Elamur Research⌕
☰
home / papers / 2025 / Preserving Comorian Linguistic Heritage: Bidirectional Transliteration Between the Latin Alphabet and the Kamar-Eddine System

Preserving Comorian Linguistic Heritage: Bidirectional Transliteration Between the Latin Alphabet and the Kamar-Eddine System

Paper · 2025 · Abdou Mohamed Naira, Abdessalam Bahafid, Zakarya Erraji et al.

Summary

Comorian (ShiKomori) is now mostly written in Latin letters. An older Arabic-based orthography, the Kamar-Eddine system from the 1960s, is still the only script some elderly speakers read, and many archival manuscripts use it. No NLP work had addressed it. The authors first collect the system's rules. Kamar-Eddine replaced diacritics with long-vowel letters, used the shadda to mark affricates, and kept letters borrowed from Persian. From these rules they build Shialifube, a rule-based tool that converts text in both directions. It detects the input script, loads the matching correspondence dictionary, and handles digraphs before single letters. No parallel Latin–Arabic corpus exists, so they evaluate by sending Latin text to Arabic and back and comparing the result with the original. After four rounds of refinement this gives 14.84% WER and 9.56% CER. They also fine-tune Whisper on transliterated data, producing the first Comorian ASR model in Arabic script. Its accuracy comes close to the Latin-script model's. The work matters because it opens a low-resource heritage script to digitization and NLP.

Key points

  • Shialifube is released as a PyPI library, with code on GitHub and a HuggingFace Space. The four dialects (ShiNgazidja, ShiMwali, ShiNdzuani, ShiMaore) are treated as one language, partly because data is scarce and partly because the Kamar-Eddine rules make no dialect distinctions.
  • Round-trip WER/CER across the four iterations:
    • no digraph handling: 68.56/34.41
    • adding digraphs and long vowels: 43.09/21.30
    • standardizing the corpus: 33.89/16.75
    • handling remaining edge cases: 14.84/9.56
  • In the Kamar-Eddine system, /e/ and /o/ get dedicated long-vowel letters. This removes ambiguities such as mezi "month" vs mizi "roots", which earlier adaptations wrote identically.
  • Whisper fine-tuning results:
    • Latin script, with Swahili as the language setting: WER 35.48%, CER 17.76%.
    • Arabic script, with Arabic as the language setting: WER 37.44%, CER 21.42%.
  • The authors give two reasons for the Arabic-script gap: the transliterated (not hand-annotated) training data is noisier, and the pretrained model's vocabulary produces more unknown tokens for that script.
  • The motivation comes partly from Maltese, where rewriting text in Arabic characters improved NLP task scores.

Interesting ideas

  • I1. Each output comes with a confidence score computed by inverting it and measuring how much of the original survives. This gives a cheap self-check when no gold data exists.
  • I2. The authors judge a data transformation by training a downstream model on the transformed data and comparing it to a model trained on the original. If the conversion damaged the data, the downstream score would drop.
  • I3. The error rate is logged after each batch of rule changes (digraphs, then normalization, then edge cases). This makes each rule family's contribution visible and explainable.
  • I4. Multi-character units are matched and replaced before single characters, which prevents greedy mis-segmentation.
  • I5. Simple rules detect the input type first and route it to the right resource, instead of using one model for every case.

For Mova

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

  • I1 (back): The text → graph → text compressor could report a per-sentence round-trip WER as a confidence signal, flagging graphs that lose information.
  • I2 (tree): When converting treebanks between UD dialects, train the perceptron parser on converted and unconverted versions and compare LAS, to catch conversion rules that corrupt structure.
  • I3 (learn): Self-induced repair rules could be added in tracked batches, each with its own before/after metric, so a rule that helps nothing can be identified and dropped.
  • I4 (tokens): The en tokenizer could apply the same longest-match-first ordering to multiword expressions and contractions before splitting.
  • I5 (tokens): A lightweight rule-based detector for genre or text register could pick specialized lexicons before tagging, keeping the routing decision explainable.

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

 Institutions

🎓Institut National de Statistique et d'Economie AppliquéeRabat · Morocco2 works · 16 citations

📚 Shelves 5 Computer Science · 9,086 works Artificial Intelligence · 8,889 works Arts and Humanities · 9,296 works Language and Linguistics · 9,206 works Literature and Literary Theory · 340 works

🧭 Topics 1 Semitic languages · 1,484 works

Cite

Cite as Abdou Mohamed Naira, Abdessalam Bahafid, Zakarya Erraji, Anass Allak, Mohamed Soibira Naoufal and Imade Benelallam. 2025. Preserving Comorian Linguistic Heritage: Bidirectional Transliteration Between the Latin Alphabet and the Kamar-Eddine System. Proceedings of the 9th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (LaTeCH-CLfL 2025).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2025.latechclfl-1.2

