Elamur Research⌕
☰
home / papers / 2026 / Latent Narratives at AR-MS NakbaNLP 2026: Reducing Character Errors in Arabic Manuscript Transcription: A CER Oriented System

Latent Narratives at AR-MS NakbaNLP 2026: Reducing Character Errors in Arabic Manuscript Transcription: A CER Oriented System

Paper · 2026 · Fatima Shaza, Sara Al-Desouky and Sarah Ayad

Summary

The authors tackle line-level transcription of historical handwritten Arabic for the AR-MS Subtask 2 of NakbaNLP 2026. This material is hard to read because of faded ink, uneven lighting, varied hands, and the script's heavy reliance on dots, hamza and tashkeel. Their system, Historic-Arabic-OCR, starts from Qari-OCR, an Arabic-adapted version of Qwen2-VL-2B-Instruct, and adds LoRA adapters. Training used about 16k lines from the Omar Al-Saleh manuscript collection: first on raw images, then on contrast-enhanced (CLAHE) images. Inference uses deterministic greedy decoding and a prompt that tells the model to transcribe literally.

Tuning the system for CER gave counterintuitive results. Beam search combined with dictionary correction made outputs worse, while contrast enhancement plus plain greedy decoding produced the lowest error. The final system reaches CER 0.10 and WER 0.28. Most remaining errors are small orthographic slips involving hamza, dots and word-final letter forms. The study suggests that visual preprocessing and literal decoding matter more than lexical "fixing" for faithful archival transcription. The model and code are publicly released.

Key points

  • Data: AR-MS Subtask 2 dataset from the Omar Al-Saleh collection, with roughly 15,962 training, 1,774 development and 2,671 test lines.
  • Setup:
    • LoRA with r=16, α=16 and no dropout, applied to the attention and feed-forward projections.
    • AdamW optimizer, learning rate 1e-5, effective batch size 16.
    • 4 epochs in total (2 on raw images, 2 on CLAHE images), trained on a single Tesla T4.
  • CER by configuration:
    • Fine-tuned baseline: 0.12
    • Beam search (5 beams) with dictionary correction and normalization: 0.14
    • CLAHE (clipLimit 2.0, 8×8 tiles) with greedy decoding: 0.10, the final system
  • Error analysis: in a 500-error sample, the most common confusions are bare alif versus hamza-bearing alif (60 and 56 cases). Ha/Kha, Ya/Alif maqsura and Ta marbuta/Ha also appear frequently.
  • Release: the weights are on Hugging Face as Shaza2004/Historic-Arabic-OCR, and the code is on GitHub (sara-aldesouky). Generalization to unseen calligraphic styles has not yet been evaluated.

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 Arabia3 works

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksComputer Vision and Pattern Recognition · 2,869 worksEngineering · 133 worksMedia Technology · 34 works

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

Cite

Cite as Fatima Shaza, Sara Al-Desouky and Sarah Ayad. 2026. Latent Narratives at AR-MS NakbaNLP 2026: Reducing Character Errors in Arabic Manuscript Transcription: A CER Oriented System. 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 10.63317/3gkghprpz2sb

@inproceedings{shaza2026latent,
  title = {{Latent Narratives at AR-MS NakbaNLP 2026: Reducing Character Errors in Arabic Manuscript Transcription: A CER Oriented System}},
  author = {Fatima Shaza, Sara Al-Desouky and Sarah Ayad},
  booktitle = {Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026},
  year = {2026},
  doi = {10.63317/3gkghprpz2sb},
  url = {https://doi.org/10.63317/3gkghprpz2sb}
}

References

11 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. R. Ahmad, S. Naz, M. Afzal, M. Liwicki, and A. Dengel. 2020. A deep learning based arabic script recognition system: Benchmark on KHATT. The International Arab Journal of Information Technology, 17(3). find
  2. Saleh Al-Sheikh, Masnizah Mohd, and L. Warlina. 2020. A review of arabic text recognition dataset. Asia-Pacific Journal of Information Technology and Multimedia, 9(1):69–81. ↗
  3. Naseem Alrobah and Saleh Albahli. 2022. Arabic handwritten recognition using deep learning: A survey. Arabian Journal for Science and Engineering, 47(8):9943–9963. ↗
  4. H. Althobaiti and C. Lu. 2017. A survey on arabic optical character recognition and an isolated handwritten arabic character recognition algorithm using encoded Freeman chain code. In Proceedings of the 2017 51st Annual Conference on Information Sciences and Systems (CISS), pages 1–6. IEEE. find
  5. Adrian Chan, Anupam Mijar, Mehreen Saeed, Chau-Wai Wong, and Akram Khater. 2024. Hatformer: Historic handwritten arabic text recognition with transformers. arXiv preprint arXiv:2410.02179. ↗
  6. A. El-Sawy, M. Loey, and H. El-Bakry. 2017. Arabic handwritten characters recognition using convolutional neural network. WSEAS Transactions on Computer Research, 5(1):11–19. find
  7. Hadi Hamoud, Ahmad Ali Chamseddine, Bilal Shalash, Firas Ben Abid, Mustafa Jarrar, Chadi Abou Chakra, Bernard Ghanem, and Fadi A. Zaraket. 2026. Nakba nlp 2026: Shared task on arabic handwritten manuscript understanding (palestine memory–omar al-saleh memoir). In Proceedings 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), Palma, Mallorca, Spain.
  8. Mahmoud Salaheldin Kasem, Mohamed Mahmoud, and Hyun-Soo Kang. 2025. Advancements and challenges in arabic optical character recognition: A comprehensive survey. ACM Computing Surveys (CSUR), 58(4). ↗
  9. Minghao Li, Tengchao Lv, Jingye Chen, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, and Furu Wei. 2023. TrOCR: Transformer-based optical character recognition with pre-trained models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 13094–13102. find
  10. 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. ↗
  11. Ahmed Wasfy, Omer Nacar, Abdelakreem Elkhateb, Mahmoud Reda, Omar Elshehy, Adel Ammar, and Wadii Boulila. 2025. Qari-OCR: High-fidelity arabic text recognition through multimodal large language model adaptation. arXiv preprint arXiv:2506.02295. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/shaza-2026-latent-narratives-ar-ms-nakbanlp/