm


50/88

gStep by step








10: until there are clusters with similarity > t

11: return A



The Algorithm


51/88

Incremental Logical Inference

The inference step is used to: Look for inconsistencies of the candidate relationship Infer logical consequences of the new axiom Possibly enforce the confidence in it

The inference is not complete(it would be EXPTIME in OWL Lite) Only a small number of steps is actually performed However this may cause inconsistencies not to be found



The Algorithm


52/88






The Algorithm


53/88






The Algorithm


54/88


The inference step is used to: Look for inconsistencies of the candidate relationship Infer logical consequences of the new axiom

Possibly enforce the confidence in it




The Algorithm


55/88

Incremental Logical Inference - Example

MusicArtist

The BlackSabbath

Mozart

Music Piece

Requiem K626

Smoke onthe water

subClassOf

type

type

ClassicRock

type

subClassOfMetal

subClassOf

Fusion

Quadrant Four

Whole lottalove

Paranoid

Orchestral

subClassOf

type

plays

type

type

type

type


subClassOf

MusicianWriter

Iron Maiden

ViennaPhilharmonic

S. Harris

subClassOf

subClassOf

type

type

type

type

ClassicalMusic


Requiem inD minor

Take the "A"Train


type

type

type

Jazz

Music

type

RockMusic

W.A. Mozart

composer-of

plays

Performer

performs

composer-of

Smoke onthe water

typeMusic

Iron Maiden

type

Billy Cobham

typecomposer-of

plays

plays


composed-by



The Algorithm


56/88


MusicArtist

The BlackSabbath

Mozart

Music Piece

Requiem K626

Smoke onthe water

subClassOf

type

type

ClassicRock

type

subClassOfMetal

subClassOf

Fusion

Quadrant Four

Whole lottalove

Paranoid

Orchestral

subClassOf

type

plays

type

type

type

type


subClassOf

MusicianWriter

Iron Maiden

ViennaPhilharmonic

S. Harris

subClassOf

subClassOf

type

type

type

type

ClassicalMusic


Requiem inD minor

Take the "A"Train


type

type

type

Jazz

Music

type

RockMusic

W.A. Mozart

composer-of

plays

Performer

performs

composer-of

Smoke onthe water

typeMusic

Iron Maiden

type

Billy Cobham

typecomposer-of

plays

plays


sameAs

composed-by



The AlgorithmI l L i l I f E l


57/88


MusicArtist

The BlackSabbath

Mozart

Music Piece

Requiem K626

Smoke onthe water

subClassOf

type

type

ClassicRock

type

subClassOfMetal

subClassOf

Fusion

Quadrant Four

Whole lottalove

Paranoid

Orchestral

subClassOf

type

plays

type

type

type

type


subClassOf

MusicianWriter

Iron Maiden

ViennaPhilharmonic

S. Harris

subClassOf

subClassOf

type

type

type

type

ClassicalMusic


Requiem inD minor

Take the "A"Train


type

type

type

Jazz

Music

type

RockMusic

W.A. Mozart

composer-of

plays

Performer

performs

composer-of

Smoke onthe water

typeMusic

Iron Maiden

type

Billy Cobham

typecomposer-of

plays

plays


sameAs

composed-by

composed-by

inverseOf

composer-of



The AlgorithmI t l L i l I f E l


58/88


MusicArtist

The BlackSabbath

Mozart

Music Piece

Requiem K626

Smoke onthe water

subClassOf

type

type

ClassicRock

type

subClassOfMetal

subClassOf

Fusion

Quadrant Four

Whole lottalove

Paranoid

Orchestral

subClassOf

type

plays

type

type

type

type


subClassOf

MusicianWriter

Iron Maiden

ViennaPhilharmonic

S. Harris

subClassOf

subClassOf

type

type

type

type

ClassicalMusic


Requiem inD minor

Take the "A"Train


type

type

type

Jazz

Music

type

RockMusic

W.A. Mozart

composer-of

plays

Performer

performs

composer-of

Smoke onthe water

typeMusic

Iron Maiden

type

Billy Cobham

typecomposer-of

plays

plays


sameAs

composed-by

composed-by

composer-of



59/88


60/88


61/88


62/88


The AlgorithmUpdate similarity score


63/88

Update similarity score

This is the key point of the algorithm A value f - the influence factorof the inference - is computed

The f factor

f =

(e1,e2)Q

sim(e1, e2)1 sim(e1, e2)

Q: the set of entity pairs that became equivalent as a consequence of inference

f is used to update the similarity score for the current couple siminf(c, c

) = min(f sim(c, c), 1)



The AlgorithmUpdate similarity score


64/88

Update similarity score

This is the key point of the algorithm A value f - the influence factorof the inference - is computed

The f factor

f =

(e1,e2)Q

sim(e1, e2)1 sim(e1, e2)

Q: the set of entity pairs that became equivalent as a consequence of inference

f is used to update the similarity score for the current couple siminf(c, c

) = min(f sim(c, c), 1)



