The MultiJEDI ERC Project: Multilingual Joint Word Sense - - PowerPoint PPT Presentation

Roberto Navigli The MultiJEDI ERC Project: Multilingual Joint Word Sense Disambiguation

5 July 2016 – META-FORUM 2016 http://lcl.uniroma1.it

ing

Simone Ponzetto Tiziano Flati Andrea Moro Daniele Vannella Taher Pilehvar Francesco Cecconi

Federico Scozzafava Alessandro Raganato Ignacio Iacobacci José Camacho Collados Claudio Delli Bovi

You may say I'm a dreamer, but I am not the only one. I hope someday you'll join us. And the world will be as one!

• John Lennon

A 5-year ERC Starting Grant (2011-2016)

• n Multilingual Word Sense Disambiguation

http://multijedi.org

INTEGRATING KNOWLEDGE

[Navigli & Ponzetto, ACL 2010; Pilehvar & Navigli, ACL 2014]

The resource diaspora

Key Objective 1: create knowledge for all languages Multilingual Joint Word Sense Disambiguation (MultiJEDI)

WordNet MultiWordNet WOLF MCR GermaNet BalkaNet

• We collect lexicalizations, definitions, translations,

images, etc. from each of the merged resources Merging entries from different resources into BabelNet

WordNet

What is BabelNet?

• A merger of resources of different kinds:

• A merger of resources of different kinds:

– WordNet: the most popular computational lexicon of English – Open Multilingual WordNet: a collection of open wordnets – WoNeF: a French WordNet – Wikipedia: the largest collaborative encyclopedia – Wikidata: the largest collaborative knowledge base – Wiktionary: the largest collaborative dictionary – OmegaWiki: a medium-size collaborative multilingual dictionary – GeoNames: a worldwide geographical database – Microsoft Terminology: a computer science thesaurus – High-quality automatic sense-based translations

What is BabelNet?

Why do we need BabelNet?

• Multilinguality: the same concept is expressed in tens of

languages

Why do we need BabelNet?

• Multilinguality: the same concept is expressed in tens of

languages

Why do we need BabelNet?

• Multilinguality: the same concept is expressed in tens of

languages

Why do we need BabelNet?

• Multilinguality: the same concept is expressed in tens of

languages

• Coverage: 271 languages and 14 million entries!

– 6M concepts and 7.7M named entities – 119M word senses – 378M semantic relations (27 relations per concept on avg.) – 11M images associated with concepts – 41M textual definitions – 2M concepts with domains associated

Why do we need BabelNet?

• Multilinguality: the same concept is expressed in tens of

languages

• Coverage: 271 languages and 14 million entries!
• Concepts and named entities together: dictionary and

encyclopedic knowledge is semantically interconnected

Why do we need BabelNet?

• Multilinguality: the same concept is expressed in tens of

languages

• Coverage: 271 languages and 14 million entries!
• Concepts and named entities together: dictionary and

encyclopedic knowledge is semantically interconnected

• "Dictionary of the future": semantic network structure

with labeled relations, pictures, multilingual synsets

Why do we need BabelNet?

• Multilinguality: the same concept is expressed in tens of

languages

• Coverage: 271 languages and 14 million entries!
• Concepts and named entities together: dictionary and

encyclopedic knowledge is semantically interconnected

• "Dictionary of the future": semantic network structure

with labeled relations, pictures, multilingual synsets

• Full-fledged taxonomy: is-a relations are available for

both concepts and named entities (Wikipedia Bitaxonomy)

– Ferrari Testarossa is-a sports car – BabelNet is-a semantic network & encyclopedic dictionary

Why do we need BabelNet?

• Multilinguality: the same concept is expressed in tens of

languages

• Coverage: 272 languages and 14 million entries!
• Concepts and named entities together: dictionary and

encyclopedic knowledge is semantically interconnected

• "Dictionary of the future": semantic network structure

with labeled relations, pictures, multilingual synsets

• Full-fledged taxonomy: is-a relations are available for

both concepts and named entities (Wikipedia Bitaxonomy)

• Easy access: Java and HTTP RESTful APIs; SPARQL

endpoint (2 billion triples); downloadable indices for research purposes

The core of the Linguistic Linked Open Data cloud!

What can we do with BabelNet?

• Search and translate:

What can we do with BabelNet?

What can we do with BabelNet?

