---
title: "Oblevit at AR-MS NAKBA NLP 2026 Subtask 2: Hybrid"
description: "This system paper addresses line-level recognition of handwritten Arabic manuscripts in the AR-MS shared task (Subtask 2 of… 14 references."
authors: ["Reem Juhaysh", "Abuelgasim Sami Abusonoun", "Sara Ayad"]
year: 2026
venue: "Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026"
source: "https://aclanthology.org/2026.nakbanlp-1.35/"
doi: "10.63317/5nwur655ha5k"
licence: "CC BY-NC 4.0 (ELRA, LREC proceedings)"
licence_url: "https://creativecommons.org/licenses/by-nc/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/juhaysh-2026-oblevit-ar-ms-nakba-nlp/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/juhaysh-2026-oblevit-ar-ms-nakba-nlp/)

[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/38/) / Oblevit at AR-MS NAKBA NLP 2026 Subtask 2: Hybrid CNN–BiLSTM–CTC Framework with Linguistic Refinement for Arabic Handwritten Manuscript Recognition

# Oblevit at AR-MS NAKBA NLP 2026 Subtask 2: Hybrid CNN–BiLSTM–CTC Framework with Linguistic Refinement for Arabic Handwritten Manuscript Recognition

**Authors:** Reem Juhaysh, Abuelgasim Sami Abusonoun and Sara Ayad · **Year:** 2026 · **Venue:** [Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026](https://md.research.elamur.ai/free/en/venues/nakbanlp/) ([International Workshop on Nakba Narratives as Language Resources @ LREC 2026](https://md.research.elamur.ai/free/en/venues/nakbanlp/2026/)) · **Publisher:** [European Language Resources Association (ELRA)](https://md.research.elamur.ai/free/en/publishers/european-language-resources-association-elra/) · **Licence:** [CC BY-NC 4.0 (ELRA, LREC proceedings)](https://creativecommons.org/licenses/by-nc/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.nakbanlp-1.35/) · [PDF](https://aclanthology.org/2026.nakbanlp-1.35.pdf) · [DOI 10.63317/5nwur655ha5k](https://doi.org/10.63317/5nwur655ha5k) · [OpenAlex](https://openalex.org/W7166529330)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
This system paper addresses line-level recognition of handwritten Arabic manuscripts in the AR-MS shared task (Subtask 2 of Nakba-NLP 2026 at LREC 2026). The data are memoir pages tied to Palestinian memory. The script is hard for OCR for several reasons: letters join cursively, their shapes change with position, many letters differ only in their dots, and historical pages add faint ink and uneven spacing. The team used only the official training lines and transcriptions. They trained a classic pipeline: a convolutional encoder turns each line image into a left-to-right feature sequence, BiLSTM layers add context, and CTC removes the need for character segmentation. Most of the gain comes after the network. The output passes through beam search that also rewards dictionary matches, rule-based fixes for dot confusions, and snapping to a lexicon built from the training data. Outputs from several such pipelines are then merged line by line. The system placed second on the leaderboard with CER 0.0925 and WER 0.3268. This shows that a light recurrent model plus script-aware cleanup can compete in low-resource manuscript OCR.

## Key points
- Architecture: CNN feature extractor, then two BiLSTM layers (256 and 128 units), then a CTC output over 139 classes (138 symbols plus blank). Inputs are resized to 1536×64, normalized, and mirrored horizontally to account for right-to-left reading.
- Training: 90/10 train/validation split, 35 epochs, batch 64, Adam at 3×10⁻⁴, early stopping, and learning-rate reduction. No external data and no pretrained language model were used.
- Post-processing:
  - beam width 5, scored by CTC probability together with lexicon fit;
  - a dot-confusion mapping with context rules;
  - edit-distance correction against roughly 41,000 unique training-set words.
- Ensembling: the seven best submission files were aligned by filename and merged line by line, using agreement between them and consistency with the training data.
- Results on the blind test set (up to 2671 images): rank 2. The winner, Misraj Ai, scored CER 0.0790 / WER 0.2440, and the baseline scored 0.3683 / 0.6905. 3reeq had a better WER (0.2996) despite a worse CER. There was no full ablation, but removing the lexicon steps visibly hurt WER.

