Elamur Research⌕
☰
home / papers / 2026 / Beyond Atomic Characters: Glyph-Aware Sub-character Alignment for Low-Resource Multilingual OCR

Beyond Atomic Characters: Glyph-Aware Sub-character Alignment for Low-Resource Multilingual OCR

Paper · 2026 · Mengxiao Zhu, Haixu Chen, Jiu Sha et al.

Summary

OCR for low-resource scripts such as Tibetan, Mongolian and Vietnamese often fails on visually similar characters. These scripts have stacked forms, cursive joins or dense diacritics, and training data and language priors are scarce. The authors present BASA, an end-to-end OCR system that pairs the AIMV2 vision encoder with a Qwen2.5 language model. Between the two sits a new adapter, GAFA, which replaces the usual linear projector. GAFA uses learnable glyph prototypes that gather local stroke- and radical-level evidence, write it back into the visual tokens and merge it through a gate. Training proceeds in two stages: structure-focused pre-training on synthetic data, then end-to-end LoRA tuning on authentic documents. The synthetic data comes from a reverse-synthesis pipeline that produces component labels automatically, with no annotation cost. The authors also release BASA-Bench, which covers 11 low-resource languages and 23 real-world scenarios. BASA beats pipeline OCR systems and strong MLLMs on that benchmark, with the largest gains on scripts with complex composition, and stays stable on OmniDocBench. This suggests that structural grounding matters more than simply adding model capacity.

Key points

  • Architecture:
    • AIMV2 encoder, then GAFA, then Qwen2.5.
    • GAFA has three steps: depth-wise convolution to sharpen detail, cross-attention from prototypes to visual tokens, and reverse attention with gated fusion.
    • An auxiliary glyph loss supervises bag-of-components labels at image level, without bounding boxes.
  • Data:
    • Characters are split into components using linguistic rules and Unicode standards.
    • Synthetic corpora are reported at hundreds of thousands of samples per language, for example roughly 700k for Chinese (per the partly garbled data table).
    • These are combined with smaller sets of real documents.
  • Benchmarks: Evaluation uses BASA-Bench (11 languages, 23 scenarios) and OmniDocBench, with the metrics CER/WER, BLEU, ANLS and METEOR. BASA consistently improves over its internal baseline Qwen2.5-VL-7B.
  • Ablations:
    • Swapping GAFA for a plain linear projector leads to a large drop, most visibly in BLEU.
    • Removing the glyph loss hurts every metric.
    • One-stage training also underperforms.
    • Training on real data only gives the worst results; mixing synthetic and real data gives the best.
  • Limitations:
    • The method relies on well-defined sub-character decompositions, which are harder for highly cursive scripts.
    • Training cost is higher than with a standard adapter.
    • Handwritten and historical documents are not yet covered.
    • Model and benchmark are planned for release at github.com/NcutLLM/BASA.

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

🎓Beijing Institute of TechnologyBeijing · China43 works · 3,652 citations🎓Minzu University of ChinaBeijing · China7 works · 2 citations🎓North China University of TechnologyBeijing · China2 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.Multilingual & cross-lingual NLP · 882 worksLow-resource languages · 287 works

Cite

Cite as Mengxiao Zhu, Haixu Chen, Jiu Sha, Jie Liu and Ge Shi. 2026. Beyond Atomic Characters: Glyph-Aware Sub-character Alignment for Low-Resource Multilingual OCR. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.1392

@inproceedings{zhu2026beyond,
  title = {{Beyond Atomic Characters: Glyph-Aware Sub-character Alignment for Low-Resource Multilingual OCR}},
  author = {Mengxiao Zhu, Haixu Chen, Jiu Sha, Jie Liu and Ge Shi},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1392},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1392}
}

