La syntaxe des expressions polylexicales: codage lexical, - - PowerPoint PPT Presentation

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

La syntaxe des expressions polylexicales: codage lexical, annotation et flexibilité syntaxique

Agata Savary

Université de Tours

Linglunch Paris Diderot, 12 avril 2018, Paris

1/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Multiword expressions

Word combinations, which exhibit lexical, syntactic, semantic, pragmatic and/or statistical irregularities. Examples: all of a sudden, a hot dog, to pay a visit, to pull

• ne’s leg

Encompass heterogeneous objects: idioms, compounds, light verb constructions, rhetorical figures, institutionalised phrases or named entities Pervasive feature: non-compositional semantics - the meaning

• f an MWE cannot be deduced from the meanings of its

components, and from its syntactic structure, in a way deemed regular for the given language. Varying degree of syntactic variability (flexibility), especially in verbal MWEs (VMWEs).

2/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Defective and restrictive properties of MWEs [Lichte et al.(2018)]

Defective properties exclude a literal interpretation of the MWE: Defective agreement: a cross-roads Defective syntactic structure: j’ai beau chercher ‘I have beautiful search.inf’⇒‘I search in vain’ Restrictive properties reduce the number of possible surface realizations

• f the MWE with respect to the literal reading, e.g.:

Restrictive lexical selection: to make ends meet vs. #to make edges come together Restrictive agreement: I cross my fingers vs. #I cross his fingers Restrictive diathesis: the die is cast vs. #one casts the die Restrictive modification: he lives a life of luxury vs. #he lives a life

3/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Scale-wise regularity of properties [Lichte et al.(2018)]

More regular (≻) = admitted by more objects (in a set) sample set: English Subj-Verb-Obj expressions (John pulled the door) “allow any head verb” ≻ “allow only the verb kick” “allow passive” ≻ “prohibt passive” “allow a possessive determiner” John pushed the/my door ≻ “impose a possessive determiner” John broke his/our fall ‘John made his/our fall less forceful’ ≻ “impose a possessive agreeing with Subj” John crossed his fingers ‘John hoped for good luck’ John held his tongue ‘John refrained from expressing his view’ Idiosyncratic = irregular (shared by no two objects) Usually only the restrictive lexical selection is truly idiosyncratic (except in polysemous MWEs: go on ‘continue/happen’)

4/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Scale-wise regularity of VMWEs [Gross(1986), Gross(1988)]

N0V (DetN)1 expression Free subject Free verb Free object Verb reduction Verb inflection Noun inflection Noun modification Passive

• Det. alternation

N prend la pomme ‘N takes an apple’ N prend une décision ‘N takes a decision’⇒‘N makes a decision’

• N tourne la page

‘N turns the page’⇒‘N stops dealing with sth.’

• ?

? N prend la porte ‘N takes the door’⇒‘N leaves (because forced)’

• 5/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Lexical encoding of MWE variability/regularity – SOA

Axis 1: Formalization of the lexicon-grammar interaction generic but insufficiently formalized MWE lexicons [Gross(1986), Mel’čuk et al.(1988), Grégoire(2010), Przepiórkowski et al.(2014), McShane et al.(2015)] formalized but grammar-bound MWE lexicons: HPSG [Sag et al.(2002), Copestake et al.(2002), Villavicencio et al.(2004), Bond et al.(2015), Herzig Sheinfux et al.(2015)], LFG [Attia(2006)], TAG [Abeillé and Schabes(1989), Abeillé and Schabes(1996), Vaidya et al.(2014), Lichte and Kallmeyer(2016)] Axis 2: Existence of factorization mechanisms no generalization of the MWE behavior [Al-Haj et al.(2014)] shallow generalization (limited degree of recursiveness) [Savary(2009), Grégoire(2010), Przepiórkowski et al.(2014), Herzig Sheinfux et al.(2015)] CFG-like metagrammar with FSs and unification for continuous MWEs [Jacquemin(2001)]

6/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Lexical encoding of MWE variability/regularity – challenges

Account for the scale-wise (ir)regularity of a MWE, while avoiding redundancy. Offer a perfectly formalized lexicon-grammar interface. Use a (relatively) theory-independent framework to mutualize VMWE lexicons.

