Elamur Research⌕
☰
home / papers / 2026 / Character Iconicity vs. Arbitrariness: An Arabic NLP Perspective

Character Iconicity vs. Arbitrariness: An Arabic NLP Perspective

Summary

Many Arabic letters share one base shape (rasm) and differ only in their dots. Earlier work showed that undotted Arabic still works well for NLP. This paper asks why. Does it work because the historical rasm groupings carry real visual or orthographic structure, or because models adapt to any reduced alphabet that is applied consistently?

The authors keep the 19 undotted rasms as the output alphabet and the original group sizes, but assign letters to groups at random. They generated 2,000 such mappings under word-level and under character-level tokenization. From these they kept the four with the highest and lowest entropy. These mappings were compared with standard dotted and dotless text on:

  • predicting text with a language model
  • assigning categories to texts
  • tagging tokens within a sequence
  • translating between languages
  • restoration of the original script

They also tested several tokenizers. The random mappings performed competitively and often shrank vocabulary, OOV rates, model size and training cost. The authors conclude that, computationally, Arabic letter form is largely arbitrary: models depend on stable distributional patterns, not on the iconicity of letter shapes. This matters for designing compact script encodings.

Key points

  • Setup: 28 Arabic letters are collapsed onto 19 rasms. The four selected random mappings are HCE/LCE (highest/lowest character-level entropy) and HWE/LWE (highest/lowest word-level entropy), each chosen from 2,000 candidates.
  • Tokenizers: word, character, Farasa morphological, disjoint-letters, and SentencePiece BPE. Corpus statistics come from Wikipedia Arabic.
  • Models: all task models are modest (non-pretrained) neural networks:
    • decoder language model
    • GRU/BiGRU classifiers for poem meter, sentiment and news topic
    • BiLSTM taggers for POS and NER
    • encoder-decoder Transformer for MT
    • BiLSTM for restoration
  • Vocabulary under Farasa: about 1.13M for dotted, 891K for dotless, and 728K for LWE. Some mappings cut vocabulary by up to 50% with little loss. HWE and HCE stayed close to the baseline, while LCE and LWE compressed more with minor degradation.
  • Limitations: restoration uses only RNNs, the task set is limited, and pretrained LLMs were not tested because of compute cost.

Interesting ideas

  • I1. The null model is a random remapping that keeps the same group sizes, so the number of symbol collisions matches the meaningful grouping and only the "meaning" of the grouping changes. This is a clean way to ask whether a system exploits real structure or mere consistency. — For Mova (learn): permute tag or lemma-class labels while preserving class sizes, and use that as a negative control to confirm the parser or tagger gains come from linguistic signal rather than label stability.
  • I2. Instead of evaluating thousands of random variants, they score all 2,000 cheaply by entropy and run full experiments only on the extremes, which brackets the plausible range at low cost. — For Mova (learn): when ablating many random perturbations of training data or rules, rank them by a cheap statistic and fully evaluate only the best and worst cases.
  • I3. The pipeline deliberately works on a lossy, reduced representation and then recovers the full form with a learned restoration model. This trades exactness for smaller vocabularies and cheaper models. — For Mova (tokens): normalize noisy surface variants (case, spelling or punctuation variants) into a reduced alphabet for tagging and morphology, then restore the original form afterwards so every step stays traceable.
  • I4. Vocabulary size, V/N, entropy, and Zipf/Heaps curves are reported for every combination of mapping and tokenizer. This exposes interactions: under the disjoint-letters tokenizer, the LWE mapping actually enlarged the vocabulary instead of shrinking it. — For Mova (tokens): when changing the tokenizer or lemma scheme, publish the same corpus statistics grid so that unexpected vocabulary blow-ups are caught before training.