The AlgorithmUpdate similarity score - Example


65/88

Update similarity score Example

MusicArtist

The BlackSabbath

Mozart

Music Piece

Requiem K626

Smoke onthe water

subClassOf

type

type

ClassicRock

type

subClassOfMetal

subClassOf

Fusion

Quadrant Four

Whole lottalove

Paranoid

Orchestral

subClassOf

type

plays

type

type

type

type


subClassOf

MusicianWriter

Iron Maiden

ViennaPhilharmonic

S. Harris

subClassOf

subClassOf

type

type

type

type

ClassicalMusic


Requiem inD minor

Take the "A"Train


type

type

type

Jazz

Music

type

RockMusic

W.A. Mozart

composer-of

plays

Performer

performs

composer-of

Smoke onthe water

typeMusic

Iron Maiden

type

Billy Cobham

typecomposer-of

plays

plays


composed-by

composed-by

composer-of



66/88


67/88


68/88


69/88




70/88











The AlgorithmBuilding the alignment


71/88

By now for each candidate pair of clusters of a given type A possible relationship has been explored A set of consequences has been inferred An inference-weighted similarity has been computed

Before restarting the loop: The axiom a(c,c) with the highest similarity score is chosen It is added to the output alignment A

If a(c,c) is an equivalence c and c are merged



The AlgorithmBuilding the alignment


72/88

By now for each candidate pair of clusters of a given type A possible relationship has been explored A set of consequences has been inferred An inference-weighted similarity has been computed

Before restarting the loop: The axiom a(c,c) with the highest similarity score is chosen It is added to the output alignment A

If a(c,c) is an equivalence c and c are merged





73/88




05:for

each couple(c

,c

)of that type

do06: Determine a candidate relationship a(c,c)07: Perform incremental inference



10: until there are clusters with similarity > t

11: return A



The AlgorithmThe End(ing)


74/88

The algorithm halts when there are no more candidate clusters When there are no clusters having similarity greater than the

threshold t Remaining clusters are not likely to share any relationship

Intuitively ILIADS is guaranteed to terminate because Previously used cluster pairs are not re-used unless their score

changes The merging process decreases the number of clusters



The AlgorithmThe End(ing)


75/88

The algorithm halts when there are no more candidate clusters When there are no clusters having similarity greater than the

threshold t Remaining clusters are not likely to share any relationship

Intuitively ILIADS is guaranteed to terminate because Previously used cluster pairs are not re-used unless their score

changes The merging process decreases the number of clusters


ILIADS Experimental Results

Comparative resultsPrecision, Recall and F-1


76/88



Tests Analysis


77/88

The interplay between structure and instance integration delivershigher quality Significant improvement of recall Tests on ontologies without instance data showed comparable

results with the other systems

Lambda-tuning allows ILIADS to adapt itself better to particoular

pairs of ontologies

Inconsistent alignments due to limited inference steps were found

only in the .5% of tests



78/88


Tests Analysis


79/88

The interplay between structure and instance integration delivershigher quality Significant improvement of recall Tests on ontologies without instance data showed comparable

results with the other systems

Lambda-tuning allows ILIADS to adapt itself better to particoular

pairs of ontologies

Inconsistent alignments due to limited inference steps were found

only in the .5% of tests


ILIADS Conclusions and Future Developments

Summary


80/88

Explicit semantics provided by ontologies can improve thequality of data integration

Instance data exploitation could significatly enhance traditional

matching techniques

ILIADS leverages both these opportunities and shows promising

results

Future developments

Inter-ontology differentiated parameters

Alignment of distant sections of the ontologies in parallel



Summary


81/88



matching techniques


results

Future developments





Summary


82/88



matching techniques


results

Future developments



Savioli, Reale, Sorbini (DEIS) UGM07 SI-LS 2009 34 / 39 ILIADS Conclusions and Future Developments

Summary


83/88



matching techniques


results

Future developments



Savioli, Reale, Sorbini (DEIS) UGM07 SI-LS 2009 34 / 39 Demo

Demo


84/88

And now an ultra-fancy demo

Savioli Reale Sorbini (DEIS) UGM07 SI-LS 2009 35 / 39 Appendix For Further Reading

For Further Reading I


85/88

W3C OWL resources.

http://www.w3.org/2004/OWL/.

D. Aumueller, H.H. Do, S. Massmann, and E. Rahm.

Schema and ontology matching with COMA++.

In Proceedings of the 2005 ACM SIGMOD international

conference on Management of data, pages 906908. ACM NewYork, NY, USA, 2005.

I. Horrocks, P.F. Patel-Schneider, and F. van Harmelen.

From SHIQ and RDF to OWL: the making of a Web Ontology

Language.

Web Semantics: Science, Services and Agents on the World Wide

Web, 1(1):726, 2003.

Savioli Reale Sorbini (DEIS) UGM07 SI-LS 2009 36 / 39 Appendix For Further Reading

For Further Reading II


86/88

Y. Kalfoglou and M. Schorlemmer.Ontology mapping: the state of the art.

The knowledge engineering review, 18(01):131, 2003.

E. Rahm and P.A. Bernstein.

A survey of approaches to automatic schema matching.

The VLDB Journal The International Journal on Very Large Data

Bases, 10(4):334350, 2001.

P. Shvaiko and J. Euzenat.

A survey of schema-based matching approaches.

Lecture Notes in Computer Science, 3730:146171, 2005.

Savioli Reale Sorbini (DEIS) UGM07 SI-LS 2009 37 / 39


87/88

Appendix For Further Reading


88/88

Thank you.

Group 15

A. Sorbini E. Savioli A. Reale

Savioli Reale Sorbini (DEIS) UGM07 SI-LS 2009 39 / 39

si-ls 2008-09 - group 15 - sorbini, savioli, reale - presentation slides

Documents