7/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

XMG [Crabbé et al.(2013), Petitjean et al.(2016)]

a language

• bject-oriented – objects, classes, inheritance

declarative – grammaticality is defined in terms of constraints rather than procedures notationally expressive - modularity, inheritance, conjunction/disjunction of tree fragments, namespaces extensible to new dimensions (semantics, frames etc.), formalisms (IG, etc.), linguistic principles (e.g. clitic ordering) a metagrammar compiler (for each target language, here FS-LTAG) – constraint solver: produces minimal tree models respecting the constraints

8/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

FrenchTAG – French XMG metagrammar [Crabbé et al.(2013)]

XMG implementation of the syntactic TAG grammar of French by [Abeillé(2002)] 285 XMG classes, 96 families (classes assigned to lexemes), compiled into 9045 TAG trees toy lexicon of 555 lexemes, including 248 verbs Example Jean prend la porte ‘John takes the door’⇒‘John leaves because he is forced to’ XMG covers literal readings (by compositionality) XMG does not cover idiomatic readings

9/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Morphology (simplified)

10/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Lemmas

Trivial classes

propename → N⋄ noun → N⋄ CliticT → CL⋄ stddeterminer → N D⋄ N∗

11/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

From metagramar to parsing: n0Vn1 (Jean prend la porte)

Metagrammar tree fragments inherited by n0Vn1

CanonicalSubject → S N↓ VN activeVerbMorphology → S VN V⋄ CanonicalObject → S VN N↓

Grammar tree Derivation tree Derived tree

12/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

From metagramar to parsing: n0Vn1 (Jean la prend)

Metagrammar tree fragments inherited by n0Vn1

CanonicalSubject → S N↓ VN activeVerbMorphology → S VN V⋄ CliticObject → S CL↓ VN . . .

Grammar tree Derivation tree Derived tree

13/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

XMG classes

14/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Class hierarchy

TopLevelClass conjunction of classes disjunction of classes VerbalArgument CanonicalArgument SubjectAgreement NonInvertedNominalSubject Clitic RealizedNonExtractedSubject CanonicalNonSubjectArg NonReflexiveClitic VerbalMorphology . . . CliticSubject CanonicalSubject CanonicalObject CliticObject3 . . . ActiveVerbMorphology Subject Object dian0Vn1Active . . . dian0Vn1Passive . . . dian0Vn1ShortPassive n0Vn1 15/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Adding MWEs to the metagrammar [Savary et al.(sub)]

Strategy add lexical entries for MWEs with co-anchors, use interface filters to express restrictive properties, reuse existing tree fragments for the (more) regular properties, decorate them with interface features, create new tree fragments for defective properties and for lexicalized arguments of various syntactic structures

16/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

MWE lemmas with co-anchors and filters

17/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Previous tree fragments decorated with interface features

18/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

New XMG classes for lexicalized arguments

19/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

From metagramar to parsing: mwen0Vn1 (Jean prend la porte)

Tree fragments inherited by mwen0Vn1

CanonicalSubject → S N↓ VN activeVerbMorphology → S VN V⋄ mweCanonicalObjectLex → S VN N

(no ↓)

mweDetNoun → N D⋄ N⋄

Grammar tree Derivation tree Derived tree

20/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Modified class hierarchy

TopLevelClass conjunction of classes disjunction of classes VerbalArgument CanonicalArgument SubjectAgreement NonInvertedNominalSubject Clitic RealizedNonExtractedSubject CanonicalNonSubjectArg NonReflexiveClitic mweCanonicalObject . . . . . . CliticSubject CanonicalSubject . . . mweObjectLex mweLexDetLexNoun . . . CanonicalObject CliticObject3 . . . VerbalMorphology Subject mweSubjectLexStruct mweObjectLexStruct Object ActiveVerbMorphology mweSubject mweObject mwedian0Vn1Active . . . mwedian0Vn1Passive . . . mwedian0Vn1ShortPassive mwen0Vn1

21/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Evaluation

Corpus French PARSEME corpus [Candito et al.(2017)] Selection of 14 VMWEs (frequent) and 52 occurrences (large syntactic variety). Simplification: (i) no subordinate sentences and coordinations, (ii) few non-lexicalized arguments/modifiers. TRAIN (26 occ.), TEST (26 occ.).

