Data Interlinking

Data interlinking

Jerome Euzenat

Montbonnot, [email protected]

June 10, 2015

Data interlinlingSimilarity-based approach

Key-based interlinkingOntology matching & data interlinking

Tools

The problem: RDF data interlinking

3

〈http://data.bnf.fr/12144801/edgar allan poe the gold bug/, dc:title, “The gold bug”〉The gold bug

title

creator

en

E. Poe

lang

firstname lastname

Writer

Work

rdf:type

rdf:type

b a1 a2

Baudelaire Malarme

The raven

orig

name namename

orig

authortranslator translator

Person

Book

rdf:type

rdf:type

≈

≥

≤

≥Jerome Euzenat Data interlinking 3 / 0



Tools

Goal of the lecture

I Provide an overview of the problem of data interlinking

I Describe broad categories of solutions

I Point to useful tools for generating links

Mostly about generating links, not on finding how to generate them

Jerome Euzenat Data interlinking 4 / 0



Tools

Outline

Data interlinling

Similarity-based approach

Key-based interlinking

Ontology matching & data interlinking

Tools




Tools

Data interlinking

I use (with the same meaning):

I instance matching

I entity linking

I data interlinking

I do not use:

I record linkage

I data deduplication

I entity reconciliation

I coreference resolution




Tools

The data interlinking problem

Data interlinking is the task of finding same entities within different datasets(RDF graphs).

Data source 1 Data source 2

interlinking

owl:sameAs




Tools

The data interlinking process

Data source

Data source

interlinking Resulting linksSample links

parameters

resources




Tools

The data interlinking process (2)

d

d ′

extraction

Linkage spec

generation l

interlinking




Tools

Approaches to data interlinking

There are two main approaches to data interlinking:

I similarity-based: resources are compared through a similarity measureand if they are similar enough, they are the same.

I key-based: sufficient conditions for two resources to be the same areinduced and used to find same entities




Tools

Classification of similarities

Data interlinking techniques may be based on:

I Data ID (URIs);

I Data keys

I External relations: (explicit or implicit) links to other resources

I Data description (content)




Tools

Manual resource matching

URI1 URI2

Manual observation

owl:sameAs

I This does not scale.

I But may be good for a first sample or reference.

I Crowdsourcing?




Tools

URI matching

URI1 URI2

URI transformation

owl:sameAs

http://dbpedia.org/resource/Johann Sebastian Bach owl:sameAs

http://www.lastfm.fr/music/Johann+Sebastian+Bach

http://rdf.insee.fr/geo/regions-2011.rdf#REG 11 ?

http://ec.europa.eu/eurostat/ramon/rdfdata/nuts2008/FR10




Tools

Id matching

id id

Finding same ids

owl:sameAs

You can find such types of ids:

I Social security numbers

I ISBN, DOI, MAC addresses, etc.

I authorities: ISO (countries, languages), IATA (airports)

Most databases are built on such identifiers. . . but they are often local to thedatabase.




Tools

Context-based similarity

URI1 URI2

VIAF

Context-based“similarity”

owl:sameAs

Process:I Project your data into another resource (DBPedia, geonames, viaf, etc.)I Assess relations between considered termsI Import the relation in the dataset

This harness the power of links!Jerome Euzenat Data interlinking 16 / 0



Tools

Content-based similarity

3

The gold bug

title

creator

E. Poe

firstname lastname

Writer

Work

rdf:type

rdf:typeb a1 a2

Baudelaire Poe

Le corbeauLe scarabe d’or

orig

name name

title

authortranslator

Person

Book

rdf:type

rdf:type

Compute similarity

owl:sameAs

Two main approaches:

I bag of text

I structured similarity




Tools

Term-based similarity

The gold bug

E. Poe

firstname lastname

Writer

Work

type

type Baudelaire Poe

Le corbeauLe scarabe d’or

orig

name name

title

authortranslator

Person

Book

type type

Compute “bag of words” similarity

owl:sameAs

Various tools:I Normalisation (Stemmer, Tokenizers)I Use of linguistic resources (Wordnet)I TranslationI Many similarity measures, especially from information retrieval




Tools

Structure similarity

title

creator

firstname lastname

type

type orig

name name

title

authortranslator

type

type

Compute structure similarity

owl:sameAs

Techniques:

I Based on graph matching techniques

I Can be used to learn weights on properties (but need matching)

I Problem: scalability




Tools

Cross-lingual RDF data interlinking

http://a.org/Mus999 France

Musee du Louvre

nom

lieu

Paris

99,rue de Rivoli

75001

adresse

ville

rue

zip

http://bb.cn/盧浮宮

盧浮宮

法國巴黎

稱號

位於

owl:sameAs ?




