Elamur Research⌕
☰
home / papers / 2026 / Miktub: A Manuscript Dataset of Historical Maltese for Handwritten Text Recognition

Miktub: A Manuscript Dataset of Historical Maltese for Handwritten Text Recognition

Paper · 2026 · Thomas Koppens and Claudia Borg

Summary

Historical Maltese has almost no handwritten resources, and the authors report that no public handwritten text recognition (HTR) dataset existed for it. They built Miktub from 35 page scans that come from the Special Collections of the University of Malta Library. Specialists transcribed the pages, and the pages were selected to cover many writers across several decades, along with archival difficulties such as fading, stains, bleed-through and struck-through text. Fully automatic line segmentation was not good enough on these mixed pages. The team therefore designed a semi-automatic workflow: horizontal projection profiling proposes the lines, light post-processing cleans them up, and a person refines the result. The line-level data comes in two annotation versions, one of which, Miktub-COR, is corrected and standardised to make it more consistent. Two public architectures were benchmarked on the data, HTR-VT and VAN. VAN trained on Miktub-COR performed best. The resource gives Maltese heritage digitisation a starting point and shows that annotation consistency has a real effect on recognition quality in low-resource settings.

Key points

  • Miktub (from the Maltese word for "written") contains 35 pages, mostly early 20th-century cursive with occasional block-letter titles, written by multiple authors over decades.
  • The best test result was 4.68% CER and 13.59% WER, achieved by VAN on the Miktub-COR variant.
  • The two baselines differ in design:
    • HTR-VT is a ViT-style model with a CNN tokenizer, span feature masking, SAM optimisation and CTC decoding.
    • VAN is a paragraph-level model whose vertical attention locates lines on its own, but it still needs line-level supervision.
  • Candidate pre-training corpora were compared, and LAM was judged the closest match. READ 2016 was considered too specialised, and Bentham was ruled out because it has a single author. IAM and RIMES were seen as too modern and too clean.
  • Maltese diacritic letters (ċ, ġ, ħ, ż) are a particular recognition risk because degradation in the scans can erase them. The data, scripts and splits are released at github.com/MLRS/Miktub.

Interesting ideas

  • I1. Rather than choosing between unreliable full automation and costly fully manual work, a cheap heuristic proposes the structure, simple rules tidy it, and humans fix only what remains.
  • I2. The dataset ships both a faithful transcription and a standardised one, so the effect of label consistency on training stability and error rates can be measured directly.
  • I3. The pre-training source is picked by matching domain properties (writer variety, time span, degradation, document genre) instead of simply taking the largest corpus.
  • I4. Evaluation pages were deliberately curated to include hard, realistic phenomena such as struck-through text and unseen writers, so that robustness gets tested rather than average-case accuracy.
  • I5. Small marks that change identity (here, diacritics) are named as the fragile point and get special attention during recognition.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): Witness-parser agreement could propose silver trees and repair rules could tidy them, so that only the leftover disagreements go to an LLM teacher or a human.
  • I2 (tree): Mova could train the parsers on both a raw treebank and its UD-dialect-normalised version, which would show how much noise comes from annotation conventions rather than from the model.
  • I3 (learn): Auxiliary treebanks or silver corpora for fables and school texts could be chosen by profiling register and genre against the target texts, not by size alone.
  • I4 (answer): Mova could build reader tests that intentionally include tricky constructions and unfamiliar story styles, which would expose failures that clean test sets hide.
  • I5 (tokens): Mova could track tagger and lemma errors that come from small surface cues such as apostrophes, hyphens and capitalisation as a separate error class.

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

🎓University of MaltaMsida · Malta32 works · 895 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArts and Humanities · 17,450 worksComputer Vision and Pattern Recognition · 3,486 worksEngineering · 188 worksLiterature and Literary Theory · 541 worksMedia Technology · 42 works

🧭 Topics 1Research subjects and languages the work is about.Semitic languages · 3,066 works

Cite

Cite as Thomas Koppens and Claudia Borg. 2026. Miktub: A Manuscript Dataset of Historical Maltese for Handwritten Text Recognition. 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 10.63317/4g7kcupa8ezp

@inproceedings{koppens2026miktub,
  title = {{Miktub: A Manuscript Dataset of Historical Maltese for Handwritten Text Recognition}},
  author = {Thomas Koppens and Claudia Borg},
  booktitle = {Proceedings of the Fourth Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA 2026) @ LREC 2026},
  year = {2026},
  doi = {10.63317/4g7kcupa8ezp},
  url = {https://doi.org/10.63317/4g7kcupa8ezp}
}

