---
title: "Cross-Lingual SynthDocs — Wang; 2025"
description: "Arabic OCR and document understanding (DU) lack large, richly annotated training data."
authors: ["Haneen Al-Homoud; Asma Ibrahim; Murtadha Al-Jubran; Fahad Al-Otaibi; Yazeed Al-Harbi; Daulet Toibazar; Kesen Wang;", "Pedro J. Moreno"]
year: 2025
venue: "arXiv:2511.04699"
source: "https://arxiv.org/abs/2511.04699"
doi: "10.48550/arxiv.2511.04699"
arxiv: "2511.04699"
licence: "CC BY 4.0 (arXiv)"
licence_url: "http://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/al-2025-homoud-cross-lingual-synthdocs-large-scale/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/al-2025-homoud-cross-lingual-synthdocs-large-scale/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/) / Cross-Lingual SynthDocs: A Large-Scale Synthetic Corpus for Any to Arabic OCR and Document Understanding

# Cross-Lingual SynthDocs: A Large-Scale Synthetic Corpus for Any to Arabic OCR and Document Understanding

**Authors:** Haneen Al-Homoud; Asma Ibrahim; Murtadha Al-Jubran; Fahad Al-Otaibi; Yazeed Al-Harbi; Daulet Toibazar; Kesen Wang;, [Pedro J. Moreno](https://md.research.elamur.ai/free/en/authors/pedro-j-moreno/) · **Year:** 2025 · **Venue:** arXiv:2511.04699 · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2511.04699) · [PDF](https://arxiv.org/pdf/2511.04699) · [DOI 10.48550/arxiv.2511.04699](https://doi.org/10.48550/arxiv.2511.04699) · [arXiv 2511.04699](https://arxiv.org/abs/2511.04699) · [OpenAlex](https://openalex.org/W7125966513)

**Citations:** 1 citations (openalex-snapshot · 2026-10-14)

## Summary
Arabic OCR and document understanding (DU) lack large, richly annotated training data. Existing Arabic sets are small, lean toward handwriting, or have thin annotations. Generic synthetic generators such as SynthDoc and DocSynth handle Arabic script poorly: right-to-left direction, diacritics and ligatures. The authors address this with Cross-Lingual SynthDocs, a synthetic corpus of more than 2.5 million samples covering text crops, full pages, tables and charts.

The key move is to reuse plentiful non-Arabic documents. Real scanned pages from the Industry Documents Library keep their layouts and backgrounds, while their text is machine-translated into Arabic and re-rendered inside the original boxes. Diacritized text crops and Arabic tables, both translated from arXiv and generated synthetically, are added to the mix.

Fine-tuning Qwen-2.5-VL on this corpus lowers WER and CER on several public Arabic OCR benchmarks. It also raises TEDS for tables and CharTeX for charts, narrowing the gap to Gemini. The work shows that realistic synthetic data built from cross-lingual sources can partly substitute for scarce human annotation in a low-resource script.

## Key points
- **Corpus contents:** over 2.5M samples, including about 1.5M textual items, roughly 270K fully annotated tables, and hundreds of thousands of charts derived from real data.
- **Text crops:** built from the arbml/tashkeela corpus. Diacritics are stripped at light, medium or heavy levels, or randomly inserted. Eastern Arabic and Western numerals are mixed at configurable rates, and crops are rendered with LaTeX in authentic Arabic fonts.
- **Full pages:**
  - Built from the IDL archive (over three million documents with line-level OCR JSON).
  - Lines are grouped into paragraphs with an adjacency graph. Two lines are linked when the baseline gap is at most the page's median line spacing and their horizontal overlap is at least 30% of the upper line's width. Each connected component becomes one paragraph.
  - Paragraphs are translated with Gemini 2.0 Flash, then re-rendered right-to-left with Pillow over the original page background.
- **Translation constraints:** abbreviations and symbols such as "$" or "%" must pass through untouched. Lines that were originally separate, like list items or headings, must come back as separate lines.
- **Tables:**
  - arXiv LaTeX tables keep their structure while an LLM translates the content into Arabic; they are then converted to HTML and re-rendered.
  - Synthetic tables come in two families: consistent-style (realistic layout) and random-style (augmentation).