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: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2608.02935 · arXiv 2608.02935

@misc{ahmad2026character,
  title = {{Character Iconicity vs. Arbitrariness: An Arabic NLP Perspective}},
  author = {Ahmad;, Dorieh Alomari; Irfan and Al-shaibani, Maged S.},
  year = {2026},
  doi = {10.48550/arxiv.2608.02935},
  eprint = {2608.02935},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2608.02935}
}

References

41 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abandah, Gheith A, Mohammed Z Khedher, Mohammad R Abdel-Majeed, Hamdi M Mansour, Salma F Hulliel, and Lara M Bisharat. 2022. Classifying and diacritizing arabic poems using deep recurrent neural networks. Journal of King Saud University-Computer and Information Sciences, 34(6):3775–3788. find
  2. Abdelali, Ahmed, 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. find our card
  3. Al-Shaibani, Maged and Irfan Ahmad. 2023. Consonant is all you need: a compact representation of english text for efficient nlp. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 11578–11588. find
  4. Al-Shaibani, Maged S. and Irfan Ahmad. 2025. Dotless arabic text for natural language processing. Computational Linguistics, 51(2):557–598. find
  5. Alhathloul, Zainab and Irfan Ahmad. 2022. Automatic dottization of arabic text (rasms) using deep recurrent neural networks. Pattern Recognition Letters, 162:47–55. find
  6. Alomari, Dorieh and Irfan Ahmad. 2024. Exploring character trigrams for robust arabic text classification: a comparative analysis in the face of vocabulary expansion and misspelled words. IEEE Access, 12:57103–57116. find
  7. Aly, Mohamed and Amir Atiya. 2013. Labr: A large scale arabic book reviews dataset. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 494–498. find
  8. Alyafeai, Zaid, Maged S Al-Shaibani, and Moataz Ahmed. 2023. Ashaar: automatic analysis and generation of arabic poetry using deep learning approaches. arXiv preprint arXiv:2307.06218. ↗ our card
  9. Alyafeai, Zaid, Maged S Al-shaibani, Mustafa Ghaleb, and Irfan Ahmad. 2023. Evaluating various tokenizers for arabic text classification. Neural Processing Letters, 55(3):2911–2933. find our card
  10. Benajiba, Yassine, Paolo Rosso, and José Miguel Benedíruiz. 2007. Anersys: An arabic named entity recognition system based on maximum entropy. In Computational Linguistics and Intelligent Text Processing: 8th International Conference, CICLing 2007, pages 143–153. find
  11. Cettolo, Mauro, Marcello Federico, Luisa Bentivogli, Jan Niehues, Sebastian Stüker, Katsuitho Sudoh, Koichiro Yoshino, and Christian Federmann. 2017. Overview of the iwslt 2017 evaluation campaign. In Proceedings of the 14th International Workshop on Spoken Language Translation, pages 2–14. find our card
  12. Changizi, Mark A., Qiong Zhang, Hao Ye, and Shinsuke Shimojo. 2006. The structures of letters and symbols throughout human history are selected to match those found in objects in natural scenes. The American Naturalist, 167(5):E117–E139. find
  13. De Carolis, Léa, Egidio Marsico, Vincent Arnaud, and Christophe Coupé. 2018. Assessing sound symbolism: Investigating phonetic forms, visual shapes and letter fonts in an implicit bouba-kiki experimental paradigm. PLOS ONE, 13(12):e0208874. find
  14. Ding, Shuoyang, Adithya Renduchintala, and Kevin Duh. 2019. A call for prudent choice of subword merge operations in neural machine translation. In Proceedings of Machine Translation Summit XVII: Research Track, pages 204–213. find
  15. Dingemanse, Mark, Damián E. Blasi, Gary Lupyan, Morten H. Christiansen, and Padraic Monaghan. 2015. Arbitrariness, iconicity, and systematicity in language. Trends in Cognitive Sciences, 19(10):603–615. find
  16. Dingemanse, Mark, Will Schuerman, Eva Reinisch, Sylvia Tufvesson, and Holger Mitterer. 2016. What sound symbolism can and cannot do: Testing the iconicity of ideophones from five languages. Language, 92(2):e117–e133. find
  17. Einea, Omar, Ashraf Elnagar, and Ridhwan Al Debsi. 2019. Sanad: Single-label arabic news articles dataset for automatic text categorization. Data in brief, 25:104076. find
  18. Font-Clos, Francesc, Gemma Boleda, and Alvaro Corral. 2013. A scaling law beyond zipf’s law and its relation to heaps’ law. New Journal of Physics, 15(9):093033. find
