---
title: "CANDLE: CTC-based Arabic Noisy-character — Alasmary 2026"
description: "Writers on social media often stretch Arabic words by repeating letters for emphasis."
authors: ["Faris Alasmary", "Taif Nono", "Orjuwan Zaafarani", "Kholood Al Tabash", "Ahmad Ghannam", "Anas Salamah", "Shouq Sadah", "Lahouari Ghouti"]
year: 2026
venue: "arXiv:2606.24758"
source: "https://arxiv.org/abs/2606.24758"
doi: "10.48550/arxiv.2606.24758"
arxiv: "2606.24758"
licence: "CC BY 4.0 (arXiv)"
licence_url: "http://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/alasmary-2026-candle-ctc-based-arabic-noisy/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/alasmary-2026-candle-ctc-based-arabic-noisy/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/3/) / CANDLE: CTC-based Arabic Noisy-character Deduplication using a Lightweight Encoder

# CANDLE: CTC-based Arabic Noisy-character Deduplication using a Lightweight Encoder

**Authors:** Faris Alasmary, Taif Nono, Orjuwan Zaafarani, Kholood Al Tabash, [Ahmad Ghannam](https://md.research.elamur.ai/free/en/authors/ahmad-ghannam/), Anas Salamah, Shouq Sadah, [Lahouari Ghouti](https://md.research.elamur.ai/free/en/authors/lahouari-ghouti/) · **Year:** 2026 · **Venue:** arXiv:2606.24758 · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2606.24758) · [PDF](https://arxiv.org/pdf/2606.24758) · [DOI 10.48550/arxiv.2606.24758](https://doi.org/10.48550/arxiv.2606.24758) · [arXiv 2606.24758](https://arxiv.org/abs/2606.24758) · [OpenAlex](https://openalex.org/W7165759561)

**Citations:** 0 citations (openalex-snapshot · 2026-10-14)

## Summary
Writers on social media often stretch Arabic words by repeating letters for emphasis. Some doubled letters, however, belong to the correct spelling, and collapsing them can turn one word into another, for example "kingdom" into "queen". CANDLE tackles this without rules, dictionaries or morphological analyzers. It treats deduplication as a Connectionist Temporal Classification (CTC) alignment problem on top of a pretrained character-level BERT encoder, and assumes that valid repeats never exceed two characters. Training text comes from clean news in the 1.5 Billion Words Arabic Corpus. The authors evaluate on three benchmarks:
- **NewsText:** clean newspaper text.
- **AmbigText:** hand-curated sentences with colliding forms.
- **WildSAText:** forum text with LLM-produced references.

The CTC model beats a classification baseline by a wide margin everywhere, reaching 5.37% sentence error rate on NewsText. A 2-layer distilled student stays close to the 6-layer teacher. Normalized text also needs up to 12.8% fewer tokens across Arabic LLM tokenizers, which lowers inference cost. Code and models are public.

## Key points
- The encoder is a character-based BERT with 6 layers, d_model = 512 and 16 heads. It was trained on Alriyadh and Alyaum news, about 480M words.
- Sentence error rate (SER), CTC versus classification baseline:

  | Benchmark | CTC | Classification |
  |---|---|---|
  | NewsText | 5.37% | 63.79% |
  | AmbigText | 23.08% | 69.96% |
  | WildSAText | 9.93% | 53.93% |

- The distilled 2-layer student scores 6.31%, 25.27% and 12.43% SER on the same benchmarks. On WildSAText sentences without repeats, its SER rises from 0.90% to 4.81%.
- The benchmarks differ in size and source:
  - AmbigText has 273 sentences (218 MSA, 55 Saudi dialect).
  - WildSAText has 142,280 sentences from the Saudi Forums Corpus. Its references were kept only where GPT 5.4 and Gemma4-it agreed.
- Normalization lowers tokenizer fertility by 4.8–12.8%, close to the gold-reference range of 4.7–12.7%.

