Elamur Research⌕
☰
home / papers / 2025 / Lost in Variation: An Unsupervised Methodology for Mining Lexico-syntactic Patterns in Middle Arabic Texts

Lost in Variation: An Unsupervised Methodology for Mining Lexico-syntactic Patterns in Middle Arabic Texts

Paper · 2025 · Julien Bezançon, Rimane Karam and Gaël Lejeune

Summary

Middle Arabic, a mix of standard, colloquial and register-specific features, has received almost no attention in NLP, and no tools exist for it. The authors study recurring "formulas" in the Damascene popular epic SĪRAT AL-MALIK AL-ẒĀHIR BAYBARS. These are stock phrases marking moments such as nightfall, mood shifts or battles, and they behave much like multiword expressions. The corpus covers the first 90 digitized booklets, grouped into 9 volumes, with 53,843 sentences and about 929k tokens. A linguist manually selected 13 such formulas. The authors adapted an earlier method for identifying frozen and unfrozen MWEs. It combines CAMeL Tools tagging, token alignment, lexico-syntactic patterns and similarity measures across several linguistic layers. With it they extracted candidate variants of each formula and ranked them. They found that variation is confined to spelling, morphology and lexical choice, while syntax and meaning stay fixed. When a variant changes meaning, it acts as a different formula with its own narrative context. This gives a reusable, unsupervised way to map variation in non-standard text.

Key points

  • Corpus statistics: 53,843 sentences, 928,935 tokens and an average sentence length of about 17 tokens. The whole-corpus TTR is very low (7.06%), which suggests heavy repetition.
  • CAMeL Tools, a multi-dialect morphological tagger covering MSA, Egyptian, Gulf and Levantine, stands in for the missing Middle Arabic tools.
  • In total, 20,386 candidate sequences were extracted and scored against the 13 formulas, and 813 of them reached a score of at least 0.7.
  • Each formula contains at least one "static" anchor word that is fixed in form and position and rare elsewhere. For example, šdīd occurs 75 times inside formula (1) and only 8 times outside it.
  • Lexical substitutions that keep the meaning are tolerated, such as swapping one strong emotion for another. Substitutions that change the meaning belong to separate formulas tied to their own contexts, such as a variant used only after poems or one used only in battle scenes.

Interesting ideas

  • I1. Variants are scored on several linguistic layers separately (surface form, lemma, POS and so on) rather than on one string distance, which shows exactly which layer a variant differs on.
  • I2. A high-precision anchor word, fixed and rarely seen outside the expression, can seed the search for a flexible multiword expression.
  • I3. The paper sets the boundary of an expression by meaning: a variant that keeps the meaning is the same formula, and one that shifts it is a different construction. Distinct usage contexts serve as the evidence for that split.
  • I4. Ranked candidates are projected into a 2D vector space and coloured by score, giving a quick visual check of how well the ranking separates true variants from noise.
  • I5. An off-the-shelf tagger for related varieties is used knowingly as an imperfect substitute, with the expectation that it will handle the shared features.

For Mova

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

  • I1 (tokens): Per-layer comparison of form, lemma and tag could help diagnose lemmatizer and tagger disagreements on misspelled or non-standard English input.
  • I2 (world): Compiled idiom and construction entries could store an anchor token and allowed variation slots, so variable fixed expressions are matched explainably.
  • I3 (frames): When a frame-evoking phrase has lexical variants, the variant's typical context can help decide whether to map it to the same frame or a new one.
  • I4 (learn): Plotting silver-data candidates by agreement score against known gold examples could serve as a cheap sanity check before adopting induced repair rules.
  • I5 (tree): When converting or parsing non-standard English dialect text, existing parsers could be applied with explicit tracking of which features they are expected to miss.

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

🎓Sorbonne UniversitéParis · France32 works · 7,270 citations🔬Sens, Texte, Informatique, HistoireParis · France4 works · 14 citations

Also on the byline: Orient & Méditerranée · Institut français du Proche-Orient · CERES - Centre d'expérimentation en méthodes numériques pour les recherches en sciences humaines et sociales.

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.Semitic languages · 2,908 works

Cite

Cite as Julien Bezançon, Rimane Karam and Gaël Lejeune. 2025. Lost in Variation: An Unsupervised Methodology for Mining Lexico-syntactic Patterns in Middle Arabic Texts. Proceedings of the 4th Workshop on Arabic Corpus Linguistics (WACL-4).

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

@inproceedings{bezancon2025lost,
  title = {{Lost in Variation: An Unsupervised Methodology for Mining Lexico-syntactic Patterns in Middle Arabic Texts}},
  author = {Julien Bezançon, Rimane Karam and Gaël Lejeune},
  booktitle = {Proceedings of the 4th Workshop on Arabic Corpus Linguistics (WACL-4)},
  year = {2025},
  url = {https://aclanthology.org/2025.wacl-1.3/}
}