FrenchTAG mweFrenchTAG with MWEs from DEV-S DEV-S + TEST-S classes 285 337 (+18%) 341 (+1.1%) MWE lemmas 5 31 37

Proof of concept Non-redundant lexical encoding of MWEs can be effectively achieved in an object-oriented metagrammar-based approach

22/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Conclusions

Tensions around MWEs Variability vs. fixedness (grammatical vs. lexical encoding) Formal-grammar integration vs. theory-independence Denotational precision vs. multilingualism Solutions Object-oriented modeling - to express degrees of variability Metagrammar - (relative) theory-independence + compilation into a precision grammar for one language Language-independent MWE variability measure (App A) Iterative unified annotation methodology development in a multilingual network (App. B)

23/23

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

(Language-independent) syntactic and linear similarity

[Pasquer et al.(2018)]

Ils ne prennent vraiment pas une bonne décision They neg take really.adv neg.adv a.det good.adj decision

nsubj adv adv adv det amod

• bj

root

Voici les sages décisions que Jean a aussitôt prises Here the wise decisions that.pron John.propn has.aux at once.adv taken

det amod

• bj
• bj

nsubj aux adv acl:relcl root

Syntactic similarity of components Sørensen–Dice coefficient: S(O1, O2) = 2 × |P(O1) ∩ P(O2)|/(|P(O1)| + |P(O2)|) On outgoing dependencies, SS(décision, décisions) =

2×|{det,amod}| |{det,amod,acl:relcl}|+|{det,amod}| = 4 5

On inserted POS. Similarity of VMWEs Syntactic: weighted average of per-component similarity. Linear: Sørensen–Dice coefficient on inserted POS. SL(V1, V2) =

2×|{adv}| |{adv,det,adj}|+|{pron,propn,aux,adv}| = 2 7

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Syntactic and linear variability [Pasquer et al.(2018)]

VMWE variability Rigidity of a VMWE: average of pairwise similarities of all variants Variability of a VMWE: inverse of rigidity Applications Linear variability is positively correlated with a linguistic variability benchmark in French [Tutin(2016)] Linear variability discriminates LVCs from VIDs. Linear variability discriminates idiomatic from literal readings Syntactic and linear similarity features are useful in VMWE variant identification [Pasquer et al.(sub)]

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

VMWEs annotation in the PARSEME scientific network

Methodology 22 languages (6 non Indo-European), Unified annotation guidelines as decision trees driven by linguistic tests (with examples in many languages), Universal categories, room for language-specific categories and tests, Close links with Universal Dependencies treebanking.

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

VMWE variability in the PARSEME guidelines

Prototypical form: head verb is in active voice, finite form; other lexicalized components depend either on the verb or on another lexicalized component. elle prend une décision Meaning-preserving variants: analytical tenses: elle a pris une décision relative clauses: la décision qu’elle prend non-finite clauses: la décision prise, en prenant une décision diathesis alternation: la décision sera prise interposed modifiers: prendre une série de décisions Canonical form: prototypical or most neutral form keeping the idiomatic reading elle prend une décision les carottes sont cuites Variant neutralization during annotation: Texts contain VMWE variants, Linguistic tests apply to the canonical form.

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

VMWE typology (v. 1.1)

Universal categories (valid for all languages): light verb constructions (LVCs) LVC.full: to give a lecture LVC.cause: to grant rights verbal idioms (VIDs) to call it a day Quasi-universal categories (valid for many languages): inherently reflexive verbs (IRVs) (FR) s’évanouir ‘to faint’ verb-particle constructions (VPCs) VPC.full to do in ‘to kill’ VPC.semi to eat up ‘to eat completely’ multi-verb constructions (MVCs) to let go Experimental (optional) category inherently adpositional verbs (IAVs) to come across sth/sb, to rely on sth/sb

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

PARSEME VMWE corpus and shared task (v. 1.0)

Corpus 1.0 [Savary et al.(2018)] Corpus Sentences Tokens VMWEs Licence Training 230,062 4,536,603 52,724 CC v4 Testing 44,314 902,601 9,494 Shared task 1.0 [Savary et al.(2017)] & 1.1 7 systems, all languages covered, Evaluation measures including VMWE variability.