  19. Hartmann, Frederik. 2019. Predicting historical phonetic features using deep neural networks: A case study of the phonetic system of proto-indo-european. In Proceedings of the 1st International Workshop on Computational Approaches to Historical Language Change, pages 98–108. find
  20. Heaps, HS. 1978. Information retrieval: Computational and theoretical aspects. Academic Press, Inc. find
  21. Hockett, Charles F. 1960. The origin of speech. Scientific American, 203(3):88–97. find
  22. Hodge, Gabrielle and Lindsay Ferrara. 2022. Iconicity as multimodal, polysemiotic, and plurifunctional. Frontiers in Psychology, 13:808896. find
  23. Kelly, Piers, James Winters, Helena Miton, and Olivier Morin. 2021. The predictable evolution of letter shapes: An emergent script of west africa recapitulates historical change in writing systems. Current Anthropology, 62(6):669–691. find
  24. Kudo, Taku and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66–71. find our card
  25. Li, Xinjian, Florian Metze, David R Mortensen, Shinji Watanabe, and Alan W Black. 2022. Zero-shot learning for grapheme to phoneme conversion with language ensemble. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2106–2115. find
  26. Loakman, Tyler, Yucheng Li, and Chenghua Lin. 2024. With ears to see and eyes to hear: Sound symbolism experiments with multimodal large language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 2849–2867. find
  27. Masrai, Ahmed and James Milton. 2016. How different is arabic from other languages? the relationship between word frequency and lexical coverage. Journal of Applied Linguistics and Language Research, 3(1):15–35. find
  28. Meister, Clara and Ryan Cotterell. 2021. Language model evaluation beyond perplexity. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5328–5339. find
  29. Miton, Helena and Olivier Morin. 2021. Graphic complexity in writing systems. Cognition, 214:104771. find
  30. Perniss, Pamela and Gabriella Vigliocco. 2014. The bridge of iconicity: from a world of experience to the experience of language. Philosophical Transactions of the Royal Society B, 369:20130300. find
  31. Porto, Alexander, Nikolai Huckle, Alexander Basalyga, Julio Santiago, and Alexander Kranjec. 2025. Glyph norming: Human and computational measurements of shape angularity in writing systems. Behavior Research Methods, 57:173. find
  32. Rom, Aviad and Kfir Bar. 2021. Supporting undotted arabic with pre-trained language models. In Proceedings of the 4th International Conference on Natural Language and Speech Processing (ICNLSP 2021), pages 89–94. find
  33. Roman, Claire and Philippe Meyer. 2024. Analysis of glyph and writing system similarities using siamese neural networks. In Proceedings of LT4HALA 2024 @ LREC-COLING 2024, pages 98–104. Saussure, Ferdinand de. 1983. Course in General Linguistics. Open Court, La Salle, IL. Originally published 1916. find
  34. Shannon, Claude Elwood. 1948. A mathematical theory of communication. The Bell system technical journal, 27(3):379–423. find
  35. Sidhu, David M. 2025. Does it matter that "bed" looks like a bed? orthographic iconicity in english. In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 47. find
  36. Smrz, Otakar, Viktor Bielickỳ, Iveta Kourilová, Jakub Krácmar, Jan Hajic, and Petr Zemánek. 2008. Prague arabic dependency treebank: A word on the million words. In Proceedings of the workshop on Arabic and local languages (LREC 2008), pages 16–23. find
  37. Taylor, Jack E., Rasmus Sinn, Cosimo Iaia, and Christian J. Fiebach. 2025. Beyond letters: Optimal transport as a model for sub-letter orthographic processing. Neurobiology of Language, 6:NOL.a.19. find
  38. Turoman, Nora and Suzy J Styles. 2017. Glyph guessing for ‘oo’and ‘ee’: Spatial frequency information in sound symbolic matching for ancient and unfamiliar scripts. Royal Society open science, 4(9):170882. find
  39. Zipf, George Kingsley. 2016. Human behavior and the principle of least effort: An introduction to human ecology. Ravenio books. find
  40. Zitouni, Imed, Jeffrey Sorensen, and Ruhi Sarikaya. 2006. Maximum entropy based restoration of arabic diacritics. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics, pages 577–584. find our card
  41. Ćwiek, Aleksandra et al. 2021. The bouba/kiki effect is robust across cultures and writing systems. Philosophical Transactions of the Royal Society B, 377:20200390. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/alomari-2026-character-iconicity-vs-arbitrariness-arabic/