Elamur Research⌕
☰
home / papers / 2018 / Traduction automatique de corpus en anglais annotés en sens pour la désambiguïsation lexicale d'une langue moins bien dotée, l'exemple de l'arabe (Automatic Translation of English Sense Annotated Corpora for Word Sense Disambiguation of a Less Well-endowed Language, the Example of Arabic)

Traduction automatique de corpus en anglais annotés en sens pour la désambiguïsation lexicale d'une langue moins bien dotée, l'exemple de l'arabe (Automatic Translation of English Sense Annotated Corpora for Word Sense Disambiguation of a Less Well-endowed Language, the Example of Arabic)

Paper · 2018 · Marwa Hadj Salah, Loïc Vial, Hervé Blanchon et al.

Summary

Arabic Word Sense Disambiguation (WSD) has little sense-annotated training data. Until recently it also had no freely available gold standard, which made systems hard to build and hard to compare. The authors address this by automatically creating Arabic training data from English. They take the 12 English corpora unified in UFSAC, all tagged with Princeton WordNet 3.0 senses, and translate them with a Moses statistical English–Arabic system. Word alignments from Moses then carry each sense tag from an English word to its Arabic counterpart. A preprocessing chain prepares the English input, and a postprocessing chain repairs the Arabic output: it fixes word order and duplicated words, merges compound pieces, adds MADAMIRA POS tags and detokenizes. To test the resulting data, they reimplement the SVM classifier from NUS-PT and evaluate on the Arabic part of OntoNotes Release 5.0. With postprocessing, the classifier reaches about 71% precision and 67% F1, roughly in line with typical English WSD scores. The approach suggests that any language with decent MT from English could get a usable supervised WSD system cheaply.

Key points

  • The source data is all 12 UFSAC corpora, including SemCor, DSO, WNGT, MASC, OMSTI, OntoNotes and several SensEval/SemEval sets. OMSTI is by far the largest, at about 28.5M words.
  • The Moses SMT system was trained on LDC-Ummah, LDC-News, News Commentary and TED Talks. Test and dev data were 3,200 held-out lines, 800 drawn from each corpus, and the system scored 24.51 BLEU.
  • The WSD classifier reimplements NUS-PT with three feature groups:
    • POS tags of the target and three words on each side;
    • 11 local collocation windows;
    • a binary vector of sentence lemmas.
  • On Arabic OntoNotes, the system assigned senses to 11,346 of 12,524 annotated words; the rest were absent from training. With a first-sense fallback it reached 67.55% precision and 65.06% F1.
  • Postprocessing raised precision to 70.86% (+3.31) and F1 to 67.36% (+2.30), with 8,040 words correct. The authors believe this is the first WSD system evaluated on Arabic OntoNotes.

Interesting ideas

  • I1. Each source token is wrapped in a tag carrying a unique id before translation, so annotations survive MT and can be reattached through alignment; this keeps provenance across a lossy transformation.
  • I2. After projection, target pieces that share one source id are merged and words inserted by the MT system are dropped; this is a cheap rule-based repair of noisy alignment.
  • I3. Postprocessing is evaluated as an ablation on the same gold test set, so its contribution is isolated as a measurable gain of about 3 precision points.
  • I4. Coverage is reported separately from accuracy (words seen in training versus not), and unseen words fall back to WordNet's first sense instead of being silently skipped.
  • I5. Heterogeneous corpora are unified into one format and one sense inventory (WordNet 3.0) before any processing, so all 12 corpora can feed a single pipeline.

For Mova

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

  • I1 (back): Translation through the meta-world could tag every graph node with a stable id that persists into generated text, making each output word traceable to its source node.
  • I2 (learn): Silver data projected or agreed between witness parsers could pass through explicit id-based merge/drop rules, and each rule could be logged as an inspectable repair.
  • I3 (learn): Every self-induced repair rule set could be scored with and without it on a fixed gold set, which keeps claimed improvements honest.
  • I4 (world): Sense assignment from WordNet could use first-sense fallback and mark it as such, so explanations distinguish learned decisions from defaults.
  • I5 (world): The compiled ontology could normalize all lexical resources to one WordNet version first, so frame and sense lookups never mix inventories.

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 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

🧭 Topics 3Research subjects and languages the work is about.English · 2,379 worksSemitic languages · 3,066 worksWord sense disambiguation · 87 works

Cite

Cite as Marwa Hadj Salah, Loïc Vial, Hervé Blanchon, Mounir Zrigui and Didier Schwab. 2018. Traduction automatique de corpus en anglais annotés en sens pour la désambiguïsation lexicale d'une langue moins bien dotée, l'exemple de l'arabe (Automatic Translation of English Sense Annotated Corpora for Word Sense Disambiguation of a Less Well-endowed Language, the Example of Arabic). Actes de la Conférence TALN. Volume 1 - Articles longs, articles courts de TALN.