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Bibliography I

Abeillé, A. and Schabes, Y. (1989). Parsing idioms in lexicalized tags. In H. L. Somers and M. M. Wood, eds., Proceedings of the 4th Conference of the European Chapter of the ACL, EACL’89, Manchester, pp. 1–9. The Association for Computer Linguistics. Abeillé, A. and Schabes, Y. (1996). Non-compositional discontinuous constituents in Tree Adjoining Grammar. In H. Bunt and A. van Horck, eds., Discontinuous Constituency, pp. 279–306. Mouton de Gruyter, Berlin, Germany. Abeillé, A. (2002). Une grammaire électronique du français. CNRS Editions. Al-Haj, H., Itai, A., and Wintner, S. (2014). Lexical Representation of Multiword Expressions in Morphologically-complex Languages. International Journal of Lexicography, 27(2), 130–170. Attia, M. A. (2006). Accommodating multiword expressions in an Arabic LFG grammar. In Proceedings of the 5th international conference on Advances in Natural Language Processing,

• pp. 87–98, Berlin. Springer.

Bond, F., Ho, J. Q., and Flickinger, D. (2015). Feeling our way to an analysis of English possessed idioms. In S. Müller, ed., Proceedings of the 22nd International Conference on Head- Driven Phrase Structure Grammar, pp. 61–74, Stanford, CA. CSLI Publications.

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Bibliography II

Candito, M., Constant, M., Ramisch, C., Savary, A., Parmentier, Y., Pasquer, C., and Antoine, J.-Y. (2017). Annotation d’expressions polylexicales verbales en français. In TALN 2017, Orléans, France. Association pour le Traitement Automatique des Langues. Copestake, A., Lambeau, F., Villavicencio, A., Bond, F., Baldwin, T., Sag, I. A., and Flickinger,

• D. (2002).

Multiword expressions: linguistic precision and reusability. In Proceedings of LREC 2002. Crabbé, B., Duchier, D., Gardent, C., Roux, J. L., and Parmentier, Y. (2013). XMG: extensible metagrammar. Computational Linguistics, 39(3), 591–629. Grégoire, N. (2010). DuELME: a Dutch electronic lexicon of multiword expressions. Language Resources and Evaluation, 44(1-2). Gross, G. (1988). Degré de figement des noms composés. Langages, 90, 57–71. Paris : Larousse. Gross, M. (1986). Lexicon-grammar: The Representation of Compound Words. In Proceedings of the 11th Coference on Computational Linguistics, pp. 1–6, Stroudsburg, PA,

• USA. Association for Computational Linguistics.

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Bibliography III

Herzig Sheinfux, L., Arad Greshler, T., Melnik, N., and Wintner, S. (2015). Hebrew verbal multi-word expressions. In S. Müller, ed., Proceedings of the 22nd International Conference on Head-Driven Phrase Structure Grammar, Nanyang Technological University (NTU), Singapore, pp. 122–135, Stanford,

• CA. CSLI Publications.

Jacquemin, C. (2001). Spotting and Discovering Terms through Natural Language Processing. MIT Press. Lichte, T. and Kallmeyer, L. (2016). Same syntax, different semantics: A compositional approach to idiomaticity in multi-word expressions. In C. Piñón, ed., Empirical Issues in Syntax and Semantics 11, pp. 111–140. Lichte, T., Petitjean, S., Savary, A., and Waszczuk, J. (2018). Lexical encoding formats for multi-word expressions: The challenge of “irregular” regularities. In Y. Parmentier and J. Waszczuk, eds., Representation and Parsing of Multiword Expressions,

• pp. 41–72. Language Science Press, Berlin.

McShane, M., Nirenburg, S., and Beale, S. (2015). The Ontological Semantic treatment of multiword expressions. Lingvisticæ Investigationes, 38(1), 73–110. Mel’čuk, I., Arbatchewsky-Jumarie, N., Dagenais, L., Elnitsky, L., Iordanskaja, L., Lefebvre, M.-N., and Mantha, S. (1988). Dictionnaire explicatif et combinatoire du français contemporain: Recherches lexico-sémantiques. Presses de l’Univ. de Montréal.