References

24 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Yoshua Bengio, Réjean Ducharme, Pascal Vincent, and Christian Jauvin. 2003. A neural probabilistic language model. Journal of machine learning research, 3(Feb):1137–1155.
  2. Théodore Bluche, Jérôome Louradour, and Ronaldo Messina. 2017. Scan, attend and read: End-to-end handwritten paragraph recognition with mdlstm attention. In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), volume 01, pages 1050– 1055. find
  3. Théodore Bluche and Ronaldo Messina. 2017. Gated convolutional recurrent neural networks for multilingual handwriting recognition. In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), volume 01, pages 646–651. find
  4. Claudia Borg and Albert Gatt. 2017. Morphological Analysis for the Maltese Language: The challenges of a hybrid system. In Proceedings of the Third Arabic Natural Language Processing Workshop, pages 25–34, Valencia, Spain. Association for Computational Linguistics. ↗
  5. Silvia Cascianelli, Vittorio Pippi, Maarand Martin, Marcella Cornia, Lorenzo Baraldi, Kermorvant Christopher, and Rita Cucchiara. 2022. The lam dataset: A novel benchmark for line-level handwritten text recognition. In International Conference on Pattern Recognition. find
  6. Stanley F Chen and Joshua Goodman. 1999. An empirical study of smoothing techniques for language modeling. Computer Speech & Language, 13(4):359–394. ↗
  7. Denis Coquenet, Clement Chatelain, and Thierry Paquet. 2023. End-to-End Handwritten Paragraph Text Recognition Using a Vertical Attention Network . IEEE Transactions on Pattern Analysis & Machine Intelligence, 45(01):508–524. find
  8. Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. 2006. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd International Conference on Machine Learning, ICML ’06, page 369–376, New York, NY, USA. Association for Computing Machinery. find
  9. Alex Graves and Jürgen Schmidhuber. 2008. Offline handwriting recognition with multidimensional recurrent neural networks. In Proceedings of the 22nd International Conference on Neural Information Processing Systems, NIPS’08, page 545–552, Red Hook, NY, USA. Curran Associates Inc. find
  10. Emmanuele Grosicki and Haikal El-Abed. 2011. Icdar 2011 - french handwriting recognition competition. In 2011 International Conference on Document Analysis and Recognition, pages 1459– 1463. find
  11. 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. Proceedings of the AAAI Conference on Artificial Intelligence, 37(11):13094–13102. find
  12. Yuting Li, Dexiong Chen, Tinglong Tang, and Xi Shen. 2025. Htr-vt: Handwritten text recognition with vision transformer. Pattern Recognition, 158:110967. ↗
  13. U.-V. Marti and H. Bunke. 2002. The iam-database: an english sentence database for offline handwriting recognition. International Journal on Document Analysis and Recognition, 5(1):39–46. S. ↗
  14. P. Muthu. 2022. Line segmentation in handwritten documents. https://github.com/muthuspark/line-segmentation-handwritten-doc. Last accessed: April 2026. ↗
  15. Arthur Flor de Sousa Neto, Byron Leite Dantas Bezerra, and Alejandro Héctor Toselli. 2020. Towards the natural language processing as spelling correction for offline handwritten text recognition systems. Applied Sciences, 10(21):7711. ↗
  16. Daniel Parres and Roberto Paredes. 2023. Fine-tuning vision encoder–decoder transformers for handwriting text recognition on historical documents. In Document Analysis and Recognition - ICDAR 2023, pages 253–268, Cham. Springer Nature Switzerland. find
  17. R. Plamondon and C.M. Privitera. 1999. The segmentation of cursive handwriting: an approach based on off-line recovery of the motor-temporal information. IEEE Transactions on Image Processing, 8(1):80–91. find
  18. R. Plamondon and S.N. Srihari. 2000. Online and off-line handwriting recognition: a comprehensive survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(1):63–84. find
  19. Daniel Povey, Arnab Ghoshal, Gilles Boulianne, Nagendra Goel, Mirko Hannemann, Yanmin Qian, Petr Schwarz, and Georg Stemmer. 2011. The kaldi speech recognition toolkit, workshop on automatic speech recognition and understanding. US IEEE Signal Processing Society, Hilton Waikoloa Village, Big Island, Hawaii. find
  20. Joan Puigcerver. 2017. Are multidimensional recurrent layers really necessary for handwritten text recognition? In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), volume 01, pages 67–72. find
  21. Shernobyl. 2024. Writer identification system. https://github.com/Shernobyl/Writer-Identification-System. Last accessed: April 2026. ↗
  22. Joan Andreu Sánchez, Verónica Romero, Alejandro H. Toselli, and Enrique Vidal. 2016. Icfhr2016 competition on handwritten text recognition on the read dataset. In 2016 15th International Conference on Frontiers in Handwriting Recognition (ICFHR), pages 630–635. find
  23. Joan Andreu Sánchez, Alejandro H. Toselli, Verónica Romero, and Enrique Vidal. 2015. Icdar 2015 competition htrts: Handwritten text recognition on the transcriptorium dataset. In 2015 13th International Conference on Document Analysis and Recognition (ICDAR), pages 1166–1170. find
  24. Long Wu, Ta Li, Li Wang, and Yonghong Yan. 2019. Improving hybrid ctc/attention architecture with time-restricted self-attention ctc for end-to-end speech recognition. Applied Sciences, 9(21):4639. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/koppens-2026-miktub-manuscript-dataset-historical-maltese/