Tools

Similarity-based data interlinking

RESOURCE RESOURCE

SIMILARITY

owl:sameAs ?

Hypothesis: ↑ similarity ↑ probability that it is the same object

DOCUMENT DOCUMENTSIMILARITY

owl:sameAs ?

Yuzhong Qu, Wei Hu, Gong Cheng: Constructing virtual documents for ontology matching. WWW 2006: 23-31.

DOCUMENT(zh) DOCUMENT(en)

DOCUMENT(en)

translation

DOCUMENT(zh)

translationSIMSIM

SIMILARITY

owl:sameAs ?

BabelNet(IDs) BabelNet(IDs)SIMILARITY

owl:sameAs ?




Tools

General cross-lingual interlinkingframework

1 VirtualDocuments

3 SimilarityComputation

4 LinkGeneration

2 LanguageNormalization




Tools

Building virtual documents by levels

http://dbpedia.org/resource/Charles Perrault

Charles Perrault

dbpedia:France

Level 1

France is a sovereigncountry in Western Eu-rope that includes over-seas regions and territo-ries. . .

Level 2




Tools

Machine translation: parameters

1 VirtualDocuments

2.1 MachineTranslation

2.2 NLPPreprocessing


4 LinkGeneration

Level 1

Level 2

ZH→ENLowercase+Tokenize+ Filter stop words

+ Stemming (Porter)

+ Bigrams (terms)

TF+cosine

TF*IDF+cosine

Greedy

Hungarian

32 settings have been explored in total




Tools

Lcase+Tokenization with TF*IDF atLevel 1

0 - 0.11

0.11 - 0.15

0.15 - 0.25

0.25 - 0.35

0.35 - 0.45

0.45 - 1




Tools

Adding noise




Tools

BabelNet method: parameters

1 VirtualDocuments

2 MultilingualKB Mapping


4 LinkGeneration

Level 1

Level 2

TF+cosine

TF*IDF+cosine

Greedy

Hungarian




Tools

Database keys

I A set of attributes which uniquely identifies elements of a relation

I e.g., Book: isbn, People: fistname, lastname, birthplace, birthdate

I usually given and used to check integrity

They may be used for identifying same entities across two databases.But they require alignments.




Tools

Example of interlinking with keys andalignments

Are the resources bnf:cb118949856 and bne:XX1721208 the same?

I if BNF ontology states foaf:Person owl:hasKey {foaf:name, dc:dates}I and we have the following alignment

foaf:Person

bnf:cb118949856

Albert Camus

07-11-1913

04-01-1960

Romancier, dramaturge et essayiste

http://id.loc.gov/vocabulary/countries/fr

Mondovi (Algerie)

1913-1960

foaf:name

rda:dateOfBirth

rda:dateOfDeath

rda:biographicalInformation

rda:countryAssociatedWithThePerson

rda:placeOfBirth

dc:dates

frbrer:C1005

bne:XX1721208

Camus, Albert

1913-1960

Aut [...]1980

frber:P3039

frber:P3040

rda:sourceConsulted

w

≡

≡

≈

≈

owl:sameAs

owl:sameAs ?


http://id.loc.gov/vocabulary/countries/fr



Tools

Key-based interlinking methods

Database keys allow for identifying entities: if they are aligned, this can beused for linking.

