Exploiting multilingual lexical resources to predict the - - PowerPoint PPT Presentation

Exploiting multilingual lexical resources to predict the compositionality of MWEs

Paul Cook University of New Brunswick

Compositionality

• Many MWEs exhibit semantic idiomaticity
• Compositionality: The extent to which the meaning of an MWE is

predictable from the meanings of its components

• Compositionality lies on a continuum
• Compositionality predictions: binary (Bannard et al., 2003), multi-

way (Fazly and Stevenson 2007), continuous (Reddy et al. 2011)

2

Compositionality

• An MWE component word is compositional if its

meaning is reflected in the meaning of the expression

• Compositionality predictions: MWE as a whole, and

individual component words (Bannard et al., 2003; Reddy et al. 2011)

3

In this talk…

• Compositionality prediction for many languages

and many kinds of MWEs via a multilingual lexical resource

• Token-level MWE identification is important for type-

level compositionality prediction

• Compositionality scores are not the end of the

story: The case of English VPCs

4

Compositionality prediction: String similarity

String similarity

• Compositionality prediction based on string similarity of MWE and

component words, under translation (Salehi and Cook, 2013)

• Applicable to many types of MWEs in many languages
• Does not require
• Language or construction specific properties (e.g., Fazly et al.,

2009)

• A parallel corpus (e.g., de Caseli et al., 2010; Salehi et al.,

2012)

• A monolingual corpus (e.g., Lin 1999; Reddy et al., 2011)
• Requires a multilingual dictionary

6

kick the bucket kick the bucket mord zad

• satl

Source language Target language Translation Source Target

make a decision make a decision tasmim gereftan sakht yek tasmim

Source language Target language Translation

Motivation

7

public service public service khadamaat omumi

• mumi

khedmat

Source language Target language Translation

Motivation

8

Translate Panlex khadamaat omumi

• mumi

khedmat Compare (LCS, LEV1, LEV2, SW) s1 Compare (LCS, LEV1, LEV2, SW) s2

Approach (1)

9

Translations

• Translate into 54 languages! (not just one)
• Panlex
• Free online translation resource
• Combines many translation dictionaries
• 20M lemmas; 9k language varieties; 1.1B

translations

• Select top-10 languages via cross-validation

10

Mean Mean public service vs. public public service vs. service Best 10 languages scores

2 1

) 1 ( s s α α − +

Compositionality score

Approach (2)

11

Results: ENC

12

Method Correlation (r) Reddy et al. (2011) 0.714 String similarity 0.649 String similarity: Best single language 0.497 String similarity + Reddy et al. 0.742

• 90 English noun compounds (ENC, Reddy et al., 2011)
• Nested 10-fold cross-validation

Results: EVPC

13

Method Accuracy Bannard et al. (2003) 0.600 String similarity 0.693

• 160 English verb-particle constructions (EVPC, Bannard, 2006)
• Binary compositionality prediction for verb component

Results: GNC

14

Method Correlation (r) String similarity 0.372 String similarity: Best single language 0.320 Schulte im Walde et al. (2013) 0.450

246 German noun compounds (GNC, von der Heide and Borgwaldt, 2009; Schulte im Walde et al. 2013)

Most-selected languages

ENC EVPC (verb) GNC Language Family Language Family Language Family Czech Slavic Basque Basque Polish Slavic Norwegian Germanic Lithuanian Baltic Lithuanian Baltic Portuguese Romance Slovenian Slavic Finnish Uralic

15

Compositionality prediction: Distributional similarity

Compositionality via distributional similarity

die kick the bucket kick bucket (Katz and Giesbrecht, 2006; Reddy et al., 2011)

17

Compositionality via distributional similarity

die kick the bucket kick bucket pail kick the pail (Katz and Giesbrecht, 2006; Reddy et al., 2011)

18

Distributional similarity of multi-way translations

• Compositionality based on distributional similarity

under translation into many languages (Salehi, Cook and Baldwin, 2014)

• Still applicable to many languages and kinds of

MWEs

• Additional requirement: Many monolingual corpora

19

Approach (1)

20

Mean Mean public service vs. public public service vs. service Best N languages scores

2 1

) 1 ( s s α α − +

Compositionality score

Approach (2)

21

Results

22

Method ENC EVPC GNC

• 1. Source language

0.700 0.177 0.141

• 2. Best-N target languages

0.434 0.398 0.113

• 3. 1 + 2

0.725 0.312 0.178

• 4. 1 + 2 + String similarity

0.732 0.417 0.364

Correlation (r) on each dataset

Compositionality prediction and token-level MWE identification

MWE identification: VPC vs V+PP

• VPCs occur in two configurations and are ambiguous

with V+PPs:

• 1. Look up the number (VPC: joined)
• 2. Look the number up (VPC: split)
• 3. Look up the chimney (V+PP)
• Token-level MWE identification strategy: Full-token n-

gram match

• For EVPC this led to poor performance of English

distributional similarity

24

MWE identification: MWEs vs literal combinations

• Expressions can be ambiguous between MWEs and literal

combinations

• 1. I think Paula might hit the roof if you start ironing
• 2. When the blood hit the roof of the car I realised it was serious
• Many verb-noun idiomatic combinations are primarily used literally!

(Fazly et al., 2009)

• blow (the) whistle (65%); pull (one’s) leg (78%); see star(s) (92%)
• Type-level compositionality prediction based on distributional similarity

can be influenced by the (possibly predominant!) literal usages of an expression

25

Beyond compositionality predictions

Limitations

• Compositionality predictions don’t indicate which meaning of a

component word is contributed

• Kangaroo court
• Court is mostly compositional (4.4/5, Reddy et al., 2011)
• Place for legal trials? Area for playing sports?
• Stir up
• 16% of annotators judged up to be entailed (Bannard, 2006)
• Particles are highly polysemous
• Is some other meaning contributed?

27

Cognitive grammar

• Cognitive linguistics: Many so-called idiomatic expressions in

fact draw on meaning of component words

• Cognitive grammar: Represent non-spatial concepts as spatial

relations

• Trajector (TR): Object that is conceptually foregrounded
• Landmark (LM): Object against which TR is foregrounded
• Schema: Abstract conceptualization of TR and LM in some initial

and final configuration, as communicated by an expression

• Lindner (1981) identifies 4 senses for up in VPCs

28

Vertical up

• TR moves away from LM in direction of

increase along a vertically-oriented axis

• Prototypical upward movement:

The balloon floated up

• Movement along an abstract vertical

axis: The price of gas jumped up

• Metaphorical extensions: Up as a path

into…

• Perceptual field: show up, spring up
• Mental field: dream up, think up
• State of activity: get up, start up

29

Goal-oriented up

• TR approaches a goal LM;

movement is not necessarily vertical

• Prototypical examples:
• The bus drew up to the stop
• He walked up to the bar
• Metaphorical extensions:
• Social domain: The intern

kissed up to his boss

• Domain of time: The deadline

is coming up quickly

30

Completive up

• A sub-sense of goal-oriented up; shares its schema
• LM represents an action being done to completion
• Corresponds to up as an aspectual marker
• Examples:
• Clean up your room!
• Penelope drank up all her milk
• I filled the car up

31

Reflexive up

• Also a sub-sense of goal-
• riented up
• Sub-parts of TR approach

each other

• Examples:
• The CEO bottled up her

anger until she burst

• He crumpled up the

piece of paper

• Tie up your skates!

32

Schematic network for up

33

Schemas vs Compositionality

• Compared annotations from Bannard (2006) and Cook and

Stevenson (2006)

• Particles judged compositional correspond to vertical up
• E.g., spring up, stay up
• Non-compositional particles include all 4 senses, e.g.,
• Vertical: speed up
• Goal-oriented: back up
• Completive: beat up
• Reflexive: roll up

34

Experiments

• Type-level classification of up VPCs by sense (Cook and

Stevenson, 2006)

• Supervised learning approach: SVM
• Features
• Verb: Relative frequencies of syntactic slots: subject,

direct object, indirect object, object of preposition

• Particle:
• Frequency of split vs joined construction
• Frequency of verb with other particles

35

Results

Features % accuracy 3-way 2-way Baseline 33 50 Verb 51 67 Particle 33 47 Verb + Particle 54 63

• 3-way: Merge goal-oriented and completive up
• 2-way: Vertical up vs rest
• Dataset: 180 up VPCs annotated for sense evenly split

into train/dev/test sets

36

Summary

• Compositionality predictions via a multilingual

lexical resource are applicable to many languages and kinds of MWEs

• MWE identification is important for compositionality

prediction based on distributional similarity

• Cognitive grammar provides an alternative way to

describe the semantics of English VPCs

37

Thanks

• Thank you to my collaborators on this work:

Timothy Baldwin, Afsaneh Fazly, Bahar Salehi, and Suzanne Stevenson

• Thank you to the Natural Sciences and Engineering

Research Council of Canada (NSERC), NICTA, the University of Toronto, and The University of Melbourne for funding this research

38