## Interesting ideas
- Instead of a generic spell checker, they built a correction table from the character pairs the model actually confuses (dot-only differences) and gated it with context rules, so each fix targets a known, systematic error.
- Decoding candidates are ranked by the network's probability combined with whether the words exist in a closed vocabulary, so symbolic validity joins the search instead of being applied only afterwards.
- Several pipelines with different decoders vote per line, and the chosen output is the one candidates agree on and that matches training-data patterns, rather than trusting any single pipeline.
- The authors note that some out-of-vocabulary outputs were real rare words, not errors, so snapping every unknown word to the nearest dictionary entry would damage correct text.
- The metrics disagree: one team beat them on WER while losing on CER, and the authors' fixes mostly affected word-level errors, so reporting both granularities exposes different failure types.

## For Mova

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

- **I1** (learn): Logging the parser's most frequent label or attachment confusions on dev data and turning each one into a context-conditioned repair rule fits the self-induced repair-rule stage, with every fix traceable to its confusion pair.
- **I2** (tree): Candidates in gparse or the perceptron parser's beam could be rescored with valency or subcategorization checks from verb classes, preferring trees that are grammatically licensed and recording why one candidate won.
- **I3** (learn): The same per-sentence or per-arc agreement among witness parsers could filter silver data, with a consistency-with-known-data check as a tiebreaker.
- **I4** (tokens): Typo normalization in the en crate should keep the raw form when the edit distance is large or the token looks like a plausible rare word, and should mark any substitution so it stays explainable.
- **I5** (learn): Evaluations should report both token-level and structure-level scores (for example LAS alongside frame or answer accuracy), because improvements and regressions can show up at only one level.

*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

- [Arab Open University](https://md.research.elamur.ai/free/en/universities/arab-open-university/) — University, Riyadh, Saudi Arabia · 3 works

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [👁️ Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) · 2,869 works · [⚙️ Engineering](https://md.research.elamur.ai/free/en/shelves/engineering/) · 133 works · [⚙️ Media Technology](https://md.research.elamur.ai/free/en/shelves/engineering/media-technology/) · 34 works

🧭 Topics (1): [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 2,908 works

## Cite

Cite as: Reem Juhaysh, Abuelgasim Sami Abusonoun and Sara Ayad. 2026. Oblevit at AR-MS NAKBA NLP 2026 Subtask 2: Hybrid CNN–BiLSTM–CTC Framework with Linguistic Refinement for Arabic Handwritten Manuscript Recognition. Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026.

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

```bibtex
@inproceedings{juhaysh2026oblevit,
  title = {{Oblevit at AR-MS NAKBA NLP 2026 Subtask 2: Hybrid CNN–BiLSTM–CTC Framework with Linguistic Refinement for Arabic Handwritten Manuscript Recognition}},
  author = {Reem Juhaysh, Abuelgasim Sami Abusonoun and Sara Ayad},
  booktitle = {Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026},
  year = {2026},
  doi = {10.63317/5nwur655ha5k},
  url = {https://doi.org/10.63317/5nwur655ha5k}
}
```

- [← previous OasisSimp: An Open-source Asian-English Sentence Simplification Dataset](https://md.research.elamur.ai/free/en/papers/liu-2026-oasissimp-open-source-asian-english/)
- [next → Obscuring Data Contamination Through Translation: Evidence from Arabic Corpora](https://md.research.elamur.ai/free/en/papers/abbas-2026-obscuring-data-contamination-through-translation/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/38/)