I AdvantagesI they are logically groundedI they allow to minimize the number of properties to compare (if we use

minimal keys)

I DrawbacksI Require alignment between properties and classesI Very few key axioms are available, and they are not necessarily useful for

interlinking

We overcome these drawbacks by introducing link keys




Tools

Link key

A link key

〈{〈p1, q1〉, . . . , 〈pn, qn〉}{〈p′1, q′1〉, . . . , 〈p′m, q′m〉} linkkey 〈c , d〉〉

holds iffFor all pairs of instances a and b belonging respectively to classes c and d ofontologies O and O′,

if a and b share at least one value (object) for each pairs ofproperties pi and qi respectively,

and a and b share all their values (objects) for each pairs ofproperties p′i and q′i respectively,

then they are the same (〈a, owl:sameAs, b〉).

Example:

〈{〈foaf:name, frbr:P3039〉}{〈dc:dates, frbr:P3040〉} linkkey 〈foaf:Person, frbr:C1005〉〉




Tools

Link key extraction

Problem: How to induce such link keys from data?

The number of set of pairs of properties is exponential

Our approach:

I discover only candidate link keys.

I evaluate them in order to select only the “good” ones




Tools

Candidate link key

A candidate link key is a set of property pairs {〈p1, q1〉, . . . , 〈pk , qk〉} that

1. would generate at least one link if used as a link key

2. is maximal for at least one link, or is the intersection of severalcandidate link keys




Tools

Supervised selection measures

If a sample of reference links is available:

I Positive examples (L+) : a set of owl:sameAs links

I Negative examples (L−) : a set of owl:differentFrom links

Idea: Approximate precision and recall on that sample

Definition (Relative precision and recall)

precision(K , L+, L−) =|L+ ∩ LD,D′(K )|

|(L+ ∪ L−) ∩ LD,D′(K )|

recall(K , L+) =|L+ ∩ LD,D′(K )|

|L+|




Tools

Unsupervised selection measures

When no reference link is available.Idea: measuring how close the extracted links would be fromone-to-one and total.

Definition (Discriminability)

disc(K ,D,D ′) =min(|{a : 〈a, b〉 ∈ LD,D′(K )}|, |{b : 〈a, b〉 ∈ LD,D′(K )}|)

|LD,D′(K )|

Definition (Coverage)

cov(K ,D,D ′) =|{a : 〈a, b〉 ∈ LD,D′(K )} ∪ {b : 〈a, b〉 ∈ LD,D′(K )}|

|{a : c(a) ∈ D} ∪ {b : d(b) ∈ D ′}|




Tools

Experimental evaluation

These selection measures were evaluated on public datasets.

Finding links between French municipalities described in two differentdatasets:

I Insee dataset: 36700 instances;

I Geonames dataset: 36552 instances.

The reference link set is composed of:

I Positive links: 36552 owl:sameAs statements;

I owl:differentFrom links derived from owl:sameAs links (closed worldassumption).




Tools

Evaluation

The algorithm extracted 11 candidate link keys:

bad F-measure≈ 0

high F-measure≈ .99

good F-measure≈ 0.89

{1} {2} {3, 4} {5, 6}

{7, 1} {2, 1} {3, 4, 1} {3, 2, 4}

{3, 7, 4, 1} {3, 2, 4, 1}

{3, 7, 2, 4, 1}

coveraged

iscr

imin

abili

ty

1 = 〈nom, name〉 2 = 〈nom, alternateName〉3 = 〈subdivisionDe, parentFeature〉 4 = 〈subdivisionDe, parentADM3〉5 = 〈codeINSEE, population〉 6 = 〈codeCommune, population〉7 = 〈nom, officialName〉




Tools

Evaluation

Correlation between the harmonic means of discriminability and coverage andF-measure:

bad F-measure≈ 0

high F-measure≈ .99

good F-measure≈ 0.89

{1} {2} {3, 4} {5, 6}

{7, 1} {2, 1} {3, 4, 1} {3, 2, 4}

{3, 7, 4, 1} {3, 2, 4, 1}

{3, 7, 2, 4, 1}

coverage

dis

crim

inab

ility

h-mean(disc.,cov)≈ .99 h-mean(disc.,cov)≈ .89 h-mean(disc.,cov) ≈ 0

1 = 〈nom, name〉 2 = 〈nom, alternateName〉3 = 〈subdivisionDe, parentFeature〉 4 = 〈subdivisionDe, parentADM3〉5 = 〈codeINSEE, population〉 6 = 〈codeCommune, population〉7 = 〈nom, officialName〉




Tools

Why using ontologies?

