Elamur Research⌕
☰
home / papers / 2022 / Annotation d'expressions polylexicales verbales en arabe : validation d'une procédure d'annotation multilingue (Annotating Verbal Multiword Expressions in Arabic : Assessing the Validity of a Multilingual)

Annotation d'expressions polylexicales verbales en arabe : validation d'une procédure d'annotation multilingue (Annotating Verbal Multiword Expressions in Arabic : Assessing the Validity of a Multilingual)

Paper · 2022 · Najet Hadj Mohamed, Cherifa Ben Khelil, Agata Savary et al.

Summary

The paper extends the PARSEME framework for verbal multiword expressions (VMWEs) to Modern Standard Arabic. An earlier Arabic PARSEME corpus was never released and did not fully follow the methodology, so Arabic was missing from the PARSEME shared tasks. The authors chose the Prague Arabic Dependency Treebank (PADT) because it is the only fully open Arabic UD corpus. They added VMWE annotations in the CUPT format, which extends CoNLL-U. Each candidate went through a fixed procedure: find a verb plus a co-occurring word, reduce it to canonical form, mark the lexicalized parts, and run the PARSEME decision trees for category.

Two native speakers independently annotated 1062 sentences to measure inter-annotator agreement (IAA). Unlike most PARSEME languages, this was done at the start of the campaign rather than the end. Agreement was satisfactory for identifying spans and good for assigning categories. The resulting corpus holds 1252 VMWEs. The guidelines worked for Arabic with no changes, which gives Arabic a usable resource for VMWE identification.

Key points

  • IAA sample (1062 PADT sentences): the two annotators marked 763 and 704 VMWEs, giving F span 0.699, κ span 0.626 and κ cat 0.864. Among 26 PARSEME IAA estimates, these rank 12th, 14th and 12th. Only Chinese used more annotations to estimate IAA.
  • Current corpus: 1847 sentences and 70,498 tokens, with about 0.68 VMWEs per sentence.
  • Category counts: LVC.full 673, VID 345, IAV 165, LVC.cause 68, MVC 1. IRV and VPC do not occur at all.
  • Universal categories: VID and LVC are confirmed as present in Arabic. LVC.full is roughly twice as frequent as VID, while LVC.cause is rare.
  • Discontinuity: about 58% of annotated VMWEs are discontinuous, which makes automatic identification hard.
  • Main source of disagreement: telling a verbal idiom apart from a literal reading of the same words. For example, "cut the road" can mean crossing a street or blocking someone.

Interesting ideas

  • I1. Measuring annotator agreement at the very beginning of a campaign checks both annotator readiness and the soundness of the guidelines before effort is wasted on flawed labels.
  • I2. Agreement is scored separately for span detection and for category labeling, and the span metric is strict: a partial overlap counts as a full miss. This shows whether errors come from boundaries or from typing.
  • I3. Multiword annotation is added as an extra layer on an existing UD treebank (CUPT on top of CoNLL-U), rather than as a separate resource. Syntax and lexicalized units then live in the same file.
  • I4. Light verb constructions are spotted by checking whether the noun, not the verb, carries the predicate. A related test asks whether the verb could be dropped, which helps separate LVC from idiom.
  • I5. Annotation follows a fixed, explainable pipeline: candidate, canonical form, lexicalized components, then a decision tree for the category.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): Run agreement checks between witness parsers and LLM teacher tiers on a small sample before generating bulk silver data, so bad specs or prompts are caught early.
  • I2 (learn): Evaluations of multiword spans or frame triggers could report exact-span F and category kappa separately, so repair rules target the right failure.
  • I3 (tree): Mova's treebanks could carry an MWE column, letting the parser output feed directly into later stages that need to know which tokens form one unit.
  • I4 (frames): "give advice" or "make a decision" should yield the frame of the noun's predicate, with the verb's subject as that predicate's agent, instead of a giving or making event.
  • I5 (world): A compiled MWE lexicon could store each entry's canonical form, fixed components and the decision-tree path, so every multiword reading can be traced and justified.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 works

🧭 Topics 3Research subjects and languages the work is about.Semitic languages · 3,066 worksMultilingual & cross-lingual NLP · 882 worksIdioms & multiword expressions · 222 works

Cite

Cite as Najet Hadj Mohamed, Cherifa Ben Khelil, Agata Savary, Iskander Keskes, Jean Yves Antoine and Lamia Hadrich Belguith. 2022. Annotation d'expressions polylexicales verbales en arabe : validation d'une procédure d'annotation multilingue (Annotating Verbal Multiword Expressions in Arabic : Assessing the Validity of a Multilingual). Actes de la 29e Conférence sur le Traitement Automatique des Langues Naturelles. Volume 1 : conférence principale.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{mohamed2022annotation,
  title = {{Annotation d'expressions polylexicales verbales en arabe : validation d'une procédure d'annotation multilingue (Annotating Verbal Multiword Expressions in Arabic : Assessing the Validity of a Multilingual)}},
  author = {Najet Hadj Mohamed, Cherifa Ben Khelil, Agata Savary, Iskander Keskes, Jean Yves Antoine and Lamia Hadrich Belguith},
  booktitle = {Actes de la 29e Conférence sur le Traitement Automatique des Langues Naturelles. Volume 1 : conférence principale},
  year = {2022},
  url = {https://aclanthology.org/2022.jeptalnrecital-taln.27/}
}

