Elamur Research⌕
☰
home / papers / 2026 / Al-Warraq at AR-MS NAKBA-NLP 2026: Adapting Vision-Language and Transformer Models for Automatic Manuscript OCR/HTR

Al-Warraq at AR-MS NAKBA-NLP 2026: Adapting Vision-Language and Transformer Models for Automatic Manuscript OCR/HTR

Paper · 2026 · Ahmad Edris Youssef, Aya Hafiz Faris, Alhasan Hamood et al.

Summary

This paper describes the Al-Warraq team's system for the NAKBA NLP 2026 shared task on Arabic manuscript OCR/HTR. The task is to convert cropped handwritten line images into Arabic text. It is hard because the script is cursive, letter shapes depend on position, dots separate many letters, and historical pages are often degraded. The team compared several OCR/HTR candidates, including Transformer and vision-language options. They dropped an early English-oriented model that handled Arabic poorly and fine-tuned the strongest candidate on roughly 16,000 training line images. Error analysis found that characters were recognised much better than whole words, with merged words and misplaced spaces as typical errors. To address this, the team added rotation-based test-time augmentation with voting and a rule-based cleanup step. The final submission ranked fifth, with CER 0.1142 and WER 0.378. The work shows that simple, targeted inference and post-processing steps help with Arabic heritage transcription. It also shows that word segmentation remains the main bottleneck.

Key points

  • Final scores: CER 0.1142, WER 0.3780, LER 0.3468, placing 5th. Training took more than 100 hours in total.
  • Development accuracy rose from about 30% to about 80% after the team swapped unsuitable models and kept training. Checkpoint-based resumption guarded against hardware limits and interruptions.
  • Reloading the trained model raised a KeyError because the tokenizer and dictionary files were missing. The team fixed this by pairing the base model's processor files with the fine-tuned weights.
  • Test-time augmentation ran each line at 0°, +2° and −2° and kept the most frequent output. Post-processing normalised letter forms, removed spaces before punctuation and split some merged words.
  • Cutting beam width from 4 to 2 reduced inference time from about 15 to about 7 hours. This let the team submit just before the deadline.

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 UniversityRiyadh · Saudi Arabia4 works

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksComputer Vision and Pattern Recognition · 3,486 worksEngineering · 188 worksMedia Technology · 42 works

🧭 Topics 2Research subjects and languages the work is about.Semitic languages · 3,066 worksVision & language · 811 works

Cite

Cite as Ahmad Edris Youssef, Aya Hafiz Faris, Alhasan Hamood, Zainab Kamil, Jana Alqasem and SARA Ali hamed”. 2026. Al-Warraq at AR-MS NAKBA-NLP 2026: Adapting Vision-Language and Transformer Models for Automatic Manuscript OCR/HTR. 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. DOI 10.63317/2fv6385csb36

@inproceedings{youssef2026al,
  title = {{Al-Warraq at AR-MS NAKBA-NLP 2026: Adapting Vision-Language and Transformer Models for Automatic Manuscript OCR/HTR}},
  author = {Ahmad Edris Youssef, Aya Hafiz Faris, Alhasan Hamood, Zainab Kamil, Jana Alqasem and SARA Ali hamed”},
  booktitle = {Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026},
  year = {2026},
  doi = {10.63317/2fv6385csb36},
  url = {https://doi.org/10.63317/2fv6385csb36}
}

References

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

  1. Carlos Garrido-Munoz, Antonio Rios-Vila, and Jorge Calvo-Zaragoza. 2025. Handwritten text recognition: A survey. arXiv preprint arXiv:2502.08417. ↗
  2. Minghao Li, Tengchao Lv, Jingye Chen, Lei Cui, Yi- juan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, and Furu Wei. 2021. Trocr: Transformer-based optical character recognition with pre-trained models. arXiv preprint arXiv:2109.10282. ↗
  3. Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. 2024. Deepseek-vl: Towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525. ↗
  4. Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. 2024. Qwen2- vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191 ↗
  5. Sarah Al Hamed, Ahmad Edris Youssef, Aya Hafiz Faris, Alhasan Hamood, Zainab Kamil, Jana Mukhles Alqasem, Adapting Vision-Language and Transformer Models for Automatic Manuscript OCR/HTR of the 2nd International Workshop on Nakba Narratives as Language Resources (Nakba-NLP 2026), co-located with the Language Resources and Evaluation Conference (LREC), 2026. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/youssef-2026-al-warraq-ar-ms-nakba/