Verbal Multiword Expression Identification:

(1)

Verbal Multiword Expression Identification:

Do We Need a Sledgehammer to Crack a Nut?

Caroline Pasquer ¹ Agata Savary ¹ Carlos Ramisch ² Jean-Yves Antoine ¹

1

University of Tours, LIFAT, France;

²

Aix Marseille University, LIS, France

Multiword expressions (MWEs)

I Definition: Word combinations with idiosyncratic properties (lexical, syntactic, semantic, etc.)

cut corners ‘do an incomplete job’

I MWE identification – challenges:

. Variability

Companies were cutting corners to save costs.

The field would look uneven if corners were cut.

. Literal and coincidental occurrences

Start with cutting one corner of the disinfectant bag.

If you cut along the lines, you get two acute corners.

I Our idea: simple method to identify verbal MWEs (VMWEs) previously seen in training data

PARSEME Shared Task

I Corpora:

. Unified typology for VMWEs in 19 languages:

I

verbal idioms (VID: cut corners ),

I

light-verb constructions (LVC.full: take a walk ),

I

inherently reflexive verbs (IRV: help oneself )

. VMWE + morpho-syntactic annotation

. train/dev/test split

. Seen VMWEs: 74–94% in dev; 50–92% in test

I Shared task on VMWE identification (edition 1.1):

. Systems are given VMWE-annotated train/dev,

. Given test with morphosyntax, predict VMWEs

. Best cross-language macro-average F1 < _0.6

MWEs’ Nature as a Guiding Principle

I MWEs are regular at the level of tokens , but idiosyncratic at the level of types

I Strong lexicalization : the combination of lexemes (not their senses) makes a MWE

. cut corners vs. trim corners , cut edges

I Critical difficulty of unseen data

I Room for improvement on seen MWEs

I Literal readings are extremely rare in practice

Hypotheses

H1 Search co-occurrences of precise lexemes (+POS) annotated as VMWEs in train

H2 Allow only for those morphosyntactic variants which were previously seen in train

H3 Not rely heavily on automatic POS tagging and syntactic parsing, which may be noisy H4 Consider syntactic coherence to eliminate coincidental occurrences

The Seen2020 system

1. Extract from dev (training) or test (testing) previously seen lemma multisets ( H1 ) 2. Filter by applying a combination of:

[f1] Components should be disambiguated by their POS ( H1 , H3 ) [f2] Components should appear in specific orders ( H2 )

[f3] Components and inserted POS should appear in specific orders ( H2 ) [f4] Components should not be too far ( H4 , H3 )

[f5] Closer components are preferred over distant ones ( H4 , H3 ) [f6] Components should be syntactically connected ( H4 )

[f7] Nominal components should appear with a seen inflection ( H2 ) [f8] Nested VMWEs should be annotated as in train

3. Select the optimal filter combination (among 2

⁸

possibilities) using dev Results

System Macro-ave. F1 seen seen

+unseen

SHOMA 0.81 0.64

TRAVERSAL 0.77 0.59 TRAPACC

^S

0.73 0.57 VarIDE 0.61 0.49 Seen2020 0.83 0.67

Seen F1 on test of Seen 2020 vs. 4 shared task systems

Conclusions

I Seen2020’s focus on (variants of) seen data boosts global performances

I Seen2020 outperforms best (neural) systems on seen and seen+unseen data

I Seen2020 is very simple : 8 morphosyntactic on/off filters (vs. millions of real numbers)

I Seen2020’s results are interpretable :

. identify filters responsible for errors → incremental development and customization

. filters reflect known, pervasive and generic linguistic properties of VMWEs

I ^drawbacks : sensitivity to scarcity and quality of VMWE annotations

I perspectives : new languages, out-of-domain generalizability

Focus on filters: example corpus

1. ^{La lumière} ^sera ^faite sur ce drame.

‘The light will be done on this drama’ → ‘Light will be shed on this drama.’

2. ^{La porte} a résolument été fermée aux initiatives.

‘The door was firmly closed on initiatives.’