References

10 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. A L -B ADRASHINY M., H AWWARI A., G HONEIM M. & D IAB M. (2016). SAMER : a semi-automatically created lexical resource for Arabic verbal multiword expressions tokens paradigm and their morphosyntactic features. In Proceedings of the 12th Workshop on Asian Language Resources (ALR12), p. 113–122.
  2. A TTIA M., T ORAL A., T OUNSI L., P ECINA P. & V AN G ENABITH J. (2010). Automatic extraction of Arabic multiword expressions. In Proceedings of the 2010 Workshop on Multiword Expressions : from Theory to Applications, p. 19–27.
  3. C ARLETTA J. (1996). Assessing agreement on classification tasks : The kappa statistic. Computational Linguistics, 22(2), 249–254.
  4. C OHEN J. (1960). Kappa : Coefficient of concordance. Educ Psych Measurement, 20(37), 37–46. C ONSTANT M., E RYI ǦIT G., M ONTI J., VAN DER P LAS L., R AMISCH C., R OSNER M. & T ODI -RASCU A. (2017). Survey : Multiword expression processing : A Survey. Computational Linguistics, 43(4), 837–892. DOI : 10.1162/COLI_a_00302. ↗
  5. G HONEIM M. & D IAB M. (2013). Multiword expressions in the context of statistical machine translation. In Proceedings of the Sixth International Joint Conference on Natural Language Processing, p. 1181–1187, Nagoya, Japan : Asian Federation of Natural Language Processing. find
  6. G RIMES S., L I X., B IES A., K ULICK S., M A X. & S TRASSEL S. (2010). Creating arabic-english parallel word-aligned treebank corpora at ldc. find
  7. H AJIC J., S MRZ O., Z EMÁNEK P., Š NAIDAUF J. & B EŠKA E. (2004). Prague arabic dependency treebank : Development in data and tools. In Proc. of the NEMLAR Intern. Conf. on Arabic Language Resources and Tools, p. 110–117. find
  8. H AWWARI A., B AR K. & D IAB M. (2012). Building an Arabic multiword expressions repository. In Proceedings of the ACL 2012 Joint Workshop on Statistical Parsing and Semantic Processing of Morphologically Rich Languages, p. 24–29.
  9. R AMISCH C., C ORDEIRO S. R., S AVARY A., V INCZE V., B ARBU M ITITELU V., B HATIA A., B ULJAN M., C ANDITO M., G ANTAR P., G IOULI V., G ÜNGÖR T., H AWWARI A., I ÑURRIETA U., K OVALEVSKAIT Ė J., K REK S., L ICHTE T., L IEBESKIND C., M ONTI J., P ARRA E SCARTÍN C., Q ASEMI Z ADEH B., R AMISCH R., S CHNEIDER N., S TOYANOVA I., V AIDYA A. & W ALSH A. (2018). Edition 1.1 of the PARSEME shared task on automatic identification of verbal multiword expressions. In Proceedings of the Joint Workshop on Linguistic Annotation, Multiword Expressions and Constructions (LAW-MWE-CxG-2018), p. 222–240, Santa Fe, New Mexico, USA : Association for Computational Linguistics.
  10. R AMISCH C., S AVARY A., G UILLAUME B., W ASZCZUK J., C ANDITO M., V AIDYA A., B ARBU M ITITELU V., B HATIA A., I ÑURRIETA U., G IOULI V., G ÜNGÖR T., J IANG M., L ICHTE T., L IEBESKIND C., M ONTI J., R AMISCH R., S TYMNE S., W ALSH A. & X U H. (2020). Edition 1.2 of the PARSEME shared task on semi-supervised identification of verbal multiword expressions. In Proceedings of the Joint Workshop on Multiword Expressions and Electronic Lexicons, p. 107–118, online : Association for Computational Linguistics. S AVARY A., C ANDITO M., M ITITELU V. B., B EJEK E., C AP F., ÉPLÖ S., C ORDEIRO S. R., E RYIIT G., G IOULI V., VAN G OMPEL M., H A C OHEN -K ERNER Y., K OVALEVSKAIT J., K REK S., L IEBESKIND C., M ONTI J., E SCARTÍN C. P., VAN DER P LAS L., Q ASEMI Z ADEH B., R AMISCH C., S ANGATI F., S TOYANOVA I. & V INCZE V. (2018). PARSEME multilingual corpus of verbal multiword expressions. In S. M ARKANTONATOU , C. R AMISCH , A. S AVARY & V. V INCZE , Éds., Multiword expressions at length and in depth : Extended papers from the MWE 2017 workshop, p. 87–147. Berlin : Language Science Press. DOI : 10.5281/zenodo.1469555. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mohamed-2022-annotation-d-expressions-polylexicales-verbales/