References

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

  1. Jean-Patrick Guillaume. 2004. Les scènes de bataille dans le roman de baybars: considérations sur le" style formulaire" dans la tradition épique arabe. Arabica, pages 55–76. find
  2. Ahmed Abdelali, Kareem Darwish, Nadir Durrani, and Hamdy Mubarak. 2016. Farasa: A fast and furious segmenter for Arabic. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations, pages 11–16, San Diego, California. Association for Computational Linguistics.
  3. N.Y. Habash. 2010. Introduction to Arabic Natural Language Processing. Synthesis digital library of engineering and computer science. Morgan & Claypool Publishers. ↗
  4. Faisal Al-Shargi, Aidan Kaplan, Ramy Eskander, Nizar Habash, and Owen Rambow. 2016. Morphologically annotated corpora and morphological analyzers for Moroccan and sanaani yemeni Arabic. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC’16), pages 1300–1306, Portorož, Slovenia. European Language Resources Association (ELRA).
  5. Graham Katz and Mona Diab. 2011. Introduction to the special issue on arabic computational linguistics. ACM Transactions on Asian Language Information Processing, 10(1). ↗
  6. Salam Khalifa, Sara Hassan, and Nizar Habash. 2017. A morphological analyzer for Gulf Arabic verbs. In Proceedings of the Third Arabic Natural Language Processing Workshop, pages 35–45, Valencia, Spain. Association for Computational Linguistics.
  7. Randah Alharbi, Walid Magdy, Kareem Darwish, Ahmed AbdelAli, and Hamdy Mubarak. 2018. Part-of-speech tagging for Arabic Gulf dialect using Bi-LSTM. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).
  8. Pierre Larcher. 2001. Moyen arabe et arabe moyen. Arabica, 48(Fasc. 4):578–609. ↗
  9. Jérôme Lentin. 2012. Reflections on middle arabic. High vs Low and Mixed Varieties: Domain, Status and Function across Time and Languages, edited by Gunvor Mejdell and Edzard Lutz, pages 32–51. find
  10. Anonymous. 2000–2021. Sīrat al-Malik al-Z.āhir Baybars. ḣasab al-riwāya al-šāmiyya. éd. G. Bohas, S. Diab, I. Hassan, K. Zakharia, Damas and Beyrouth, Presses de l’Ifpo. find
  11. Jérôme Lentin. 2008. Middle Arabic. Volume 3 of (Versteegh et al., 2008). find
  12. Julien Bezançon and Gaël Lejeune. 2023. Reconnaissance de défigements dans des tweets en français par des mesures de similarité sur des alignements textuels. In 30e Conférence sur le Traitement Automatique des Langues Naturelles, TALN, pages 56– 67, Paris, France. ATALA. find
  13. Salah Mejri. 2009. Figement, défigement et traduction. Problématique théorique. In figement, défigement et traduction (Fijación, desautomatización y traducción), pages 153–163. Universidad de Alicante. find
  14. Joshua Blau. 1982. The state of research in the field of the linguistic study of middle arabic. In Études de Linguistique Arabe, pages 187–203. Brill. ↗
  15. Ossama Obeid, Go Inoue, and Nizar Habash. 2022. Camelira: An Arabic multi-dialect morphological disambiguator. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 319–326, Abu Dhabi, UAE. Association for Computational Linguistics.
  16. Mathieu Constant, Gülşen Eryiǧit, Johanna Monti, Lonneke van der Plas, Carlos Ramisch, Michael Rosner, and Amalia Todirascu. 2017. Survey: Multiword Expression Processing: A Survey. Computational Linguistics, 43(4):837–892. Place: Cambridge, MA Publisher: MIT Press.
  17. Milman Parry. 1930. Studies in the epic technique of oral verse-making. i. homer and homeric style. Harvard Studies in Classical Philology, 41:73–147. ↗
  18. Kareem Darwish, Hamdy Mubarak, Ahmed Abdelali, Mohamed Eldesouki, Younes Samih, Randah Alharbi, Mohammed Attia, Walid Magdy, and Laura Kallmeyer. 2018. Multi-Dialect Arabic POS Tagging: A CRF Approach. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).
  19. Arfath Pasha, Mohamed Al-Badrashiny, Mona T Diab, Ahmed El Kholy, Ramy Eskander, Nizar Habash, Manoj Pooleery, Owen Rambow, and Ryan Roth. 2014. Madamira: A fast, comprehensive tool for morphological analysis and disambiguation of arabic. In Lrec, volume 14, pages 1094–1101. is transliterated as ġd.b in the present paper.
  20. Ivan A. Sag, Timothy Baldwin, Francis Bond, Ann Copestake, and Dan Flickinger. 2002. Multiword Expressions: A Pain in the Neck for NLP. In Computational Linguistics and Intelligent Text Processing, Lecture Notes in Computer Science, pages 1–15, Berlin, Heidelberg. Springer. For our experiment, we used sci-kit learn’s vectorization features with the following parameters: Younes Samih, Mohammed Attia, Mohamed Eldesouki, Ahmed Abdelali, Hamdy Mubarak, Laura Kallmeyer, and Kareem Darwish. 2017. A neural architecture for dialectal Arabic segmentation. In Proceedings of the Third Arabic Natural Language Processing Workshop, pages 46–54, Valencia, Spain. Association for Computational Linguistics. • CountV ectorizer • ngram_range = (1, 1)
  21. Kees Versteegh et al. 2008. Encyclopedia of Arabic Language and Linguistics, volume 3. Brill. • encoding = ”utf − 8” Liesbeth Zack and Arie Schippers. 2012. Middle Arabic and Mixed Arabic: Diachrony and Synchrony. Brill, Leiden, The Netherlands. • lowercase = T rue Nasser Zalmout, Alexander Erdmann, and Nizar Habash. 2018. Noise-robust morphological disambiguation for dialectal Arabic. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 953–964, New Orleans, Louisiana. Association for Computational Linguistics. • analyzer = lambda x : x.split(” ”) • stop_words = N one

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/bezancon-2025-lost-variation-unsupervised-methodology-mining/