## References

14 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. H. Bunke, S. Bengio, and A. Vinciarelli. 2009. Offline recognition of unconstrained handwritten texts using hmms and statistical language models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(5):855–868. [find](https://scholar.google.com/scholar?q=H%2E+Bunke%2C+S%2E+Bengio%2C+and+A%2E+Vinciarelli%2E+2009%2E+Offline+recognition+of+unconstrained+handwritten+texts+using+hmms+and+statistical+language+models%2E+IEEE+Transactions+on+Pattern+Analysis+and+Machine+Inte)
2. A. Dosovitskiy, L. Beyer, A. Kolesnikov, and D. Weissenborn. 2020. [An image is worth 16×16 words: Transformers for image recognition at scale](https://md.research.elamur.ai/free/en/papers/dosovitskiy-2020-image-worth-16x16-words-transformers/). arXiv preprint arXiv:2010.11929. [↗](https://arxiv.org/abs/2010.11929)
3. A. Graves, M. Liwicki, S. Fernandez, R. Bertolami, H. Bunke, and J. Schmidhuber. 2009. A novel connectionist system for unconstrained handwriting recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(5):855– 868. [↗](https://doi.org/10.1109/tpami.2008.137)
4. A. Graves and J. Schmidhuber. 2008. Offline handwriting recognition with multidimensional recurrent neural networks. In Advances in Neural Information Processing Systems. [find](https://scholar.google.com/scholar?q=A%2E+Graves+and+J%2E+Schmidhuber%2E+2008%2E+Offline+handwriting+recognition+with+multidimensional+recurrent+neural+networks%2E+In+Advances+in+Neural+Information+Processing+Systems%2E)
5. H. Hamoud, A. A. Chamseddine, B. Shalash, F. B. Abid, M. Jarrar, C. A. Chakra, B. Ghanem, and F. A. Zaraket. 2026. [Nakba nlp 2026: Shared task on arabic handwritten manuscript understanding (palestine memory—omar al-saleh memoir)](https://md.research.elamur.ai/free/en/papers/hamoud-2026-nakba-nlp-2026-shared-task/). In Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources (Nakba-NLP 2026), co-located with LREC 2026, Palma, Mallorca, Spain.
6. D. Kass and E. Vats. 2022. Attentionhtr: Handwritten text recognition based on attention encoder–decoder networks. In International Conference on Document Analysis Systems. [find](https://scholar.google.com/scholar?q=D%2E+Kass+and+E%2E+Vats%2E+2022%2E+Attentionhtr%3A+Handwritten+text+recognition+based+on+attention+encoder%E2%80%93decoder+networks%2E+In+International+Conference+on+Document+Analysis+Systems%2E)
7. A. Krizhevsky, I. Sutskever, and G. Hinton. 2012. [Imagenet classification with deep convolutional neural networks](https://md.research.elamur.ai/free/en/papers/krizhevsky-2012-imagenet-classification-deep-convolutional/). In Advances in Neural Information Processing Systems, pages 1097–1105.
8. M. Li, T. Lv, L. Cui, Y. Lu, D. Florencio, C. Zhang, Z. Li, and F. Wei. 2021. Trocr: Transformer-based optical character recognition with pre-trained models. arXiv preprint arXiv:2109.10282. [↗](https://arxiv.org/abs/2109.10282)
9. J. Michael, R. Labahn, T. Gruning, and J. Zollner. 2019. Evaluating sequence-to-sequence models for handwritten text recognition. In International Conference on Document Analysis and Recognition, pages 1286–1293. [find](https://scholar.google.com/scholar?q=J%2E+Michael%2C+R%2E+Labahn%2C+T%2E+Gruning%2C+and+J%2E+Zollner%2E+2019%2E+Evaluating+sequence%2Dto%2Dsequence+models+for+handwritten+text+recognition%2E+In+International+Conference+on+Document+Analysis+and+Recognition%2C+page)
10. T. Ploetz and G. Fink. 2009. Markov models for offline handwriting recognition. International Journal on Document Analysis and Recognition, 12:269–298. [find](https://scholar.google.com/scholar?q=T%2E+Ploetz+and+G%2E+Fink%2E+2009%2E+Markov+models+for+offline+handwriting+recognition%2E+International+Journal+on+Document+Analysis+and+Recognition%2C+12%3A269%E2%80%93298%2E)
11. A. Poznanski and L. Wolf. 2016. Cnn-n-gram for handwriting word recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2305–2314. [find](https://scholar.google.com/scholar?q=A%2E+Poznanski+and+L%2E+Wolf%2E+2016%2E+Cnn%2Dn%2Dgram+for+handwriting+word+recognition%2E+In+Proceedings+of+the+IEEE+Conference+on+Computer+Vision+and+Pattern+Recognition%2C+pages+2305%E2%80%932314%2E)
12. C. Shorten and T. Khoshgoftaar. 2019. A survey on image data augmentation for deep learning. Journal of Big Data, 6(1). [↗](https://doi.org/10.1186/s40537-019-0197-0)
13. A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. Gomez, L. Kaiser, and I. Polosukhin. 2017. [Attention is all you need](https://md.research.elamur.ai/free/en/papers/vaswani-2017-attention-all-need/). In Advances in Neural Information Processing Systems, pages 6000–6010.
14. C. Wigington, S. Stewart, B. Davis, B. Barrett, B. Price, and S. Cohen. 2017. Data augmentation for recognition of handwritten words and lines using a cnn–lstm network. In International Conference on Document Analysis and Recognition, pages 639–645. [find](https://scholar.google.com/scholar?q=C%2E+Wigington%2C+S%2E+Stewart%2C+B%2E+Davis%2C+B%2E+Barrett%2C+B%2E+Price%2C+and+S%2E+Cohen%2E+2017%2E+Data+augmentation+for+recognition+of+handwritten+words+and+lines+using+a+cnn%E2%80%93lstm+network%2E+In+International+Conference+on+)

*Updated 1 November 2026*