Elamur Research⌕
☰
home / papers / 2026 / A Parameter-Efficient and Data-Centric Framework for Ancient Chinese Text Recognition and Layout Analysis

A Parameter-Efficient and Data-Centric Framework for Ancient Chinese Text Recognition and Layout Analysis

Paper · 2026 · Yuchun Meng

Summary

Digitizing historical Chinese books is difficult because of irregular layouts, mixed printed and handwritten styles, and frequent variant characters. This system paper describes a Closed Track entry to the EvaHan 2026 shared task, which covers printed character recognition (Task A), layout analysis (Task B) and handwritten character recognition (Task C). The author uses the same backbone as the organizers' baseline, Qwen2.5-VL-7B-Instruct, and adds a data-centric layer on top of it. Dataset geometry (aspect ratios, orientation) is profiled first, and horizontally aligned handwritten slips are rotated so that all inputs share a vertical orientation. Prompts then encode fixed dataset-level priors: slip shape, character-count range, top-to-bottom reading order and an instruction to keep variant forms. Three independent rank-16 LoRA adapters, one per task, are trained on a single 24GB GPU, with 4-bit QLoRA used for layout. Against the rank-32 baseline, the system halves CER on printed text and markedly improves layout classification. Handwriting scores stay slightly below the baseline. The results suggest that input normalization and structured prompts can matter more than adapter capacity, although no ablation study isolates the individual contributions.

Key points

  • Task A (printed): CER dropped from 0.0618 to 0.0305, F1 rose to 0.9719, and the Overall score reached 0.9703 (baseline 0.9397).
  • Task B (layout): Macro F1 rose from 0.1530 to 0.4162 (+172%), mAP from 0.2006 to 0.2749 (+37%), and Avg Match IoU reached 0.6740. The author attributes the low absolute mAP to coordinates being discretized into 1000 bins and to precision loss from 4-bit quantization.
  • Task C (handwritten): F1 was 0.9018 versus the baseline's 0.9099. Scores were higher when variant characters were counted than when they were excluded (F1 0.8873), which the author reads as a "semantic correction" effect, mapping variant glyphs to their normative equivalents.
  • Training setup: Alibaba Cloud PAI-DSW with an NVIDIA A10 (24GB). LoRA used rank 16, alpha 32 and dropout 0.05 on the attention and MLP projections, with learning rate 2e-4, effective batch size 16, 3 epochs and roughly 1.2M max pixels. Training ran in the Swift framework.
  • Stated limitations: no ablation studies, no targeted handling of degraded glyphs, and separate per-task adapters instead of a single unified model.

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

🎓Nanjing UniversityNanjing · China74 works · 7,814 citations

 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

Cite

Cite as Yuchun Meng. 2026. A Parameter-Efficient and Data-Centric Framework for Ancient Chinese Text Recognition and Layout Analysis. Proceedings of the Fourth Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA 2026) @ LREC 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/4inrcp772rid

@inproceedings{meng2026parameter,
  title = {{A Parameter-Efficient and Data-Centric Framework for Ancient Chinese Text Recognition and Layout Analysis}},
  author = {Yuchun Meng},
  booktitle = {Proceedings of the Fourth Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA 2026) @ LREC 2026},
  year = {2026},
  doi = {10.63317/4inrcp772rid},
  url = {https://doi.org/10.63317/4inrcp772rid}
}

References

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

  1. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., ... & Schmidt, L. (2023). OpenFlamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390.
  2. Chen, T., Saxena, S., Li, L., Fleet, D. J., & Hinton, G. (2021). Pix2seq: A language modeling framework for object detection. International Conference on Learning Representations. find
  3. Dettmers, T., Pagnoni, A., Holtzman, A., & Zettlemoyer, L. (2023). Qlora: Efficient finetuning of quantized llms. Advances in Neural Information Processing Systems, 36.
  4. Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., ... & Chen, W. (2021). LoRA: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685. \#\# 5. Conclusion

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/meng-2026-parameter-efficient-data-centric-framework/