Elamur Research⌕
☰
home / papers / 2026 / LV-ROVER-MLT: Low-Resource Maltese OCR by Synthetic Fine-Tuning and Multi-Stream Arbitration

LV-ROVER-MLT: Low-Resource Maltese OCR by Synthetic Fine-Tuning and Multi-Stream Arbitration

Summary

Maltese has large text corpora and pretrained language models, but very few labelled page images exist for training OCR. The NOMOCRAT project provides only 57 annotated pages. The script is also hard to recognise. Fonts and pipelines can quietly replace the letters ċ, ġ, ħ and ż with their plain ASCII bases. The hyphen that attaches the definite article (il-kelb) is the same glyph as a line-break hyphen.

LV-ROVER-MLT fine-tunes Tesseract 5 on line images rendered from korpus_malti text. The rendering is tuned to match the competition crops in resolution, JPEG compression and hyphenation. At inference, five deterministic Tesseract configurations read each paragraph. They differ in language chain, stock versus fine-tuned model, and image scale. One anchor stream fixes the output, and the other streams may only swap in lexicon-eligible words, with a preference for diacritic-richer forms.

The system won the DocEng 2026 Maltese OCR competition with a held-out CER of 0.0074. That is less than half the error of the runner-up (0.0161) and of NOMOCRAT (0.0163). The author also releases a 36,803-pair paragraph corpus. The work shows that compact, CPU-only OCR can be strong for small languages.

Key points

  • Development-set results: the full pipeline reaches CER 0.00700, against 0.0234 for the organizers' fine-tuned Tesseract baseline. The fine-tuned anchor alone, with the rule-based joiner, scores 0.01605.
  • Recogniser training: 49,369 synthetic line crops, 2,000 iterations. Checkpoint 1400 was chosen on the 422-paragraph development set, which the author notes is not an independent validation.
  • Training text: drawn from korpus_malti v4.2 (about 470M tokens), keeping 11 of 19 configurations.
    • press_mt was dropped because its diacritic density is about 0.01, compared with 0.12–0.20 for the retained sources.
    • Maltese Wikipedia and the Maltese Universal Dependencies treebank serve as fallback text.
    • English code-switch material makes up 12% of the synthetic text.
  • Fonts and joiner: 68 permissively licensed fonts were screened for Maltese canary glyphs, and five candidates were rejected. The hyphen joiner achieves 99.51% round-trip accuracy on synthetic multi-line paragraphs.
  • Portability and corpus:
    • Luxembourgish CER improved from 0.25520 to 0.16927 (p < 0.0001).
    • Hungarian moved from 0.13543 to 0.13438 (p = 0.656, inconclusive).
    • The corpus contains 94 EUR-Lex and 36,709 Wikipedia pairs, filtered at 0.7 word coverage. It is born-digital rather than scanned and was built after the competition.

Interesting ideas

  • I1. Asymmetric arbitration: one anchor hypothesis fixes the structure, and the other streams can only substitute at aligned positions if a validity gate admits them; out-of-vocabulary anchor words survive. This keeps output well-formed while still harvesting fixes from weaker systems. — For Mova (learn): witness parsers producing silver data could follow this pattern, with a primary parser fixing the tree and witnesses only proposing per-token head or label swaps that pass UD validity checks.
  • I2. Hyphen tagging at generation time: the renderer labels every hyphen as soft, structural or compound, so the joiner can be scored on the genuinely ambiguous cases alone. This makes evaluation sensitive exactly where errors matter. — For Mova (tokens): synthetic test sentences could carry tags on ambiguous items such as 's (possessive vs "is") and hyphenated compounds, so the tokenizer and tagger are scored on hard cases separately.
  • I3. Ablation with opposing-regression checks: each post-processing rule is removed one at a time, and a gain on the development set counts only if no held-out synthetic set degrades. Benchmark-convention fixes are kept apart from real recognition gains. — For Mova (learn): self-induced repair rules could be admitted only after leave-one-out on development data shows no regression on independent negative-control sets, with formatting rules tracked separately from linguistic gains.
  • I4. Canary checks for silent data loss: fonts are screened for required glyphs, and corpus subsets are filtered by diacritic density, before any training. This catches data corruption that would otherwise poison labels unnoticed. — For Mova (tokens): input normalisation could run canary strings with curly quotes, accented names and dashes to detect pipelines that silently flatten characters before tagging and lemmatisation.
  • I5. Cheap diversity without retraining: different streams come from varying the inputs and configuration of one engine rather than training new models. This makes their errors complementary at almost no cost. — For Mova (tree): the perceptron parser and gparse could be run on input variants, such as alternative tagger outputs or sentence splits, to obtain extra low-cost witnesses for agreement voting.

