Elamur Research⌕
☰
home / papers / 2012 / Utilisation de la translittération arabe pour l'amélioration de l'alignement de mots à partir de corpus parallèles français-arabe (Using Arabic Transliteration to Improve Word Alignment from French-Arabic Parallel Corpora) [in French]

Utilisation de la translittération arabe pour l'amélioration de l'alignement de mots à partir de corpus parallèles français-arabe (Using Arabic Transliteration to Improve Word Alignment from French-Arabic Parallel Corpora) [in French]

Paper · 2012 · Houda Saadane and Nasredine Semmar

Summary

The paper tackles a weakness in French–Arabic word alignment: proper names. Arabic names reach Latin script in many competing spellings, so a bilingual lexicon alone often misses them. The authors build a transliterator that turns Arabic names, with or without vowel marks, into the range of plausible Latin forms. They do not impose a single standard. Instead, they assembled the letter equivalences bottom-up from published norms and from spellings actually seen online, in a corpus covering the target languages.

This transliterator feeds the cognate-matching step of an existing linguistic aligner. That aligner pairs single words using a bilingual lexicon plus cues such as position, part of speech, named-entity status and cognates. It then pairs compound words through the syntactic dependency relations that connect their parts.

The system was evaluated in two ways: against a manually built reference alignment, and through Moses translation quality. Both evaluations improve, including a gain of +0.32 BLEU. This matters for building bilingual lexicons, for machine translation and for cross-language information retrieval, where name variation is a major source of misses.

Key points

  • Name variation is severe. Mouammar Kadhafi has more than 60 Latin spellings, and Lawrence of Arabia spelled Jeddah three different ways in a single 1926 book (25, 6 and 1 times).
  • Standard Arabic script has 28 letters (25 consonants, 3 vowels). Short vowels and consonant doubling (shadda, tanwin) are usually left unwritten, which is the main source of ambiguity.
  • The corpus study found that informal Latin-script Arabic (SMS, social sites) uses digits such as 3, 7, 9 and 6 for letters with no Latin equivalent.
  • The authors attribute spelling variation partly to colonial history (French influence in the Maghreb, British/American influence in the Middle East) and partly to the absence of any common transliteration norm.
  • Adding transliteration-based cognate alignments to the Moses training data raised BLEU by +0.32. The authors note this is hard to call significant, since training used only 10,000 sentence pairs for the translation model and about 11,000 sentences for the language model.

Interesting ideas

  • I1. Treat a name as a set of plausible spellings generated by a one-to-many mapping, and accept a match when any variant fits, instead of forcing one canonical form.
  • I2. Derive character correspondences empirically from how people actually write, including informal conventions like digits for missing letters, rather than from an official norm.
  • I3. Align in two stages: first anchor single words with a lexicon and cheap cues (position, POS, entity type, cognates), then align multiword units by following the dependency relations between their components.
  • I4. Evaluate a component both intrinsically against a hand-made reference and extrinsically through a downstream task, and openly flag when a small downstream gain may not be significant.

For Mova

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

  • I1 (world): Wikidata-derived entities in the knowledge graph could carry generated spelling-variant sets, so mentions like "Gadhafi" and "Qaddafi" resolve to the same entity.
  • I2 (tokens): a normalization layer for noisy English input (leetspeak, phonetic or nonstandard spellings) could be learned from observed usage before lemmatization.
  • I3 (back): translation through the meta-world could match multiword expressions as UD subtrees anchored on already-aligned head words, which keeps every alignment explainable.
  • I4 (learn): self-induced repair rules could be scored on treebank accuracy and on answer-task accuracy, with small gains on small test sets treated as unproven.

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

📚 Shelves 4 Computer Science · 9,086 works Artificial Intelligence · 8,889 works Arts and Humanities · 9,296 works Language and Linguistics · 9,206 works

🧭 Topics 2 Semitic languages · 1,484 works French · 65 works

Cite

Cite as Houda Saadane and Nasredine Semmar. 2012. Utilisation de la translittération arabe pour l'amélioration de l'alignement de mots à partir de corpus parallèles français-arabe (Using Arabic Transliteration to Improve Word Alignment from French-Arabic Parallel Corpora) [in French]. Proceedings of the Joint Conference JEP-TALN-RECITAL 2012, volume 2: TALN.