3. ^Il ferme la porte à une loi. ‘He closes the door on a law.’

4. Dorothée [...] prit la fuite ‘Dorothée took escape’ → ‘Dorothée absconded.’

5. ^Le ^fossé entre les riches et les pauvres se creusait.

‘The gap between the rich and the poor widened.’

train

6. L’enquête n’a pas fait de lumière sur les causes du sinistre.

‘The inquiry shed no light on the disaster’s causes.’

7. Sa lumière la fait briller. ‘Its light makes it shine.’

8. Celui-ci met en lumière le constat fait plus haut.

‘This one puts into light the observation made above.’

9. ^{La lumière} a enfin été faite sur ce drame.

‘The light was finally shed on this drama.’

10. ^Faire une partie de la lumière sur le projet de réforme.

‘To shed part of the light on this reform project.’

11. la lumière nocturne, qui n’a pas fait l’objet [...] que la lumière diurne.

‘nocturnal light, which has not been subject to [...] as diurnal light.’

12. Une enquête a été ouverte pour faire les lumières sur ce drame.

‘An inquiry was launched to shed the lights on this drama.’

13. ^{La lumière} sur son rôle dans cette affaire doit être faite.

‘The light on its role in this case must be shed.’

14. ^{La lumière} ^est ^faite sur ce drame. ‘The light is shed on this drama.’

15. La lumière est faite de diode LED. ‘The light is made of LED diodes.’

test

Focus on filters: selected candidates

Ex. Candidate f1f2f3f4f5f6f7bf8

7 lumière

_noun

la

_pron

fait

_aux

X X X X X X

8 met en lumière

_noun

le

_det

constat

_noun

fait

_verb

X X X X X

9 La

_det

lumière

_noun

a

_aux

enfin

_adv

été

_aux

faite

_verb

X X X X X X X X

10 Faire

_verb

une

_det

partie

_noun

de

_adp

la

_det

lumière

_noun

X X X X X X

11-a La

_det

lumière

_noun

fait

_verb

X X X X X X

11-b [...] lumière

_noun

fait

_verb

l’

_det

[...] X X X X

11-c [...] fait

_verb

[...] la

_det

lumière

_noun

[...] X X X X X

12 [...] faire

_verb

les

_det

lumières

_noun._plur

[...] X X X X X X X

13 La

_det

lumière

_noun

sur

_adp

son

_det

rôle

_noun

X X X X X X

dans

_adp

cette

_det

affaire

_noun

doit

_verb

être

_aux

faite

_verb

14 La

_det

lumière

_noun

est

_aux

faite

_verb

X X X X X X X X

15 La

_det

lumière

_noun

est

_aux

faite

_verb

Verbal Multiword Expression Identification:

Verbal Multiword Expression Identification:

Do We Need a Sledgehammer to Crack a Nut?

Caroline Pasquer 1 Agata Savary 1 Carlos Ramisch 2 Jean-Yves Antoine 1

University of Tours, LIFAT, France;

Aix Marseille University, LIS, France

Multiword expressions (MWEs)

I Definition: Word combinations with idiosyncratic properties (lexical, syntactic, semantic, etc.)

cut corners ‘do an incomplete job’

I MWE identification – challenges:

. Variability

Companies were cutting corners to save costs.

The field would look uneven if corners were cut.

. Literal and coincidental occurrences

Start with cutting one corner of the disinfectant bag.

If you cut along the lines, you get two acute corners.

I Our idea: simple method to identify verbal MWEs (VMWEs) previously seen in training data

PARSEME Shared Task

I Corpora:

. Unified typology for VMWEs in 19 languages:

verbal idioms (VID: cut corners ),

light-verb constructions (LVC.full: take a walk ),

inherently reflexive verbs (IRV: help oneself )

. VMWE + morpho-syntactic annotation

. train/dev/test split

. Seen VMWEs: 74–94% in dev; 50–92% in test

I Shared task on VMWE identification (edition 1.1):

