A Robust Approach to Aligning Heterogeneous Lexical Resources

A Robust Approach to Aligning Heterogeneous Lexical Resources

Mohammad Taher PilehvarRoberto Navigli

MultiJEDI ERC 259234

Lexical Resource

WordNet

BabelNet

UBY

Lexical Resource

WordNet

BabelNet

UBY

Lexical Resource

WordNet

BabelNet

UBY

Improved word and concept coveragee.g., named entities, new senses

Improved domain coverage

Improved multilingualitydozens of new languages

Expert-made relations preservede.g., Hypernymy, meronymy, etc.

Why combine resources?

Provides complementary knowledge

Applications:

Semantic Parsing Shi and Mihalcea, 2005

Semantic Role LabelingPalmer et al., 2010

WSD and entity linkingMoro et al., TACL 2014

Why combine resources?

Difficulty of resource alignment

Fine granularity of lexical resources

plant

WordNet

4 senses 15 senses

How resource alignment works?

WordNetplant#n#1plant#n#1

• Usually measures the similarity of two concepts

WKT: WN:


• Usually measures the similarity of two concepts

• And aligns two concepts if their similarity exceeds a certain threshold

?WordNetplant#n#1plant#n#1WKT: WN:


Alignment approaches differ in the way they calculate this similarity

?WordNetplant#n#1plant#n#1WKT: WN:

Gloss similarity WordNet

Denfinitional similarity

• Strong baseline

• Fall short when• Different wordings are used for same concepts• When two words lack quality glosses

plant -- Buildings for carrying on industrial labor.

plant -- The necessary infrastructure used in support and maintenance of a given facility.


Contributions

A novel concept similarity measure


A robust technique foralignment of resourcesA robust technique for

alignment of heterogeneous resources

An effective ontologization approach

Our approach: SemAlign


Definition similarity


Structural similarity

SemAlign: structural similarity

WordNet

SemAlign: core

Modeling concepts through Semantic Signatures

conceptsome

Personalized PageRankSemantic Signature of a concept concept

Distributional representationover all concepts in the semantic network

Personalized PageRankSemantic Signature of a concept concept

SemAlign

Definition similarity

WordNet

SemAlign

WordNet

SemAlign: signature unification

conc

ept_4

conc

ept_5

conc

ept_2

conc

ept_6

conc

ept_1

conc

ept_3

conc

ept_4

conc

ept_5

conc

ept_2

conc

ept_6

conc

ept_1

conc

ept_3

WordNet

SemAlign: signature unificationFind concepts associated with monosemous words

conc

ept_4

conc

ept_5

conc

ept_2

conc

ept_6

conc

ept_1

conc

ept_3

conc

ept_4

conc

ept_5

conc

ept_2

conc

ept_6

conc

ept_1

conc

ept_3

WordNet

SemAlign: signature unificationTruncate vectors to the overlapping concepts

SemAlign: signature unification

WordNet Synsets containing at least one monosemous word

~ 60% (72,000)

The reliability of leveraging monosemous words

Semantic Signature Comparison

Weighted Overlap(Pilehvar et al., ACL 2013)

Semantic Signature Comparison

SemAlign: score combination

Ontologizationof lexical resources

WordNet


WordNet




WordNet


WordNet




Definition page for sail

WordNet


WordNet



WordNet


Definition page for windmill

1. A machine which translates linear motion of wind to rotational motion by means of adjustable vanes called sails.




windmill.n.1linear motion.n.1

rotational motion.n.1translate.v.3

wind.n.1

machine.n.1sail.n.3

adjustable vanes.v.4


Ontologization: similarity-based disambiguation




1. A trip in a boat, especially a sailboat.

2. A sailing vessel; a vessel of any kind; a craft.

3. The blade of a windmill.

4. A tower-like structure found on the dorsal (topside) surface of submarines.




























Ontologization:evaluation

Our methodWKTWSD

0.7

0.8

0.9

F1

Acc

(Meyer & Gurevych, 2012)

Alignment Experiments:Datasets

(Matuschek & Gurevych, TACL 2013)

WordNetWN synsetsmanually mapped to their corresponding concepts

320

484315

Alignment Experiments:System configurations

Parameter 1:

score combination parameter

𝜷𝜷

𝜷


Parameter 2:

similarity threshold 𝜽


𝜽 𝜷

0 1

0.5

Setting and

a. Unsupervised system


𝜽 𝜷

0 1


Setting and

b. Tuning on subset

0.42 0.53


𝜽 𝜷


Setting and

b. Tuning on subsetc. Cross validation

0 1

0.44 0.51

Comparison systemDijkstra-WSA

(Matuschek & Gurevych, TACL 2013)

Alignment Experiments

WN-WP WN-WT WN-OW0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.805

0.712 0.709

0.833

0.73 0.733

0.84

0.722

0.749

Unsupervised Tuning on subset Cross-validation

SB

SB+DWSAF1



0.65

0.7

0.75

0.8

0.850.824

0.736

0.824

0.723

0.684 0.684

Tuning on WN-WP Tuning on WN-WT Tuning on WN-OW

F1SB

SB+DWSA

SemAlign: structural similarity



0.4

0.5

0.6

0.7

0.8

0.9

0.71

0.39

0.473

0.83

0.623 0.627

DWSA SemAlign

F1

Conclusions

• A novel approach for aligning lexical resources– Accurate even in the absence of training data

– Robust across different resources

• An effective ontologization approach

• Experiments on aligning – WN to WP, WT, and OW

Future directions

• Integrating the approach into BabelNet for boosting the alignment accuracy

• Alignment across different languages

• Updating lexicons with novel terms

1. thanks -- an acknowledgment of appreciation

2. thanks -- with the help of or owing to;

2. thanks -- an expression of gratitude.

2. thanks -- because of; normally used with a positive connotation, though it can be used sarcastically.

WordNet

Thanks!

Documents

A Robust Approach to Aligning Heterogeneous Lexical Resources