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

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2607.00250 · arXiv 2607.00250

@misc{darmanin2026lv,
  title = {{LV-ROVER-MLT: Low-Resource Maltese OCR by Synthetic Fine-Tuning and Multi-Stream Arbitration}},
  author = {Darmanin, Adam},
  year = {2026},
  doi = {10.48550/arxiv.2607.00250},
  eprint = {2607.00250},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2607.00250}
}

References

22 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Merveilles Agbeti-Messan, Pierrick Tranouez, Stéphane Nicolas, Clément Chatelain, and Thierry Paquet. 2026. A Benchmark of State-Space Models vs. Transformers and BiLSTM-based Models for Historical Newspaper OCR. arXiv:2604.00725. arXiv:2604.00725 ↗
  2. Jeonghun Baek, Geewook Kim, Junyeop Lee, Sungrae Park, Dongyoon Han, Sangdoo Yun, Seong Joon Oh, and Hwalsuk Lee. 2019. What Is Wrong With Scene Text Recognition Model Comparisons? Dataset and Model Analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, Seoul, Korea (South), 4715–4723. doi:10.1109/ICCV.2019.00481 ↗
  3. Yoav Benjamini and Yosef Hochberg. 1995. Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. Journal of the Royal Statistical Society: Series B (Methodological) 57, 1 (1995), 289– 300. doi:10.1111/j.2517-6161.1995.tb02031.x ↗
  4. Claudia Borg and Albert Gatt. 2017. Morphological Analysis for the Maltese Language: The Challenges of a Hybrid System. In Proceedings of the https://github.com/adamd1985/lv-rover-mlt https://huggingface.co/radmada/lv-rover-mlt Third Arabic Natural Language Processing Workshop (WANLP). Association for Computational Linguistics, Valencia, Spain, 25–34. doi:10.18653/ v1/W17-1304 ↗ our card
  5. Bradley Efron and Robert J. Tibshirani. 1993. An Introduction to the Bootstrap. Number 57 in Monographs on Statistics and Applied Probability. Chapman and Hall/CRC, New York. doi:10.1201/9780429246593 ↗
  6. Jonathan G. Fiscus. 1997. A Post-Processing System to Yield Reduced Word Error Rates: Recognizer Output Voting Error Reduction (ROVER). In Proceedings of the IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU). IEEE, Santa Barbara, CA, 347–354. doi:10. 1109/ASRU.1997.659110 find
  7. Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. 2016. Synthetic Data for Text Localisation in Natural Images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, Las Vegas, NV, 2315–2324. doi:10.1109/CVPR.2016.254 ↗
  8. Philipp Koehn. 2004. Statistical Significance Tests for Machine Translation Evaluation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, Barcelona, Spain, 388–395. ACL Anthology has not assigned a DOI to this paper. https://aclanthology.org/W04-3250 ↗
  9. Jonathan Li, Zoltan Csaki, Nidhi Hiremath, Etash Guha, Fenglu Hong, Edward Ma, and Urmish Thakker. 2025. Synthetic Document Question Answering in Hungarian. arXiv:2505.23008. arXiv:2505.23008 ↗
  10. 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 AAAI Conference on Artificial Intelligence. find
  11. Daniel Lopresti and Jiangying Zhou. 1997. Using Consensus Sequence Voting to Correct OCR Errors. Computer Vision and Image Understanding 67, 1 (1997), 39–47. find
