Populating DBpedia FR and using it for Extracting Information

12
Julien Plu [email protected] @julienplu Populating DBpedia FR and using it for Extracting Information

Transcript of Populating DBpedia FR and using it for Extracting Information

Julien Plu

[email protected]

@julienplu

Populating DBpedia FR and using it for

Extracting Information

Agenda

Mapping the French infoboxes

How is DBpedia FR used at Orange?

Presentation of the Orange challenge

Project: ExtSem

Module 1: ParseText

Module 2: BuildDepGraph

Module 3: ExtractRDF

Module 4: SelectRDF

Experiments

09/02/2015 - 3rd DBpedia Community Meeting – Dublin, Ireland - 2

Mapping the French infoboxes

The set of mappings has grown significantly

during the last three years (2012-2015)

208 infoboxes have mappings

I contribute to 100 mappings

This amounts to 50% of the articles in the French

Wikipedia which have an infobox

Example:

Infobox Communes de France (mapping): 36765

occurrences

Infobox Musique (œuvre) (mapping): 29429 occurrences

09/02/2015 - 3rd DBpedia Community Meeting – Dublin, Ireland - 3

How is DBpedia FR used at Orange?

Used as a knowledge graph for the in-house

Web search engine

Used to interlink background knowledge with

internal data about films (AlloCine) and music

(Deezer)

Used as a knowledge provider for public tools

in IPTV

Used for recommendation system in VOD

service

09/02/2015 - 3rd DBpedia Community Meeting – Dublin, Ireland - 4

Presentation of the Orange challenge

Team members:

Guillaume Viland

Jonathan Marchand

Julien Plu

Internal challenge for getting new research

projects

Only two weeks to get something to present

09/02/2015 - 3rd DBpedia Community Meeting – Dublin, Ireland - 5

Project : ExtSem

Goal: extracting relations among named

entities in raw text

Example:

L'excentrique Lady Gaga est au coeur de l'actu depuis

qu'elle a dévoilé son single "Applause" issu de son

quatrième album à découvrir à partir du 11 novembre.

Results:

Subject predicate object

Lady Gaga etre aucoeurdeactu

Lady Gaga devoiler Applause (chanson)

09/02/2015 - 3rd DBpedia Community Meeting – Dublin, Ireland - 6

Module 1: ParseText

.txt

Tokenizer

et PoS

Tagger :

Melt

.conll06

.inmalt

Parser :

MaltParser

• Part of Speech Tagger and

Parser are stochastic and

trained with the French

Dependency Treebank

• Deep syntactic analysis with

dependencies

09/02/2015 - 3rd DBpedia Community Meeting – Dublin, Ireland - 7

Module 2: BuildDepGraph

.conll06 .nerd

buildDe

pGraph

.depnt

• This module merges

the output from the

NERD framework with

the syntactic analysis

• The output is in RDF

modeled with a

vocabulary mapped on

French POS tags

09/02/2015 - 3rd DBpedia Community Meeting – Dublin, Ireland - 8

Module 3: ExtractRDF

.depnt example

.depnt

extractRdf .fullnt

09/02/2015 - 3rd DBpedia Community Meeting – Dublin, Ireland - 9

Module 4: selectRDF

.fullnt

selectRd

f

.nt

• This module enables to select

the triples who has a URI as

subject

• One can also customize this

module according to a topic

to map the predicate to

properties from well-known

vocabularies

09/02/2015 - 3rd DBpedia Community Meeting – Dublin, Ireland - 10

Experiments

We have processed, for one month, the (480) daily

articles from the “Closer” Magazine.

Some statistics:

2800 triples extracted

971 distinct entities

657 distinct predicates

At least 4 triples extracted per articles

Qualitative analysis:

57% of the triples are about relationship between

celebrities (wedding, cheating, rumors, etc.)

43% of the triples are about diverse topics such as sport,

fashion or politics09/02/2015 - 3rd DBpedia Community Meeting – Dublin, Ireland - 11

Conclusion

Good results for two weeks of work (3rd

position on 7 participants for this challenge)

The idea behind this project has been taken by

Orange Labs for being exploited

Possible evolutions:

Automatic mapping of the predicates

Add more grammar rules to get more triples

Improve the performance (slow and long process)

Machine learning algorithm to classify which triple can be

useful (interesting) or not.

09/02/2015 - 3rd DBpedia Community Meeting – Dublin, Ireland - 12