Automated WordNet Construction Using Word Embeddings · (WOLF) [4] Universal Wordnet [5] Extended...

Preview:

Citation preview

(pistol)

(shooting)

(to shoot)

(species)

(crossbow)(kind)

(family)

(plant)

(coriander)

(cilantro)

(celery)

(garlic)

Isometric mapping of sense-clustersfound for "лук"("bow", "onion")

Automated WordNet ConstructionUsing Word Embeddings

Mikhail Khodak, Andrej Risteski, Christiane Fellbaum, Sanjeev AroraComputer Science Department, Princeton University

assign a scoreto each

candidate synset

flagstone

flag

slab

TargetWord

get translations of target word using a

bilingual dictionary or machine translation (MT)

get set of candidate synsetsby querying PrincetonWordNet (PWN) using

translations of

Translations

MT + PWN SynsetScoring

ThresholdMatching

return all synsetswith score abovea threshold

Candidates SynsetScore

Threshold 0.41

flag.n.01

flag.n.04

flag.n.06

flag.n.07

iris.n.01

masthead.n.01

pin.n.08

slab.n.01

score = 0.280

score = 0.222

score = 0.360

score = 0.161

score = 0.200

score = 0.195

score = 0.251

score = 0.521

Contributions:

Wordnet Libre du Français (WOLF) [4]

Universal Wordnet [5]

Extended Open Multilingual

Wordnet [6]

Synset Representation

Synset Representation

+ Sense Clusters

F-Score of Synset Matching for French and Russian

Isometric mapping of sense-clusters found for "fox"

[1] Fellbaum, MIT Press.[2] Arora et al., ICLR 2017.[3] Arora et al., arXiv 2016.

[4] Sagot and Fišer, LREC 2008.[5] de Melo and Weikum, CIKM 2009.[6] Bond and Foster, ACL 2013.

PrincetonWordNet

French-EnglishDictionary

French word

'dalle'

��������

correct synsets:flag.n.06 -slab.n.01 �

retrieved synsets:slab.n.01 �

Synset Representation:

Sense Clusters:

Recommended