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Bibliography IV

Pasquer, C., Savary, A., Antoine, J.-Y., and Ramisch, C. (2018). Towards a Variability Measure for Multiword Expressions. In Proceedings of NAACL. Accepted paper. Pasquer, C., Savary, A., Antoine, J.-Y., and Ramisch, C. (sub.). If you’ve seen some, you’ve seen them all: Identifying variants of multiword expressions. In Proceedings of COLING 2018. Petitjean, S., Duchier, D., and Parmentier, Y. (2016). XMG 2: Describing description languages. In M. Amblard, P. de Groote, S. Pogodalla, and C. Retoré, eds., Logical Aspects of Computational Linguistics. Celebrating 20 Years of LACL (1996-2016) - 9th International Conference, LACL 2016, Nancy, France, December 5-7, 2016, Proceedings, pp. 255–272. Przepiórkowski, A., Hajnicz, E., Patejuk, A., and Woliński, M. (2014). Extended phraseological information in a valence dictionary for NLP applications. In Proceedings of the Workshop on Lexical and Grammatical Resources for Language Processing (LG-LP 2014), pp. 83–91, Dublin, Ireland. Association for Computational Linguistics and Dublin City University. Sag, I., Baldwin, T., Bond, F., Copestak, A., and Flickinger, D. (2002). Multiword expressions: A pain in the neck for NLP. In Proceedings of the 3rd International Conference on Intelligent Text Processing and Computational Linguistics (CICLing-2002), p. 1–15, Mexico City,Mexico.

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Bibliography V

Savary, A. (2009). Multiflex: A Multilingual Finite-State Tool for Multi-Word Units. In S. Maneth, ed., Implementation and Application of Automata, pp. 237–240. Springer Berlin Heidelberg. preprint: http://www.info.univ-tours.fr/~savary/English/papersASgb.html#CIAA09. Savary, A., Ramisch, C., Cordeiro, S., Sangati, F., Vincze, V., QasemiZadeh, B., Candito, M., Cap, F., Giouli, V., Stoyanova, I., and Doucet, A. (2017). The PARSEME Shared Task on Automatic Identification of Verbal Multiword Expressions. In Proceedings of the EACL’17 Workshop on Multiword Expressions. Savary, A., Candito, M., Mititelu, V., Bejček, E., Cap, F., Čéplö, S., Cordeiro, S. R., Eryiğit, G., Giouli, V., van Gompel, M., HaCohen-Kerner, Y., Kovalevskait˙ e, J., Krek, S., Liebeskind, C., Monti, J., Parra Escartín, C., van der Plas, L., QasemiZadeh, B., Ramisch, C., Sangati, F., Stoyanova, I., and Vincze, V. (2018). Selected Extended Papers from the MWE 2017 Workhop, chapter The PARSEME multilingual corpus of verbal multiword expressions. Savary, A., Petitjean, S., Lichte, T., Kallmeyer, L., and Waszczuk, J. (sub.). Object-oriented lexical encoding of multiword expressions: Short and sweet. In Proceedings of COLING 2018. Tutin, A. (2016). Comparing morphological and syntactic variations of support verb constructions and verbal full phrasemes in French: a corpus based study. In PARSEME COST Action. Relieving the pain in the neck in natural language processing: 7th final general meeting, Dubrovnik, Croatia.

MWEs Metagrammar Adding MWEs Evaluation Conclusions App A: Variability measure App B: annotation Bibliography

Bibliography VI

Vaidya, A., Rambow, O., and Palmer, M. (2014). Light verb constructions with ‘do’ and ‘be’ in Hindi: A TAG analysis. In Proceedings of the Workshop on Lexical and Grammatical Resources for Language Processing,

• pp. 127–136.

Villavicencio, A., Copestake, A., Waldron, B., and Lambeau, F. (2004). Lexical Encoding of MWEs. In ACL Workshop on Multiword Expressions: Integrating Processing, July 2004, pp. 80–87.