## Interesting ideas
- **Per-page adaptive grouping:** paragraphs are recovered from noisy line boxes using a page-specific threshold (median line spacing) plus an overlap rule, with connected components as the units. This adapts to each document instead of relying on one global constant. — For Mova (tokens): a similar adjacency-and-components step could rebuild paragraphs and sentences from PDF-extracted or line-broken input before tokenization, so the parser does not receive broken sentence fragments.
- **Keep the skeleton, swap the content:** translating only the text while preserving a known structure (table layout, bounding boxes) yields new annotated data for free, because the gold structure carries over. — For Mova (learn): one could rewrite lexical content in sentences whose UD trees are already agreed, keeping the tree shape. This would produce extra silver training pairs whose structure is known by construction.
- **Protected tokens in generation:** the LLM is told to leave abbreviations and symbols unchanged and to return discontiguous lines separately, so the output stays aligned unit-by-unit with the source. — For Mova (back): when generating or translating from the meaning graph, numbers, units and named abbreviations could be locked as literal slots. Keeping a one-to-one mapping between graph units and output segments would make each output piece traceable.
- **Graded corruption levels:** diacritics are removed at distinct light, medium and heavy intensities rather than with a single noise setting. This gives controllable variation and makes it possible to measure robustness level by level. — For Mova (learn): parser and reader evaluation could use graded perturbations, such as dropped punctuation, lowercasing or spelling noise at fixed strengths. Plotting accuracy against strength would act as a stress-test and negative-control curve.
- **Two styles of synthetic data:** consistent-style tables imitate real documents, while random-style tables push diversity for augmentation. This separates realism from coverage as deliberate goals. — For Mova (learn): self-play or synthetic sentence generation could keep a realistic tier and a deliberately odd tier as separate pools. That would show whether gains come from realism or from variety.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/)

## Cite

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

```bibtex
@misc{wang2025cross,
  title = {{Cross-Lingual SynthDocs: A Large-Scale Synthetic Corpus for Any to Arabic OCR and Document Understanding}},
  author = {Wang;, Haneen Al-Homoud; Asma Ibrahim; Murtadha Al-Jubran; Fahad Al-Otaibi; Yazeed Al-Harbi; Daulet Toibazar; Kesen and Moreno, Pedro J.},
  year = {2025},
  doi = {10.48550/arxiv.2511.04699},
  eprint = {2511.04699},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2511.04699}
}
```

- [← previous Cross-Genre Learning for Old English Poetry POS Tagging](https://md.research.elamur.ai/free/en/papers/aclbulk-2025-miani-cross-genre-learning-old-english/)
- [next → DAPO: An Open-Source LLM Reinforcement Learning System at Scale](https://md.research.elamur.ai/free/en/papers/yu-2025-dapo-open-source-llm-reinforcement/)
- [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/)