## Interesting ideas
- Each benchmark reports the subset with repeated characters separately from the subset without them, so the authors can see whether the model damages text that needs no change. — For Mova (learn): Evaluation could include an identity control set where any edit by a repair rule or normalizer counts as an error, exposing over-eager rules.
- AmbigText is built from word pairs that collapse to the same reduced form but mean different things, so only context can resolve them. — For Mova (tokens): A small minimal-pair test set of forms that share a surface after normalization could check whether the English normalizer and lemmatizer keep legitimate doubled letters.
- Reference outputs for noisy text came from two different LLMs, and only the sentences where both agreed were kept. — For Mova (learn): This fits the witness-agreement approach to silver data, letting LLM teacher tiers serve as independent witnesses whose agreement filters training items.
- The authors define two explicit string operations, one collapsing runs to a single character and one doubling every character after that, and a hard cap of two valid repeats. Together these give a clear, inspectable vocabulary for describing the task. — For Mova (tokens): Elongations like "sooooo" could be reduced by a documented, explainable step before tagging, with the original surface kept for pragmatic cues such as emphasis.
- Besides accuracy, the authors measure how many tokens the normalized text needs. They use this as a cheap downstream proxy for usefulness. — For Mova (back): The text → graph → text compressor could report length or token savings next to fidelity, making the cost side of compression visible.

*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](https://md.research.elamur.ai/free/en/shelves/linguistics/)

## Cite

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

```bibtex
@misc{alasmary2026candle,
  title = {{CANDLE: CTC-based Arabic Noisy-character Deduplication using a Lightweight Encoder}},
  author = {Alasmary, Faris and Nono, Taif and Zaafarani, Orjuwan and Tabash, Kholood Al and Ghannam, Ahmad and Salamah, Anas and Sadah, Shouq and Ghouti, Lahouari},
  year = {2026},
  doi = {10.48550/arxiv.2606.24758},
  eprint = {2606.24758},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2606.24758}
}
```

- [← previous Can You Be More Explicit? A Task and Dataset on Explicitations of Implicit Meaning](https://md.research.elamur.ai/free/en/papers/zeidler-2026-can-you-be-more-explicit/)
- [next → CANVAS: A Multimodal Dataset of Chinese Textbook Images for Bias and Representation Analysis](https://md.research.elamur.ai/free/en/papers/zhu-2026-canvas-multimodal-dataset-chinese-textbook/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/3/)