Because it is obvious that we must compare the instances of equivalentclasses based on equivalent properties.

More precisely:

I For reducing the search space for finding link keys and similarities

I For reducing the scope of linkage specifications

I Because not the same linkage rules work for the same classes

I Because classes and properties are hint like others of the similaritybetween resources

Ex. With similarity and with keys




Tools

Data interlinking through a commonontology

o

URI1 URI2

Resource matching

of datasets

described by the

same ontology

owl:sameAs




Tools

Matching with a common ontology

+ Focus the search: only match instances of the same class;

– Not sufficient: it remains to identify corresponding entities

+ If keys are defined (OWL 2), this is done;+ At least we know which properties to compare;– Inferring secondary keys may be useful;– Correcting discrepancies: record linkage.




Tools

Record linkage

Name Johann

Date 1665-03-21

Place Munchen

NameJohannes

Date31/03/1665

PlaceMonaco di Bavaria

Having a common ontology does not solve all problems.




Tools

Different types of mismatch

I Different domains, connected (BIM, Energy demand)⇒ few correspondences, any type

I Same domain, different models (engineer, policy maker)⇒ many correspondences, mostly equivalence

I Same domain, different granularity (city management, building design)⇒ many correspondences, mostly subsumption




Tools

Data interlinking with differentontologies (implicit alignment)

o o ′

URI1 URI2

Resource matching

of datasets

described by

different ontologies

owl:sameAs




Tools

Data interlinking with differentontologies (explicit alignment)

o o ′

URI1 URI2

A

Resource matching

of datasets

described by

different ontologies

owl:sameAs




Tools

Ontology matching for data interlinking

o o ′

URI1 URI2

Ontology matching

A

Data interlinking

owl:sameAs




Tools

Heterogeneity problem

Resources being expressed in different ways must be reconciled before beingused.Mismatch between formalized knowledge can occur when:

I different languages are used (OWL vs. Topic maps);I different terminologies are used:

I English vs. Chinese;I Book vs. Monograph.

I different models are used:I different classes: Autobiography vs. Paperback;I classes vs. property: Essay vs. literarygenre;I classes vs. instances: One physical book as an instance vs. one work as

an instance.

I different scopes and granularity are used.I Only books vs. cultural items vs. any product;I Books detailed to the print and translation level vs. books as works.




Tools

Ontology alignment

Item

DVD

Book

Paperback

Hardcover

CD

pricetitledoicreatorpp

author

integer

string

uri

Person

Monograph

Essay

Literary critics

Politics

Biography

Autobiography

Literature

pages

isbnauthor

title

subject

Human

Writer

≥

≥

≥

≤

≥




Tools

Expressive alignments (EDOAL)

Pocket

Booktopic

author=

Volume

size14≥

Autobiography

v

=

∀x ,Pocket(x)⇐ Volume(x) ∧ size(x , y) ∧ y ≤ 14

∀x ,Book(x) ∧ author(x , y) ∧ topic(x , y) ≡ Autobiography(x)




Tools

Example: INSEE dataset

Region table:

code nom chef-lieu

11 Ile-de-France 7505621 Champagne-Ardenne 5110822 Picardie 80021

Sous-region table:

region departement

11 7511 7711 7811 9111 9211 93




Tools

Example: Administrative ontology

Territoire FR

Pays

Region

Departement

Arrondissement

Commune

codenom

chef-lieusubdivision

integer

string




Tools

Example: NUTS dataset

NUTSRegion table:

level code name hasParentRegion

0 FR FRANCE

1 FR1 ILE DE FRANCE FR

2 FR10 Ile de France FR13 FR101 Paris FR103 FR104 Essonne FR10




Tools

Example: Linking INSEE and NUTS

NUTS: Nomenclature of territorial units for statistics

#INSEE INSEE name NUTS Level #NUTS1 Pays 0 34

1 14226 Region 2 344

100 Departement 3 1488342 Arrondissement

4036 Canton 452422 Commune 5




Tools

Example: Linking INSEE and NUTS

Territoire FR

Pays

Region

Departement

Commune

PAYS FR

REG 11

DEP 75

DEP 77

DEP 78

COM 75056

Region

Country

NUTSRegion

LAURegion

FR

UK

FR1

FR10

FR101

FR102

FR103

owl:sameAs

owl:sameAs

owl:sameAs

owl:sameAs

owl:sameAs




Tools

Example: Linksets

Specific data sets containing URIs.