## References

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

1. W. Zaghouani, N. Habash, O. Obeid, B. Mohit, H. Bouamor, and K. Oflazer, “Building an arabic machine translation post-edited corpus: Guidelines and annotation,” in Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC’16), 2016, pp. 1869–1876. [find](https://scholar.google.com/scholar?q=W%2E+Zaghouani%2C+N%2E+Habash%2C+O%2E+Obeid%2C+B%2E+Mohit%2C+H%2E+Bouamor%2C+and+K%2E+Oflazer%2C+%E2%80%9CBuilding+an+arabic+machine+translation+post%2Dedited+corpus%3A+Guidelines+and+annotation%2C%E2%80%9D+in+Proceedings+of+the+Tenth+Internation)
2. H. Al-Muhtaseb and R. Qahwaji, “Arabic optical character recognition: recent trends and future directions,” Applied Signal and Image Processing: Multidisciplinary Advancements, pp. 324–346, 2011. [find](https://scholar.google.com/scholar?q=H%2E+Al%2DMuhtaseb+and+R%2E+Qahwaji%2C+%E2%80%9CArabic+optical+character+recognition%3A+recent+trends+and+future+directions%2C%E2%80%9D+Applied+Signal+and+Image+Processing%3A+Multidisciplinary+Advancements%2C+pp%2E+324%E2%80%93346%2C+2011%2E)
3. A. W. Harley, A. Ufkes, and K. G. Derpanis, “Evaluation of deep convolutional nets for document image classification and retrieval,” in 2015 13th international conference on document analysis and recognition (ICDAR). IEEE, 2015, pp. 991–995. [find](https://scholar.google.com/scholar?q=A%2E+W%2E+Harley%2C+A%2E+Ufkes%2C+and+K%2E+G%2E+Derpanis%2C+%E2%80%9CEvaluation+of+deep+convolutional+nets+for+document+image+classification+and+retrieval%2C%E2%80%9D+in+2015+13th+international+conference+on+document+analysis+and+reco)
4. M. Mathew, D. Karatzas, and C. Jawahar, “Docvqa: A dataset for vqa on document images,” in Proceedings of the IEEE/CVF winter conference on applications of computer vision, 2021, pp. 2200–2209. [find](https://scholar.google.com/scholar?q=M%2E+Mathew%2C+D%2E+Karatzas%2C+and+C%2E+Jawahar%2C+%E2%80%9CDocvqa%3A+A+dataset+for+vqa+on+document+images%2C%E2%80%9D+in+Proceedings+of+the+IEEE%2FCVF+winter+conference+on+applications+of+computer+vision%2C+2021%2C+pp%2E+2200%E2%80%932209%2E) [our card](https://md.research.elamur.ai/free/en/papers/mathew-2020-docvqa-dataset-vqa-document-images/)
5. W. Antoun, F. Baly, and H. Hajj, “Arabert: Transformer-based model for arabic language understanding,” arXiv preprint arXiv:2003.00104, 2020. [↗](https://arxiv.org/abs/2003.00104) [our card](https://md.research.elamur.ai/free/en/papers/antoun-2020-arabert-transformer-based-model-arabic/)
6. O. Obeid, N. Zalmout, S. Khalifa, D. Taji, M. Oudah, B. Alhafni, G. Inoue, F. Eryani, A. Erdmann, and N. Habash, “Camel tools: An open source python toolkit for arabic natural language processing,” in Proceedings of the twelfth language resources and evaluation conference, 2020, pp. 7022–7032. [find](https://scholar.google.com/scholar?q=O%2E+Obeid%2C+N%2E+Zalmout%2C+S%2E+Khalifa%2C+D%2E+Taji%2C+M%2E+Oudah%2C+B%2E+Alhafni%2C+G%2E+Inoue%2C+F%2E+Eryani%2C+A%2E+Erdmann%2C+and+N%2E+Habash%2C+%E2%80%9CCamel+tools%3A+An+open+source+python+toolkit+for+arabic+natural+language+processing%2C%E2%80%9D+in) [our card](https://md.research.elamur.ai/free/en/papers/obeid-2020-camel-tools-open-source-python/)