## References

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

1. Ahmed bin Ali Al-Qalqashandi. 1987. Subh al-A’asha fi Sina’at al-Insha, volume 9. Dar al-Kutub al-’Ilmiyah, Beirut, Lebanon. [find](https://scholar.google.com/scholar?q=Ahmed+bin+Ali+Al%2DQalqashandi%2E+1987%2E+Subh+al%2DA%E2%80%99asha+fi+Sina%E2%80%99at+al%2DInsha%2C+volume+9%2E+Dar+al%2DKutub+al%2D%E2%80%99Ilmiyah%2C+Beirut%2C+Lebanon%2E)
2. Faris Alasmary, Orjuwan Zaafarani, and Ahmad Ghannam. 2024. Catt: Character-based arabic tashkeel transformer. In Proceedings of the Second Arabic Natural Language Processing Conference, pages 250–257. [find](https://scholar.google.com/scholar?q=Faris+Alasmary%2C+Orjuwan+Zaafarani%2C+and+Ahmad+Ghannam%2E+2024%2E+Catt%3A+Character%2Dbased+arabic+tashkeel+transformer%2E+In+Proceedings+of+the+Second+Arabic+Natural+Language+Processing+Conference%2C+pages+250%E2%80%93257)
3. Abubakar Ahmad Aliero, Sulaimon Adebayo Bashir, Hamzat Olanrewaju Aliyu, Amina Gogo Tafida, Umar Kangiwa Bashar, and Muhammad Dankolo Nasiru. 2023. Systematic review on text normalization techniques and its approach to non-standard words. International Journal of Computer Applications. [find](https://scholar.google.com/scholar?q=Abubakar+Ahmad+Aliero%2C+Sulaimon+Adebayo+Bashir%2C+Hamzat+Olanrewaju+Aliyu%2C+Amina+Gogo+Tafida%2C+Umar+Kangiwa+Bashar%2C+and+Muhammad+Dankolo+Nasiru%2E+2023%2E+Systematic+review+on+text+normalization+techniques+a)
4. Anab Maulana Barik, Rahmad Mahendra, and Mirna Adriani. 2019. Normalization of indonesian-english code-mixed twitter data. In Proceedings of the 5th Workshop on Noisy User-generated Text (W-NUT 2019), pages 417–424. [find](https://scholar.google.com/scholar?q=Anab+Maulana+Barik%2C+Rahmad+Mahendra%2C+and+Mirna+Adriani%2E+2019%2E+Normalization+of+indonesian%2Denglish+code%2Dmixed+twitter+data%2E+In+Proceedings+of+the+5th+Workshop+on+Noisy+User%2Dgenerated+Text+%28W%2DNUT+2019%29%2C)
5. Kareem Darwish, Walid Magdy, and Ahmed Mourad. 2012. Language processing for arabic microblog retrieval. In Proceedings of the 21st ACM international conference on Information and knowledge management, pages 2427–2430. [find](https://scholar.google.com/scholar?q=Kareem+Darwish%2C+Walid+Magdy%2C+and+Ahmed+Mourad%2E+2012%2E+Language+processing+for+arabic+microblog+retrieval%2E+In+Proceedings+of+the+21st+ACM+international+conference+on+Information+and+knowledge+management)
6. Seniz Demir and Berkay Topcu. 2022. Graph-based turkish text normalization and its impact on noisy text processing. Engineering Science and Technology, an International Journal, 35:101192. [find](https://scholar.google.com/scholar?q=Seniz+Demir+and+Berkay+Topcu%2E+2022%2E+Graph%2Dbased+turkish+text+normalization+and+its+impact+on+noisy+text+processing%2E+Engineering+Science+and+Technology%2C+an+International+Journal%2C+35%3A101192%2E)
7. Neelmay Desai and Meera Narvekar. 2015. Normalization of noisy text data. Procedia Computer Science, 45:127–132. [find](https://scholar.google.com/scholar?q=Neelmay+Desai+and+Meera+Narvekar%2E+2015%2E+Normalization+of+noisy+text+data%2E+Procedia+Computer+Science%2C+45%3A127%E2%80%93132%2E)
8. Ibrahim Abu El-Khair. 2016. 1.5 billion words arabic corpus. arXiv preprint arXiv:1611.04033. [↗](https://arxiv.org/abs/1611.04033)
9. Sanchit Gandhi, Patrick Von Platen, and Alexander M Rush. 2023. Distil-whisper: Robust knowledge distillation via large-scale pseudo labelling. arXiv preprint arXiv:2311.00430. [↗](https://arxiv.org/abs/2311.00430)
10. Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. 2006. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, pages 369–376. [find](https://scholar.google.com/scholar?q=Alex+Graves%2C+Santiago+Fern%C3%A1ndez%2C+Faustino+Gomez%2C+and+J%C3%BCrgen+Schmidhuber%2E+2006%2E+Connectionist+temporal+classification%3A+labelling+unsegmented+sequence+data+with+recurrent+neural+networks%2E+In+Proceedings)
11. Awni Y Hannun, Andrew L Maas, Daniel Jurafsky, and Andrew Y Ng. 2014. First-pass large vocabulary continuous speech recognition using bi-directional recurrent dnns. arXiv preprint arXiv:1408.2873. [↗](https://arxiv.org/abs/1408.2873)
12. Mohamed Osman Hegazi, Yasser Al-Dossari, Abdullah Al-Yahy, Abdulaziz Al-Sumari, and Anwer Hilal. 2021. Preprocessing arabic text on social media. Heliyon, 7(2). [find](https://scholar.google.com/scholar?q=Mohamed+Osman+Hegazi%2C+Yasser+Al%2DDossari%2C+Abdullah+Al%2DYahy%2C+Abdulaziz+Al%2DSumari%2C+and+Anwer+Hilal%2E+2021%2E+Preprocessing+arabic+text+on+social+media%2E+Heliyon%2C+7%282%29%2E)
13. Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531. [↗](https://arxiv.org/abs/1503.02531) [our card](https://md.research.elamur.ai/free/en/papers/hinton-2015-distilling-knowledge-neural-network/)
14. Tuan Anh Le, David Moeljadi, Yasuhide Miura, and Tomoko Ohkuma. 2016. Sentiment analysis for low resource languages: A study on informal indonesian tweets. In Proceedings of the 12th workshop on Asian language resources (ALR12), pages 123–131. [find](https://scholar.google.com/scholar?q=Tuan+Anh+Le%2C+David+Moeljadi%2C+Yasuhide+Miura%2C+and+Tomoko+Ohkuma%2E+2016%2E+Sentiment+analysis+for+low+resource+languages%3A+A+study+on+informal+indonesian+tweets%2E+In+Proceedings+of+the+12th+workshop+on+Asian)
15. Ilya Loshchilov and Frank Hutter. 2018. Decoupled weight decay regularization. In International Conference on Learning Representations. [find](https://scholar.google.com/scholar?q=Ilya+Loshchilov+and+Frank+Hutter%2E+2018%2E+Decoupled+weight+decay+regularization%2E+In+International+Conference+on+Learning+Representations%2E)
16. Ismini Lourentzou, Kabir Manghnani, and ChengXiang Zhai. 2019. Adapting sequence to sequence models for text normalization in social media. In Proceedings of the international AAAI conference on web and social media, volume 13, pages 335–345. [find](https://scholar.google.com/scholar?q=Ismini+Lourentzou%2C+Kabir+Manghnani%2C+and+ChengXiang+Zhai%2E+2019%2E+Adapting+sequence+to+sequence+models+for+text+normalization+in+social+media%2E+In+Proceedings+of+the+international+AAAI+conference+on+web+a)
17. Zainab Mansur, Nazlia Omar, Sabrina Tiun, and Eissa M Alshari. 2024. A normalization model for repeated letters in social media hate speech text based on rules and spelling correction. Plos one, 19(3):e0299652. [find](https://scholar.google.com/scholar?q=Zainab+Mansur%2C+Nazlia+Omar%2C+Sabrina+Tiun%2C+and+Eissa+M+Alshari%2E+2024%2E+A+normalization+model+for+repeated+letters+in+social+media+hate+speech+text+based+on+rules+and+spelling+correction%2E+Plos+one%2C+19%283%29)
18. Niko Partanen, Mika Hämäläinen, and Khalid Alnajjar. 2019. Dialect text normalization to normative standard finnish. In Proceedings of the 5th Workshop on Noisy User-generated Text (W-NUT 2019), pages 141–146. [find](https://scholar.google.com/scholar?q=Niko+Partanen%2C+Mika+H%C3%A4m%C3%A4l%C3%A4inen%2C+and+Khalid+Alnajjar%2E+2019%2E+Dialect+text+normalization+to+normative+standard+finnish%2E+In+Proceedings+of+the+5th+Workshop+on+Noisy+User%2Dgenerated+Text+%28W%2DNUT+2019%29%2C+pages)
19. Faisal Qarah. 2024. Saudibert: A large language model pretrained on saudi dialect corpora. arXiv preprint arXiv:2405.06239. [↗](https://arxiv.org/abs/2405.06239) [our card](https://md.research.elamur.ai/free/en/papers/qarah-2024-saudibert-large-language-model-pretrained/)
20. Jitendra Kumar Rout, Kim-Kwang Raymond Choo, Amiya Kumar Dash, Sambit Bakshi, Sanjay Kumar Jena, and Karen L Williams. 2018. A model for sentiment and emotion analysis of unstructured social media text. Electronic Commerce Research, 18(1):181– 199. [find](https://scholar.google.com/scholar?q=Jitendra+Kumar+Rout%2C+Kim%2DKwang+Raymond+Choo%2C+Amiya+Kumar+Dash%2C+Sambit+Bakshi%2C+Sanjay+Kumar+Jena%2C+and+Karen+L+Williams%2E+2018%2E+A+model+for+sentiment+and+emotion+analysis+of+unstructured+social+media+tex)
21. Sam Shleifer and Alexander M Rush. 2020. Pre-trained summarization distillation. arXiv preprint arXiv:2010.13002. [↗](https://arxiv.org/abs/2010.13002)
22. Phavanh Sosamphan. 2016. Snet: a statistical normalisation method for twitter. Master’s thesis, Unitec Institute of Technology. [find](https://scholar.google.com/scholar?q=Phavanh+Sosamphan%2E+2016%2E+Snet%3A+a+statistical+normalisation+method+for+twitter%2E+Master%E2%80%99s+thesis%2C+Unitec+Institute+of+Technology%2E)
23. Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. In Advances in neural information processing systems, pages 3104–3112. [find](https://scholar.google.com/scholar?q=Ilya+Sutskever%2C+Oriol+Vinyals%2C+and+Quoc+V+Le%2E+2014%2E+Sequence+to+sequence+learning+with+neural+networks%2E+In+Advances+in+neural+information+processing+systems%2C+pages+3104%E2%80%933112%2E)