@inproceedings{naira2025preserving,
  title = {{Preserving Comorian Linguistic Heritage: Bidirectional Transliteration Between the Latin Alphabet and the Kamar-Eddine System}},
  author = {Abdou Mohamed Naira, Abdessalam Bahafid, Zakarya Erraji, Anass Allak, Mohamed Soibira Naoufal and Imade Benelallam},
  booktitle = {Proceedings of the 9th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (LaTeCH-CLfL 2025)},
  year = {2025},
  doi = {10.18653/v1/2025.latechclfl-1.2},
  url = {https://doi.org/10.18653/v1/2025.latechclfl-1.2}
}

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. Moneim Abdourahamane, Christian Boitet, Valérie Bellynck, Lingxiao Wang, and Hervé Blanchon. 2016. Construction d’un corpus parallèle français-comorien en utilisant de la TA français-swahili. In TALAf (Traitement Automatique des Langues africaines), Paris, France. find
  2. Kurt Micallef, Fadhl Eryani, Nizar Habash, Houda Bouamor, and Claudia Borg. 2023. Exploring the impact of transliteration on nlp performance: Treating maltese as an arabic dialect. In Proceedings of the Workshop on Computation and Written Language (CAWL 2023), page 22–32. Association for Computational Linguistics. find
  3. Mohamed Ahmed Chamanga. 2022. ShiKomori, the Bantu Language of the Comoros: Status and Perspectives, page 79–98. BRILL. find
  4. John Mugane. 2017. The odyssey of ajami and the swahili people. Islamic Africa, 8(1–2):193–216. find
  5. Claude Allibert. 2015. L’archipel des comores et son histoire ancienne. essai de mise en perspective des chroniques, de la tradition orale et des typologies de céramiques locales et d’importation. Afriques, 06. ↗
  6. Abdou Mohamed Naira, Benelallam Imade, Bahafid Abdessalam, and Erraji Zakarya. 2024. Datasets creation and empirical evaluations of cross-lingual learning on extremely low-resource languages: A focus on comorian dialects. In Proceedings of The 18th Linguistic Annotation Workshop (LAW-XVIII), pages 140– 149, St. Julians, Malta. Association for Computational Linguistics. find
  7. Mohamed Ahmed Chamanga and Noël Jacques Gueunier. 1977. Recherches sur l’instrumentalisation du comorien : problèmes d’adaptation lexicale (d’après la version comorienne de la loi du 23 novembre 1974). Cahiers d’Études africaines, 66-67:213–239. ↗
  8. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. Robust speech recognition via large-scale weak supervision. Preprint, arXiv:2212.04356. ↗ our card
  9. Galien Currah. 2015. Orthographe wolofal. find
  10. Soufiane Hajbi, Omayma Amezian, Nawfal El Moukhi, Redouan Korchiyne, and Younes Chihab. 2024. Moroccan arabizi-to-arabic conversion using rule-based transliteration and weighted levenshtein algorithm. Scientific African, 23:e02073. ↗
  11. Annachiara Raia. 2021. One text, many forms: a comparative view of the variability of swahili manuscripts. In R.H. Samsom and C. Vierke, editors, Manuscript Cultures, 17, pages 65–86. find
  12. Yova Kementchedjhieva and Anders Søgaard. 2023. Grammatical error correction through round-trip machine translation. In Findings of the Association for Computational Linguistics: EACL 2023, page 2208–2215. Association for Computational Linguistics. ↗
  13. Maurizio Serva and Michele Pasquini. 2021. The sabaki languages of comoros. ↗
  14. Ali Shazal, Aiza Usman, and Nizar Habash. 2020. A unified model for Arabizi detection and transliteration using sequence-to-sequence models. In Proceedings of the Fifth Arabic Natural Language Processing Workshop, pages 167– 177, Barcelona, Spain (Online). Association for Computational Linguistics. find
  15. B. Klouche and S. M. Benslimane. 2020. Arabizi Chat Alphabet Transliteration to Algerian Dialect, page 790–797. Springer International Publishing. find
  16. Michel Lafon. 2007. Le système Kamar-Eddine : une tentative originale d’écriture du comorien en graphie arabe. Ya Mkobe, 14-15:29–48. find
  17. Yu-Hsiang Lin, Chian-Yu Chen, Jean Lee, Zirui Li, Yuyan Zhang, Mengzhou Xia, Shruti Rijhwani, Junxian He, Zhisong Zhang, Xuezhe Ma, Antonios Anastasopoulos, Patrick Littell, and Graham Neubig. 2019. Choosing transfer languages for cross-lingual learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, page 3125–3135. Association for Computational Linguistics. ↗ our card

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/naira-2025-preserving-comorian-linguistic-heritage-bidirectional/