7. S. A. Mahmoud, I. Ahmad, M. Alshayeb, W. G. Al-Khatib, M. T. Parvez, G. A. Fink, V. Märgner, and H. El Abed, “Khatt: Arabic offline handwritten text database,” in 2012 International conference on frontiers in handwriting recognition. IEEE, 2012, pp. 449–454. [find](https://scholar.google.com/scholar?q=S%2E+A%2E+Mahmoud%2C+I%2E+Ahmad%2C+M%2E+Alshayeb%2C+W%2E+G%2E+Al%2DKhatib%2C+M%2E+T%2E+Parvez%2C+G%2E+A%2E+Fink%2C+V%2E+Ma%CC%88rgner%2C+and+H%2E+El+Abed%2C+%E2%80%9CKhatt%3A+Arabic+offline+handwritten+text+database%2C%E2%80%9D+in+2012+International+conference+on+fro)
8. C. Ding, X. Liu, W. Tang, J. Li, X. Wang, R. Zhao, C.-T. Nguyen, and F. Tan, “Synthdoc: Bilingual documents synthesis for visual document understanding,” in Proceedings of the 2nd Workshop on Large Generative Models Meet Multimodal Applications, 2024, pp. 16–25. [find](https://scholar.google.com/scholar?q=C%2E+Ding%2C+X%2E+Liu%2C+W%2E+Tang%2C+J%2E+Li%2C+X%2E+Wang%2C+R%2E+Zhao%2C+C%2E%2DT%2E+Nguyen%2C+and+F%2E+Tan%2C+%E2%80%9CSynthdoc%3A+Bilingual+documents+synthesis+for+visual+document+understanding%2C%E2%80%9D+in+Proceedings+of+the+2nd+Workshop+on+Large+Ge)
9. S. Biswas, P. Riba, J. Lladós, and U. Pal, “Docsynth: a layout guided approach for controllable document image synthesis,” in International Conference on Document Analysis and Recognition. Springer, 2021, pp. 555–568. [find](https://scholar.google.com/scholar?q=S%2E+Biswas%2C+P%2E+Riba%2C+J%2E+Llado%CC%81s%2C+and+U%2E+Pal%2C+%E2%80%9CDocsynth%3A+a+layout+guided+approach+for+controllable+document+image+synthesis%2C%E2%80%9D+in+International+Conference+on+Document+Analysis+and+Recognition%2E+Springer%2C+)
10. O. Nacar, Y. Al-Habashi, S. Sibaee, A. Ammar, and W. Boulila, “Sard: A large-scale synthetic arabic ocr dataset for book-style text recognition,” arXiv preprint arXiv:2505.24600, 2025. [↗](https://arxiv.org/abs/2505.24600)
11. M. Rashad, “Arabic-nougat: Fine-tuning vision transformers for arabic ocr and markdown extraction,” arXiv preprint arXiv:2411.17835, 2024. [↗](https://arxiv.org/abs/2411.17835) [our card](https://md.research.elamur.ai/free/en/papers/rashad-2024-arabic-nougat-fine-tuning-vision/)
12. A. Wasfy, O. Nacar, A. Elkhateb, M. Reda, O. Elshehy, A. Ammar, and W. Boulila, “Qari-ocr: High-fidelity arabic text recognition through multimodal large language model adaptation,” arXiv preprint arXiv:2506.02295, 2025. [↗](https://arxiv.org/abs/2506.02295)
13. R. Smith, “An overview of the tesseract ocr engine,” in Ninth international conference on document analysis and recognition (ICDAR 2007), vol. 2. IEEE, 2007, pp. 629–633. [find](https://scholar.google.com/scholar?q=R%2E+Smith%2C+%E2%80%9CAn+overview+of+the+tesseract+ocr+engine%2C%E2%80%9D+in+Ninth+international+conference+on+document+analysis+and+recognition+%28ICDAR+2007%29%2C+vol%2E+2%2E+IEEE%2C+2007%2C+pp%2E+629%E2%80%93633%2E)
14. A. Facchin et al., “Whole-part-whole reading instruction in the teaching and learning of arabic as a foreign language at beginner level,” ANNALI DI CA’FOSCARI. SERIE ORIENTALE, vol. 56, pp. 171–190, 2020. [find](https://scholar.google.com/scholar?q=A%2E+Facchin+et+al%2E%2C+%E2%80%9CWhole%2Dpart%2Dwhole+reading+instruction+in+the+teaching+and+learning+of+arabic+as+a+foreign+language+at+beginner+level%2C%E2%80%9D+ANNALI+DI+CA%E2%80%99FOSCARI%2E+SERIE+ORIENTALE%2C+vol%2E+56%2C+pp%2E+171%E2%80%93190%2C+2)