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

@inproceedings{salah2018traduction,
  title = {{Traduction automatique de corpus en anglais annotés en sens pour la désambiguïsation lexicale d'une langue moins bien dotée, l'exemple de l'arabe (Automatic Translation of English Sense Annotated Corpora for Word Sense Disambiguation of a Less Well-endowed Language, the Example of Arabic)}},
  author = {Marwa Hadj Salah, Loïc Vial, Hervé Blanchon, Mounir Zrigui and Didier Schwab},
  booktitle = {Actes de la Conférence TALN. Volume 1 - Articles longs, articles courts de TALN},
  year = {2018},
  url = {https://aclanthology.org/2018.jeptalnrecital-court.15/}
}

References

15 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. F. B ENARMARA , N. H ATOUT , P. M ULLER & S. O ZDOWSKA , Eds. (2007). find
  2. Actes de TALN 2007 (Traitement automatique des langues naturelles), Toulouse. ATALA, IRIT. find
  3. C AI J. F., L EE W. S. & T EH Y. W. (2007). Nus-ml : Improving word sense disambiguation using topic features. In Proceedings of the 4th International Workshop on Semantic Evaluations, SemEval ’07, p. 249–252, Stroudsburg, PA, USA : Association for Computational Linguistics. find
  4. C HAN Y. S., N G H. T. & Z HONG Z. (2007). Nus-pt : exploiting parallel texts for word sense disambiguation in the english all-words tasks. In Proceedings of the 4th International Workshop on Semantic Evaluations, p. 253–256 : Association for Computational Linguistics.
  5. Actes de TALN 2015 (Traitement automatique des langues naturelles), Caen. ATALA, HULTECH. find
  6. H ADJ S ALAH M., B LANCHON H., Z RIGUI M. & S CHWAB D. (2016). Amélioration de la traduction automatique d’un corpus annoté. In JEP-TALN-RECITAL 2016. find
  7. I ACOBACCI I., P ILEHVAR M. T. & N AVIGLI R. (2016). Embeddings for word sense disambiguation : An evaluation study. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1 : Long Papers), p. 897–907, Berlin, Germany : Association for Computational Linguistics. K OEHN P., H OANG H., B IRCH A., C ALLISON -B URCH C., F EDERICO M., B ERTOLDI N., C OWAN B., S HEN W., M ORAN C., Z ENS R. et al. (2007). Moses : Open source toolkit for statistical machine translation. In Proceedings of the 45th annual meeting of the ACL on interactive poster and demonstration sessions, p. 177–180 : Association for Computational Linguistics. find
  8. L AIGNELET M. & R IOULT F. (2009). Repérer automatiquement les segments obsolescents à l’aide d’indices sémantiques et discursifs. In Actes de TALN 2009 (Traitement automatique des langues naturelles), Senlis : ATALA LIPN. find
  9. L ANGLAIS P. & P ATRY A. (2007). Enrichissement d’un lexique bilingue par analogie. In (Benarmara et al., 2007), p. 101–110. find
  10. M ILLER G. A. (1995). Wordnet : a lexical database for english. Communications of the ACM, 38(11), 39–41.
  11. N ASIRUDDIN M., T CHECHMEDJIEV A., B LANCHON H. & S CHWAB D. (2015). Création rapide et efficace d’un système de désambiguïsation lexicale pour une langue peu dotée. In TALN 2015-22ème Conférence sur le Traitement Automatique des Langues Naturelles„ Caen, France. find
  12. N OVISCHI A., S RIKANTH M. & B ENNETT A. (2007). Lcc-wsd : System description for english coarse grained all words task at semeval 2007. In Proceedings of the 4th International Workshop on Semantic Evaluations, SemEval ’07, p. 223–226, Stroudsburg, PA, USA : Association for Computational Linguistics.
  13. S CHWAB D. (2017). Cours master mosig. find
  14. S ERETAN V. & W EHRLI E. (2007). Collocation translation based on sentence alignment and parsing. In (Benarmara et al., 2007), p. 401–410. find
  15. V IAL L., L ECOUTEUX B. & S CHWAB D. (2017). Uniformisation de corpus anglais annotés en sens. In 24ème Conférence sur le Traitement Automatique des Langues Naturelles, Orléans, France. W EISCHEDEL R., P ALMER M., M ARCUS M., H OVY E., P RADHAN S., R AMSHAW L., X UE N., T AYLOR A., K AUFMAN J., F RANCHINI M., E L -B ACHOUTI M., B ELVIN R. & H OUSTON A. (2015). Ontonotes release 5.0. LDC2013T19. Web Download. Philadelphia : Linguistic Data Consortium. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/salah-2018-traduction-automatique-de-corpus-en/