<http://www.example.org/linkset/INSEE-NUTS>

a void:Linkset ;

void:target <http://rdf.insee.fr/geo/regions-2011.rdf>;

void:target <http://nuts.psi.enakting.org/id/>;

insee:PAYS FR owl:sameAs nuts:FR

insee:REG 11 owl:sameAs nuts:FR10

insee:DEP 75 owl:sameAs nuts:FR101






Tools

Example: interesting sets

nuts

onsordnance s. igninsee

geonames dbpedia freebase




Tools

A simple algorithm

I Find matching concepts [concept matching];

I For each of them, determine matching properties based on the similaritybetween their values in both datasets [property matching];

I From them find property combinations identifying corresponding entities[key extraction];

I Link corresponding entities [link generation].

For instance, nom/RegionINSEE ⊆ name/NUTSRegionNUTS and moreoverthey are unambiguous.




Tools

INSEE and NUTS: ontology alignment

Territoire FR

Pays

Region

Departement

Arrondissement

Canton

Commune

codenom

chef-lieusubdivision

integer

string

Region

Country

NUTSRegion

LAURegion

name

level

code

hasSubRegion=

≤

≤

≤

≤

≤

=




Tools

Simple alignments are not sufficient

Territoire FR

Region

Departement

Commune

nom

DEP 75

nom

COM 75056

nom

Region

NUTSRegion

name

FR101

name

Paris

=

=

=

≤≤

≤

=

=

=




Tools

Expressive alignments are necessary

Region

NUTSRegion

level

hasParentRegion

2 =

FR=

=

subdivision hasSubRegion=

nom name=




Tools

What does this mean?

I Ontology alignments are schema-level expression of correspondences;

I They are useful for focussing the search;

I Expressive alignments are necessary;

I They can be turned into SPARQL-based link generators.

but it is also necessary to express instance level constraints:

I for converting data (e.g., mph vs. m/s);

I for expressing matching constraint on data (e.g., similarity).




Tools

Data interlinking and ontology matching

d

o

d ′

o ′Matcher

A

Generator

l




Tools

Tools for data interlinking

Linkage spec extraction generation

similarity LIMES Silk, LIMES, OpenRefine

key LinkKeyDisco SPARQL




Tools

Silk

Silk is a robust software for interlinking data sets.

It relies on an expressive specification of linking conditions:

I Declare data sources (DataSource);

I Circumscribe entities to compare (Source/TargetDataset);I Describe how to compare them (LinkageRule):

I Select properties to compare through paths (Input);I Compute distances between them (Compare+threshold);I Aggregate all comparisons (Aggregate);

I Select those pairs of entities to be linked (Filter);

I Generate links (Output+thresholds).




Tools

A Silk script

Consider a linking script between INSEE and NUTS:

<Silk>

<Prefix id="nuts"

namespace="http://ec.europa.eu/.../geographic.rdf#" />

<Prefix id="insee"

namespace="http://rdf.insee.fr/geo/" />

<DataSource id="nuts2008"

type="sparqlEndpoint">

<Param name="endpointURI"

value="http://localhost:9091/.../internal"/>

<Param name="graph"

value="http://localhost:9091/.../nuts2008-complete-1"/>

</DataSource>

<DataSource id="insee2010"

type="sparqlEndpoint">

<Param name="endpointURI"

value="http://localhost:9091/.../internal"/>

<Param name="graph"

value="http://localhost:9091/.../source/regions-2010-1"/>

</DataSource>

<Thresholds accept="0.9" verify="0.7" />

<Outputs>

<Output type="sparul">

<Param name="graphUri"

value="http://localhost:9091/.../source/insee-nuts-silk"/>

<Param name="uri"

value="http://localhost:9091/.../lifted/"/>

<Param name="parameter" value="update"/>

</Output>

</Outputs>

<Interlinks>

<Interlink id="linkingNUTS">

<LinkType>owl:sameAs</LinkType>

<SourceDataset dataSource="nuts2008" var="s">

<RestrictTo>?s rdf:type nuts:NUTSRegion.