15. S. Faizullah, M. S. Ayub, S. Hussain, and M. A. Khan, “A survey of ocr in arabic language: applications, techniques, and challenges,” Applied Sciences, vol. 13, no. 7, p. 4584, 2023. [find](https://scholar.google.com/scholar?q=S%2E+Faizullah%2C+M%2E+S%2E+Ayub%2C+S%2E+Hussain%2C+and+M%2E+A%2E+Khan%2C+%E2%80%9CA+survey+of+ocr+in+arabic+language%3A+applications%2C+techniques%2C+and+challenges%2C%E2%80%9D+Applied+Sciences%2C+vol%2E+13%2C+no%2E+7%2C+p%2E+4584%2C+2023%2E)
16. S. Guan, C. Xu, M. Lin, and D. Greene, “Effective synthetic data and test-time adaptation for OCR correction,” in Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Y. Al-Onaizan, M. Bansal, and Y.-N. Chen, Eds. Miami, Florida, USA: Association for Computational Linguistics, Nov. 2024, pp. 15 412–15 425. [Online]. Available: https://aclanthology.org/2024. emnlp-main.862/ [↗](https://aclanthology.org/2024)
17. P. Verma, “S2 chunking: A hybrid framework for document segmentation through integrated spatial and semantic analysis,” arXiv preprint arXiv:2501.05485, 2025. [↗](https://arxiv.org/abs/2501.05485)
18. L. McInnes, J. Healy, and S. Astels, “hdbscan: Hierarchical Density Based Spatial Clustering of Applications with Noise,” Journal of Open Source Software, vol. 2, no. 11, p. 205, 2017. [find](https://scholar.google.com/scholar?q=L%2E+McInnes%2C+J%2E+Healy%2C+and+S%2E+Astels%2C+%E2%80%9Chdbscan%3A+Hierarchical+Density+Based+Spatial+Clustering+of+Applications+with+Noise%2C%E2%80%9D+Journal+of+Open+Source+Software%2C+vol%2E+2%2C+no%2E+11%2C+p%2E+205%2C+2017%2E)
19. G. Binmakhashen and S. Mahmoud, “Document Layout Analysis: A Comprehensive Survey,” ACM Computing Surveys, vol. 52, no. 6, pp. 1–34, 2019. [find](https://scholar.google.com/scholar?q=G%2E+Binmakhashen+and+S%2E+Mahmoud%2C+%E2%80%9CDocument+Layout+Analysis%3A+A+Comprehensive+Survey%2C%E2%80%9D+ACM+Computing+Surveys%2C+vol%2E+52%2C+no%2E+6%2C+pp%2E+1%E2%80%9334%2C+2019%2E)
20. I. D. Library, “Idl-wds: Industry documents library word data set,” [https://huggingface.co/datasets/pixparse/idl-wds](https://huggingface.co/datasets/pixparse/idl-wds), 2024, version v1.0, accessed 15 July 2025. [↗](https://huggingface.co/datasets/pixparse/idl-wds)
21. A. F. Biten, R. Tito, L. Gomez, E. Valveny, and D. Karatzas, “Ocr-idl: Ocr annotations for industry document library dataset,” arXiv preprint arXiv:2202.12985, 2022. [↗](https://arxiv.org/abs/2202.12985)
22. G. Team, “Gemini: A family of highly capable multimodal models,” 2025. [Online]. Available: https://arxiv.org/abs/2312.11805 [↗](https://arxiv.org/abs/2312.11805) [our card](https://md.research.elamur.ai/free/en/papers/gemini-2023-team-gemini-family-highly-capable-multimodal/)
23. A. Heakl, A. Sohail, M. Ranjan, R. Hossam, G. Ahmed, M. El-Geish, O. Maher, Z. Shen, F. Khan, and S. Khan, “Kitab-bench: A comprehensive multi-domain benchmark for arabic ocr and document understanding,” arXiv preprint arXiv:2502.14949, 2025. [↗](https://arxiv.org/abs/2502.14949)
24. X. Zhong, E. ShafieiBavani, and A. Jimeno-Yepes, “Image-based table recognition: data, model, and evaluation,” CoRR, vol. abs/1911.10683, 2019. [Online]. Available: http://arxiv.org/abs/1911.10683 [↗](https://arxiv.org/abs/1911.10683)