. Systems are given VMWE-annotated train/dev,

. Given test with morphosyntax, predict VMWEs

. Best cross-language macro-average F1 < 0.6

MWEs’ Nature as a Guiding Principle

I MWEs are regular at the level of tokens , but idiosyncratic at the level of types

I Strong lexicalization : the combination of lexemes (not their senses) makes a MWE

. cut corners vs. trim corners , cut edges

I Critical difficulty of unseen data

I Room for improvement on seen MWEs

I Literal readings are extremely rare in practice

Hypotheses

H1 Search co-occurrences of precise lexemes (+POS) annotated as VMWEs in train

H2 Allow only for those morphosyntactic variants which were previously seen in train

H3 Not rely heavily on automatic POS tagging and syntactic parsing, which may be noisy H4 Consider syntactic coherence to eliminate coincidental occurrences

The Seen2020 system

1. Extract from dev (training) or test (testing) previously seen lemma multisets ( H1 ) 2. Filter by applying a combination of:

[f1] Components should be disambiguated by their POS ( H1 , H3 ) [f2] Components should appear in specific orders ( H2 )

[f3] Components and inserted POS should appear in specific orders ( H2 ) [f4] Components should not be too far ( H4 , H3 )

[f5] Closer components are preferred over distant ones ( H4 , H3 ) [f6] Components should be syntactically connected ( H4 )

[f7] Nominal components should appear with a seen inflection ( H2 ) [f8] Nested VMWEs should be annotated as in train

3. Select the optimal filter combination (among 2

possibilities) using dev Results

System Macro-ave. F1 seen seen

+unseen

SHOMA 0.81 0.64

TRAVERSAL 0.77 0.59 TRAPACC

0.73 0.57 VarIDE 0.61 0.49 Seen2020 0.83 0.67

Seen F1 on test of Seen 2020 vs. 4 shared task systems

Conclusions

I Seen2020’s focus on (variants of) seen data boosts global performances

I Seen2020 outperforms best (neural) systems on seen and seen+unseen data

I Seen2020 is very simple : 8 morphosyntactic on/off filters (vs. millions of real numbers)

I Seen2020’s results are interpretable :

. identify filters responsible for errors → incremental development and customization

. filters reflect known, pervasive and generic linguistic properties of VMWEs

I drawbacks : sensitivity to scarcity and quality of VMWE annotations

I perspectives : new languages, out-of-domain generalizability

Focus on filters: example corpus

1. La lumière sera faite sur ce drame.

‘The light will be done on this drama’ → ‘Light will be shed on this drama.’

2. La porte a résolument été fermée aux initiatives.

‘The door was firmly closed on initiatives.’

3. Il ferme la porte à une loi. ‘He closes the door on a law.’

4. Dorothée [...] prit la fuite ‘Dorothée took escape’ → ‘Dorothée absconded.’

5. Le fossé entre les riches et les pauvres se creusait.

‘The gap between the rich and the poor widened.’

train

6. L’enquête n’a pas fait de lumière sur les causes du sinistre.

‘The inquiry shed no light on the disaster’s causes.’

7. Sa lumière la fait briller. ‘Its light makes it shine.’

8. Celui-ci met en lumière le constat fait plus haut.

‘This one puts into light the observation made above.’

9. La lumière a enfin été faite sur ce drame.

Caroline Pasquer ¹ Agata Savary ¹ Carlos Ramisch ² Jean-Yves Antoine ¹

. Best cross-language macro-average F1 < _0.6

I ^drawbacks : sensitivity to scarcity and quality of VMWE annotations

1. ^{La lumière} ^sera ^faite sur ce drame.

2. ^{La porte} a résolument été fermée aux initiatives.

3. ^Il ferme la porte à une loi. ‘He closes the door on a law.’

5. ^Le ^fossé entre les riches et les pauvres se creusait.

9. ^{La lumière} a enfin été faite sur ce drame.

10. ^Faire une partie de la lumière sur le projet de réforme.

13. ^{La lumière} sur son rôle dans cette affaire doit être faite.

14. ^{La lumière} ^est ^faite sur ce drame. ‘The light is shed on this drama.’