References

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

  1. Milind Agarwal and Antonios Anastasopoulos. 2024. A concise survey of ocr for low-resource languages. In Proceedings of the 4th Workshop on Natural Language Processing for Indigenous Languages of the Americas (AmericasNLP 2024), pages 88–102. find
  2. Isuri Anuradha, Chamila Liyanage, and Ruvan Weerasinghe. 2021. Estimating the effects of text genre, image resolution and algorithmic complexity needed for sinhala optical character recognition. International Journal on Advances in ICT for Emerging Regions (ICTer), 14(3). find
  3. Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, and 1 others. 2025. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923. — Qwen2.5-VL Technical Report
  4. Satanjeev Banerjee and Alon Lavie. 2005. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65–72.
  5. Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. 2019. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4291–4301. find
  6. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arX…
  7. Cheng Cui, Ting Sun, Manhui Lin, Tingquan Gao, Yubo Zhang, Jiaxuan Liu, Xueqing Wang, Zelun Zhang, Changda Zhou, Hongen Liu, and 1 others. 2025. Paddleocr 3.0 technical report. arXiv preprint arXiv:2507.05595. — [arXiv:2507.05595](https://arxiv.org/abs/2507.05595) find
  8. Hao Feng, Shu Wei, Xiang Fei, Wei Shi, Yingdong Han, Lei Liao, Jinghui Lu, Binghong Wu, Qi Liu, Chunhui Lin, and 1 others. 2025. Dolphin: Document image parsing via heterogeneous anchor prompting. arXiv preprint arXiv:2505.14059. find
  9. Enrico Fini, Mustafa Shukor, Xiujun Li, Philipp Dufter, Michal Klein, David Haldimann, Sai Aitharaju, Victor G. Turrisi da Costa, Louis Béthune, Zhe Gan, Alexander Toshev, Marcin Eichner, Moin Nabi, Yinfei Yang, Joshua M. Susskind, and Alaaeldin El-Nouby. 2025. Multimodal autoregressive pre-training… find
  10. Humane Intelligence Lab (HiLab), Xiaohongshu. 2024. dots.ocr: A multilingual document ocr system. https://dotsocr.xiaohongshu.com. Code available at https://github.com/ rednote-hilab/dots.ocr. ↗
  11. Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276.
  12. Pooja Jain, Kavita Taneja, and Harmunish Taneja. 2021. Which ocr toolset is good and why: A comparative study. Kuwait Journal of Science, 48(2). find
  13. Zhang Li, Yuliang Liu, Qiang Liu, Zhiyin Ma, Ziyang Zhang, Shuo Zhang, Zidun Guo, Jiarui Zhang, Xinyu Wang, and Xiang Bai. 2025. Monkeyocr: Document parsing with a structure-recognition-relation triplet paradigm. arXiv preprint arXiv:2506.05218. find
  14. Jit Yan Lim, Kian Ming Lim, Chin Poo Lee, and Yong Xuan Tan. 2024. Ssl-protonet: Self-supervised learning prototypical networks for few-shot learning. Expert Systems with Applications, 238:122173. find
  15. Fatemeh Naiemi, Vahid Ghods, and Hassan Khalesi. 2022. Scene text detection and recognition: a survey. Multimedia Tools and Applications, 81(14):20255– 20290. find
  16. ChatDOC Team. 2024a. Ocrflux. https://ocrflux. pdfparser.io/\#/blog. GitHub repository. find
  17. Qwen Team. 2024b. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2. find
  18. Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, Fukai Shang, and 1 others. 2024. Mineru: An open-source solution for precise document content extraction. arXiv preprint arXiv:2409.18839.
  19. Haoran Wei, Chenglong Liu, Jinyue Chen, Jia Wang, Lingyu Kong, Yanming Xu, Zheng Ge, Liang Zhao, Jianjian Sun, Yuang Peng, and 1 others. 2024. General ocr theory: Towards ocr-2.0 via a unified end-to-end model. arXiv preprint arXiv:2409.01704. — [arXiv:2409.01704](https://arxiv.org/abs/2409.01704) find
  20. Moonbin Yim, Yoonsik Kim, Hancheol Cho, and Sungrae Park. 2021. Synthtiger: Synthetic text image generator towards better text recognition models. In 16th International Conference on Document Analysis and Recognition, volume 12824, pages 109–124. find
  21. Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. 2020. Image-based table recognition: data, model, and evaluation. In European conference on computer vision, pages 564–580. Springer. find
  22. Xu Zhong, Jianbin Tang, and Antonio Jimeno Yepes. 2019. Publaynet: largest dataset ever for document layout analysis. In 2019 International conference on document analysis and recognition (ICDAR), pages 1015–1022. IEEE. find
  23. Meharuniza Nazeem, R Anitha, S Navaneeth, and 1 others. 2024. Open-source ocr libraries: A comprehensive study for low resource language. In Proceedings of the 21st International Conference on Natural Language Processing (ICON), pages 416–421. find
  24. Clemens Neudecker, Konstantin Baierer, Mike Gerber, Christian Clausner, Apostolos Antonacopoulos, and Stefan Pletschacher. 2021. A survey of ocr evaluation tools and metrics. In Proceedings of the 6th International Workshop on Historical Document Imaging and Processing, pages 13–18. find
  25. Linke Ouyang, Yuan Qu, Hongbin Zhou, Jiawei Zhu, Rui Zhang, Qunshu Lin, Bin Wang, Zhiyuan Zhao, Man Jiang, Xiaomeng Zhao, and 1 others. 2025. Omnidocbench: Benchmarking diverse pdf document parsing with comprehensive annotations. In Proceedings of the Computer Vision and Pattern Recognition Conferen… find
  26. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhu-2026-beyond-atomic-characters-glyph-aware/