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

@inproceedings{saadane2012utilisation,
  title = {{Utilisation de la translittération arabe pour l'amélioration de l'alignement de mots à partir de corpus parallèles français-arabe (Using Arabic Transliteration to Improve Word Alignment from French-Arabic Parallel Corpora) [in French]}},
  author = {Houda Saadane and Nasredine Semmar},
  booktitle = {Proceedings of the Joint Conference JEP-TALN-RECITAL 2012, volume 2: TALN},
  year = {2012},
  url = {https://aclanthology.org/F12-2010/}
}

References

17 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. A BDULJALEEL , N. et L ARKEY , L. (2003). Statistical transliteration for English-Arabic Cross Language Information Retrieval. In Proceedings of the Twelfth ACM International Conference on Information and Knowledge Management, New Orleans, Louisiana, pages 139–146. find
  2. A LGHAMDI , M. (2005). Alghorithms for Romanizing Arabic names. In Journal of King Saud University: Computer Sciences and Information, n° 17, 2005, Riyadh, pages 1–27. find
  3. A LSALMAN , A., A LGHAMDI , M., A LHUQAYL , K. et A LSUBAI , S. (2007). Romanization System for Arabic Names. In Proceedings of The First International Symposium on Computer and Arabic Language (ISCAL – 07), Riyadh, pages 214–227. find
  4. B ARKAT -D EFRADAS , M., H AMDI , R. et P ELLEGRINO , F. (2004). De la caractérisation linguistique à l’identification automatique des dialectes arabes. In Proceedings of MIDL 2004. find
  5. B ESANÇON , R., D E C HALENDAR , G., F ERRET , O., G ARA , F., L AIB , M., M ESNARD , O. et S EMMAR , N. (2010). Lima: A multilingual framework for linguistic analysis and linguistic resources development and evaluation. In Proceedings of LREC 2010, Malta. find
  6. G UIDERE , M. (2004). Le traitement de la parole et la détection des dialectes arabes. In Langues stratégiques et défense nationale, Publications du CREC, Saint-Cyr, pages 53–75. find
  7. K NIGHT , K. et G RAEHL , J. (1997). Machine transliteration. In Journal version Computational Linguistics, 24(4), 1997, pages 599–612. find
  8. 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 ORGAN , C., Z ENS , R., D YER , C., B OJAR , O., C ONSTANTIN , A. et H ERBST , E. (2007). Moses: Open source toolkit for statistical machine translation. In Proceedings of ACL 2007, demo session, Prague. find
  9. K OEHN , P. et H OANG , H. (2007). Factored Translation Models. In Proceedings of the Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (ACL 2007), Prague, pages 868–876. find
  10. M IHALCEA , R. et P EDERSEN , T. (2003). An evaluation exercise for word alignment. In Proceedings of the Workshop on Building and Using Parallel Texts: Data Driven Machine Translation and Beyond, Edmonton, pages 10–10. our card
  11. S EMMAR , N. et L AIB , M. (2010). Using a Hybrid Word Alignment Approach for Automatic Construction and Updating of Arabic to French Lexicons. In Proceedings of LREC 2010: Workshop on Language Resources and Human Technologies for Semitic Languages, Malta. find
  12. S HAO , L. et N G , H. T. (2004). Mining new word translations from comparable corpora. In Proceedings of the 20th International Conference on Computational Linguistics (COLING’04), Stroudsburg, pages 618–624. find
  13. S HERIF , T. et K ONDRAK , G. (2007). Bootstrapping a stochastic transducer for Arabic-English transliteration extraction. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics (ACL 2007), Prague, pages 864–871. find
  14. S TALLS , B. et K NIGHT , K . (1998). Translating names and technical terms in arabic text. In Proceedings of the COLING/ACL Workshop on Computational Approches to Semitic Languages, Montreal. find
  15. T AO , T., Y OON , S. Y., F ISTER , A., S PROAT , R. et Z HAI , C. (2006). Unsupervised named entity transliteration using temporal and phonetic correlation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP’06), Sydney, pages 250–257. find
  16. V ERONIS , J., H AMON , O., A YACHE , C., B ELMOUHOUB , R., K RAIF , O., L AURENT , D., N GUYEN , T. M. H., S EMMAR , N., S TUCK , F. et Z AGHOUANI , W. (2008). Arcade II Action de recherche concertée sur l’alignement de documents et son évaluation. In Chapitre 2, Editions Hermès. find
  17. Y ASER , A. O. et K NIGHT , K. (2002). Translating named entities using monolingual and bilingual resources. In Proceedings of the 40th Annual Meeting of the Association of Computational Linguistics (ACL’02), Philadelphia, pages 400–408. our card

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/saadane-2012-utilisation-de-la-translitteration-arabe/