?s nuts:level 2.

</RestrictTo>

</SourceDataset>

<TargetDataset dataSource="insee2010" var="ss">

<RestrictTo>?ss rdf:type insee:Region</RestrictTo>

</TargetDataset>

<LinkageRule>

<Aggregate type="max">

<Compare metric="levenshteinDistance"

threshold=".2">

<Input path="?s/nuts:name"/>

<Input path="?ss/insee:nom"/>

</Compare>

</Aggregate>

</LinkageRule>

</Interlinks>

</Interlink>

</Silk>




Tools

Silk: prefix and sources

<Silk>

<Prefix id="nuts" namespace="http://ec.europa.eu/.../geographic.rdf#" />

<Prefix id="insee" namespace="http://rdf.insee.fr/geo/" />

<DataSource id="nuts2008" type="sparqlEndpoint">

<Param name="endpointURI" value="http://localhost:9091/.../internal"/>

<Param name="graph" value="http://localhost:9091/.../nuts2008-complete-1"/>

</DataSource>

<DataSource id="id1" type="file">

<Param name="file" value="/Skratch/TutoLinking/admin/regions-2010.rdf"/>

<Param name="format" value="RDF/XML" />

</DataSource>

Sources can be files or SPARQL endpoint.




Tools

Silk rules

<Interlinks>

<Interlink id="linkingNUTS">

<LinkType>owl:sameAs</LinkType>

<SourceDataset dataSource="nuts2008" var="s">

<RestrictTo>?s rdf:type nuts:NUTSRegion.

?s nuts:level 2.

</RestrictTo>

</SourceDataset>

<TargetDataset dataSource="insee2010" var="ss">

<RestrictTo>?ss rdf:type insee:Region</RestrictTo>

</TargetDataset>

<Thresholds accept="0.9" verify="0.7" />

<Outputs>

<Output type="sparul">

<Param name="graphUri" value="http://localhost:9091/.../source/insee-nuts-silk"/>

<Param name="uri" value="http://localhost:9091/.../lifted/"/>

<Param name="parameter" value="update"/>

</Output>

</Outputs>

Restrictions are given in SPARQL graph patternsOutput can be file (in various format, including the Alignment API) or aSPARQL endpoint.They can be made dependent on thresholds.




Tools

Silk rules (cont’ed)

<LinkageRule>

<Aggregate type="max">

<Compare metric="levenshteinDistance" threshold=".2">

<Input path="?s/nuts:name"/>

<Input path="?ss/insee:nom"/>

</Compare>

</Aggregate>

</LinkageRule>

</Interlinks>

</Interlink>

</Silk>

They can:

I transform the data (lowercase, tokenize, to integers, etc.),

I use comparison metrics (equality, levenshtein, Jaro-Winkler, etc.), and

I aggregate their values (average, min, max, etc.).




Tools

Silk workbench




Tools

EDOAL Alignments

<Cell>

<entity1><e:Class rdf:about="&insee;Region"/></entity1>

<entity2>

<e:Class>

<e:and rdf:parseType="Collection">

<e:Class rdf:about="&nuts;NUTSRegion"/>

<e:AttributeValueRestriction>

<e:onAttribute><e:Property rdf:about="&nuts;level"/></e:onAttribute>

<e:comparator rdf:resource="&edoal;equals"/>

<e:value><e:Literal e:type="&xsd;integer" e:string="2" /></e:value>

</e:AttributeValueRestriction>

<e:AttributeValueRestriction>

<e:onAttribute>

<e:Relation rdf:about="&nuts;hasParentRegion" />

</e:onAttribute>

<e:comparator rdf:resource="&edoal;equals"/>

<e:value><e:Instance rdf:about="&esdata;FR" /></e:value>

</e:AttributeValueRestriction>

</e:and>

</e:Class>

</entity2>

<relation>equivalence</relation>

<measure>1.0</measure>

...