• Explore the network:

WordNet-Wikipedia mapping accuracy

• Quality lower bound of the mapping: 87%

– On the 6000 lowest-confidence mappings – Note: this concerns only 50k synsets in the intersection

Creating Datasets with BabelNet: all in one!

• Annotating with BabelNet implies annotating with

WordNet, Wikipedia, OmegaWiki, Open Multilingual WordNet, Wikidata and Wiktionary

Key fact!

BabelNet

7

ADDRESSING AMBIGUITY

[Moro, Raganato & Navigli, TACL 2014]

Motivation (1): hungry computers

• EN - The mouse ate the cheese

Motivation (1): hungry computers

• EN - The mouse ate the cheese
• FR - La souris a mangé le fromage.

Motivation (1): hungry computers

• EN - The mouse ate the cheese
• FR - La souris a mangé le fromage.
• IT - Il mouse ha mangiato il formaggio

Multilingual Joint Word Sense Disambiguation (MultiJEDI) Key Objective 2: use all languages to disambiguate one

So what?

• The first (and only) system that performs

Word Sense Disambiguation (common nouns, verbs, adjectives) and Entity Linking together

• In arbitrary languages (270+ languages)
• In multiple languages at once

Step 4: Select the most reliable meanings

“Thomas and Mario are strikers playing in Munich”

Thomas (novel) Seth Thomas Thomas Müller Mario Gómez Mario (Album) Mario (Character) Striker (Movie) Striker (Video Game) striker (Sport) Munich (City) FC Bayern Munich Munich (Song)

Experimental Results: Fine-grained (Multilingual) Disambiguation

Senseval-3 SemEval-2007 task 17 SemEval-2013 task 12

Experimental Results: KORE50, AIDA-CoNLL

• Two gold-standard Entity Linking datasets:

Babelfy "understands" 'the mouse ate the cheese'!

WSD and Entity Linking together win!

The Crazy Polyglot!

Live demo (2) – Crazy polyglot! EN In todayʼs knowledge and information society FR le paysage lexicographique est plus hétérogène que jamais. IT Possono le risorse stand-alone competere ES con múltiples funciones, portale lexicográficas multilingüe y servicios web, ZH Web服，定 制 的 喜 好 和 个 人 用 的 个 人 料 ？

BabelNet 3.6 is now a knowledge base!

• Semantic relations from Wikidata + Infoboxes

(superset of DBpedia) + relations extracted with Open Information Extraction techniques

SENSE AND CONCEPT REPRESENTATIONS

[Iacobacci et al., ACL 2015; Camacho-Collados et al., NAACL+ACL 2015]

Latent representation of word senses: SensEmbed

Iacobacci, Pilehvar, Navigli (ACL 2015)

Problem: word representations cannot capture polysemy

Problem: word representations cannot capture polysemy

Problem: word representations cannot capture polysemy

Our solution: distinct representation for each word’s meaning

Embeddings + Semantic Knowledge = SensEmbed

• Achieved by training word2vec with text disambiguated

with Babelfy with high precision, low recall

Explicit representation of concepts: NASARI

Camacho Collados, Pilehvar and Navigli NAACL 2015 + ACL 2015 + Artificial Intelligence Journal 2016

Motivation

NASARI: human-interpretable semantic vectors

• A vector represents the meaning of Babel synset
• Its components are

– words – Babel synsets (concepts and named entities)

NASARI: human-interpretable semantic vectors

• The semantic part of the vector enables cross-lingual

semantic alignments and comparison of text

Semantic similarity: results

• NASARI is the concatenation of lexical and unified

vectors:

Cross-lingual Word similarity: Results

Spearman (ρ) and Pearson (r) correlation performance of different systems on multilingual editions of the RG-65 datasets.

Babelscape: bringing our multilingual technologies to the market

• The MultiJEDI project is now over
• However, much work still to be done in this direction

– BabelNet live – Increase coverage

• Babelscape brings our multilingual technologies to the

market

• Key objective: making BabelNet and all other HLT

sustainable

• BabelNet will always be available (and downloadable)

for research purposes

Summarizing…

+ latent and explicit representations of meanings + sustainability plan for improving our systems over time

Thanks or…

m i

(grazie)

Roberto Navigli

Linguistic Computing Laboratory http://lcl.uniroma1.it

@RNavigli