  12. William B. Lund and Eric K. Ringger. 2009. Improving Optical Character Recognition through Efficient Multiple System Alignment. In Proceedings of the 9th ACM/IEEE-CS Joint Conference on Digital Libraries (JCDL). ACM, 231–240. find
  13. Maltese Language Resource Server. 2025. malti: A Python library for processing text in the Maltese language. https://github.com/MLRS/malti. Accessed 2026-07-02. ↗
  14. Kurt Micallef, Albert Gatt, Marc Tanti, Lonneke van der Plas, and Claudia Borg. 2022. Pre-training Data Quality and Quantity for a Low-Resource Language: New Corpus and BERT Models for Maltese. In Proceedings of the Third Workshop on Deep Learning for Low-Resourced Natural Language Processing (DeepLo). Association for Computational Linguistics, Dublin, Ireland, 90–101. doi:10.18653/v1/2022.deeplo-1.10 ↗
  15. MLRS. 2026. MLRS/korpus_malti: General Maltese corpus, version 4.2. https://huggingface.co/datasets/MLRS/korpus_malti. HuggingFace dataset card. Accessed 2026-07-02. ↗
  16. Ray Smith. 2007. An Overview of the Tesseract OCR Engine. In Proceedings of the Ninth International Conference on Document Analysis and Recognition (ICDAR). IEEE, Curitiba, Brazil, 629–633. doi:10.1109/ ICDAR.2007.4376991 find
  17. Ray Smith, Daria Antonova, and Dar-Shyang Lee. 2009. Adapting the Tesseract Open Source OCR Engine for Multilingual OCR. In Proceedings of the International Workshop on Multilingual OCR. ACM, Barcelona, Spain, 1–8. doi:10.1145/1577802.1577804 ↗
  18. Bruno Stuner, Clément Chatelain, and Thierry Paquet. 2020. LV-ROVER: Lexicon Verified Recognizer Output Voting Error Reduction. International Journal on Document Analysis and Recognition (IJDAR) 23, 1 (2020), 1– 13. doi:10.1007/s10032-019-00344-9 ↗
  19. Marc Tanti, Alexandra Bonnici, and Stefania Cristina. 2026. NOMOCRAT: New Open Maltese OCR Annotated Text. https://www.systemsandcontrol.com/post/nomocrat-new-open-maltese-ocr-annotated-text. University of Malta project, Xjenza Malta grant REP-2024-057; project post, no formal proceedings venue. Post undated; the grant number reflects the award year, not the post’s date. The post itself references the then-upcoming DocEng 2026 competition, indicating 2026 authorship. Accessed 2026-07-04. ↗
  20. Nik Vaessen. 2025. jiwer: Evaluate automatic speech recognition systems. https://github.com/jitsi/jiwer. v4.0.0, Python package for CER/WER computation. Accessed 2026-07-02. ↗
  21. Christoph Wick, Christian Reul, and Frank Puppe. 2020. Calamari - A High-Performance TensorFlow-based Deep Learning Package for Optical Character Recognition. Digital Humanities Quarterly 14, 2 (2020). http://www.digitalhumanities.org/dhq/vol/14/2/000451/000451.html ↗
  22. Moonbin Yim, Yoonsik Kim, Han-Cheol Cho, and Sungrae Park. 2021. SynthTIGER: Synthetic Text Image Generator Towards Better Text Recognition Models. In Proceedings of the International Conference on Document Analysis and Recognition (ICDAR). Springer, Cham, 109–124. doi:10.1007/978-3-030-86337-1_8 ↗

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/darmanin-2026-lv-rover-mlt-low-resource/