</Cell>




Tools

Link keys in the Alignment API

<e:linkkey>

<e:Linkkey>

<e:binding>

<e:Intersects>

<e:property1><e:Property rdf:about="&insee;nom" /></e:property1>

<e:property2><e:Property rdf:about="&nuts;name" /></e:property2>

</e:Intersects>

<e:Equals>

<e:property1>

<e:Property>

<e:inverse><e:Property rdf:about="&insee;subdivision" /></e:inverse>

</e:property1>

<e:property2><e:Property rdf:about="&nuts;hasParentRegion" /></e:property2>

</e:Equals>

</e:binding>

</e:Linkkey>

</e:linkkey>




Tools

Query generation

PREFIX insee: <http://rdf.insee.fr/ontologie-geo-2006.rdf#>

PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#>

SELECT ?r

FROM <http://rdf.insee.fr/geo/regions-2011.rdf>

WHERE {?r rdf:type insee:Region .

}

PREFIX nuts: <http://ec.europa.eu/eurostat/ramon/ontologies/geographic.rdf#>


SELECT ?n

FROM <http://ec.europa.eu/eurostat/ramon/rdfdata/nuts2008/>

WHERE {?n rdf:type nuts:NUTSRegion .

?n nuts:level 2^^xsd:integer .

?n nuts:hasParentRegion nuts:FR .

}




Tools

Data transformation




CONSTRUCT {?r rdf:type nuts:NUTSRegion .

?r nuts:level 2^^xsd:integer .

?r nuts:hasParentRegion nuts:FR .

}FROM <http://rdf.insee.fr/geo/regions-2011.rdf>


}




Tools

SameAs link generation generation


PREFIX nuts: <http://ec.europa.eu/eurostat/ramon/ontologies/geographic.rdf#>


CONSTRUCT { ?r owl:sameAs ?n . }FROM <http://rdf.insee.fr/geo/regions-2011.rdf>

FROM <http://ec.europa.eu/eurostat/ramon/rdfdata/nuts2008/>


?r insee:nom ?l .

?n rdf:type nuts:NUTSRegion .

?n nuts:name ?l .

?n nuts:level 2^^xsd:integer .

?n nuts:hasParentRegion nuts:FR .

}




Tools

Other issue: performances

n ×m n3 ×

m3 + n

3 ×m3 + n

3 ×m3

10× 10 = 1001000× 1000 = 1000000

100000× 100000 = 10000000000




Tools


Blocking: index+cluster

Dataset 1 Dataset 2




Tools


Blocks can be obtained from:

I clustering values in index

I predefined block (based on equality)

I classes in an ontology (blocks are defined as class expressions)




Tools

Other issue: evaluation

d

d ′

interlinking l

Reference links

evaluation

Precision

Recall

F-measure




Tools

Other issue: learning

d

d ′

Training links interlinking l

evaluation




Tools

Conclusion

I Data interlinking is one of the most critical task in linked data

I . . . but not only, e.g. smart citiesI If faces many problems due to:

I heterogeneity (format, languages, convention)I size

I Interlinking can be based on similarities or keys

I There is active work to infer such interlinking pattern




Tools

Further reading

I T. Heath, C. Bizer, Linked Data: Evolving the Web into a Global DataSpace, Morgan & Claypool (US), 2011 http://linkeddatabook.com/

I J. Euzenat, P. Shvaiko, Ontology matching, 2nd ed., Springer,Heildelberg (DE), 2013 http://book.ontologymatching.org

I K. Stefanidis, V. Efthymiou, M. Herschel, V. Christophides, EntityResolution in the Web of Data, Tutorial, WWW conference, Seoul(KR), 2014 http://www.csd.uoc.gr/~vefthym/er/

Silk http://silk-framework.com/

Alignment API http://alignapi.gforge.inria.fr

Al 4 SC http://al4sc.inrialpes.fr


http://linkeddatabook.com/

http://book.ontologymatching.org

http://www.csd.uoc.gr/~vefthym/er/

http://silk-framework.com/

http://alignapi.gforge.inria.fr

http://al4sc.inrialpes.fr



Tools

Thanks

I To my colleagues Manuel Atencia, Jerome David, Nicolas Guillouet andFrancois Scharffe

I The Datalift and Lindicle projects

I The Ready4SmartCities project


http://exmo.inria.fr

Jerome . Euzenat @ inria . fr

Data Interlinking

Science

Transcript of Data Interlinking