Elamur Research⌕
☰
home / papers / 2026 / Data Matters: Looking for High-Quality Corpora to Build Robust and Reliable Models for Humanists

Data Matters: Looking for High-Quality Corpora to Build Robust and Reliable Models for Humanists

Paper · 2026 · Jaione Macicior-Mitxelena and Ana García-Serrano

Summary

Digitized Spanish historical newspapers are hard to process automatically. Scans are often degraded, fonts vary widely, pages use dense multi-column layouts, and the language differs from modern Spanish. The authors argue that the main bottleneck is the quality and coherence of the annotated data, not the model architecture. They study how corpus design, especially layout variability, affects OCR and layout analysis. Fine-tuning Tesseract on PastReader 2025 (IberLEF) data did not outperform the default Spanish model. In response, they built los101, a hand-transcribed subset of PastReader annotated in Transkribus with text, regions and links between regions. A second corpus, annotated for layout only and built from more regularly structured pages, was then used to isolate the effect of structural consistency. Tesseract, TrOCR, Granite and DocLayoutYOLO serve as diagnostic probes of corpus adequacy rather than as contributions in themselves. Heterogeneous data destabilized both transcription and segmentation, while consistently structured pages made layout detection noticeably more reliable. The study presents careful corpus construction as a prerequisite for trustworthy digital-humanities tools.

Key points

  • The PastReader 2025 corpus contains over 12,000 scanned pages from eight publications covering cultural, scientific, satirical and literary genres. Fine-tuning Tesseract with tesstrain raised CER from 0.36 (baseline spa v5.3) to 0.38, and BLEU dropped slightly.
  • los101 consists of 101 manually selected documents transcribed in Transkribus and split 80/10/10 into train, validation and test sets. Its guidelines keep the original orthography but drop purely paleographic features. Annotations also encode structure such as paragraphs, headings, marginalia, title–body links and image–caption links.
  • The second, layout-homogeneous corpus (Obispo et al., 2026, PageXML) covers pages from nine newspapers with more uniform column arrangements. It contains layout annotations only.
  • OCR was evaluated with CER, WER, BLEU and ROUGE-L. Layout was evaluated with mAP@0.5, mAP@0.5:0.95 and a newly proposed composite layout score that balances spatial and categorical accuracy.
  • The experiments run in two phases: (1) joint OCR and layout tests on the heterogeneous los101 corpus, and (2) layout-only DocLayoutYOLO training on each corpus separately.

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, topics and concepts

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSocial Sciences · 1,772 worksGeneral Social Sciences · 227 worksLiterature and Literary Theory · 519 works

Cite

Cite as Jaione Macicior-Mitxelena and Ana García-Serrano. 2026. Data Matters: Looking for High-Quality Corpora to Build Robust and Reliable Models for Humanists. Proceedings of the First Workshop on Creating Interoperable Corpora of Historical Newspapers.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/28atsf4aiory

@inproceedings{maciciormitxelena2026data,
  title = {{Data Matters: Looking for High-Quality Corpora to Build Robust and Reliable Models for Humanists}},
  author = {Jaione Macicior-Mitxelena and Ana García-Serrano},
  booktitle = {Proceedings of the First Workshop on Creating Interoperable Corpora of Historical Newspapers},
  year = {2026},
  doi = {10.63317/28atsf4aiory},
  url = {https://doi.org/10.63317/28atsf4aiory}
}

References

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

  1. Michele Alberti, Mathias Seuret, Vinaychandran Pondenkandath, Rolf Ingold, and Marcus Liwicki. 2017. Historical document image segmentation with lda-initialized deep neural networks. In Proceedings of the 4th international workshop on historical document imaging and processing, pages 95–100. find
  2. Mark J Hill and Simon Hengchen. 2019. Quantifying the impact of dirty ocr on historical text analysis: Eighteenth century collections online as a case study. Digital Scholarship in the Humanities, 34(4):825–843. find
  3. Austrian National Library Labs. 2024. Esperanto newspaper excerpts. GitLab repository. Accessed 2026-02-25. find
  4. Areej Jaber, Israa Bahati, and Paloma Martínez. 2025. Leveraging pre-trained embeddings in an ensemble machine learning approach for arabic sentiment analysis. Frontiers in Artificial Intelligence, Volume 8 - 2025. find
  5. Joan Benavent, Xaro Benavent, Esther de Ves, Ruben Granados, and Ana García-Serrano. 2010. Experiences at imageclef 2010 using CBIR and TBIR mixing information approaches. In CLEF 2010 LABs and Workshops, Notebook Papers, 22-23 September 2010, Padua, Italy, volume 1176 of CEUR Workshop Proceedings.… find
  6. Jaione Macicior Mitxelena and Ana Garcia-Serrano. 2026. From Paper To Pixel: Experimental Framework for Access to Historical Spanish Documents. https://github.com/jaionemacicior/ from-paper-to-pixel. Software/Code. ↗
  7. Kimmo Kettunen, Heikki Keskustalo, Sanna Kumpulainen, Tuula Pääkkönen, and Juha Rautiainen. 2022. Ocr quality affects perceived usefulness of historical newspaper clippings–a user study. arXiv preprint arXiv:2203.03557. — [arXiv:2203.03557](https://arxiv.org/abs/2203.03557) find
  8. Yoshua Bengio, Aaron Courville, and Pascal Vincent. 2013. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798–1828.
  9. Juan J. Lastra-Díaz, Josu Goikoetxea, Mohamed Ali Hadj Taieb, Ana Garcia-Serrano, Mohamed Ben Aouicha, Eneko Agirre, and David Sánchez. 2021. A large reproducible benchmark of ontology-based methods and word embeddings for word similarity. Information Systems, 96:101636. find
  10. Kai Chen, Mathias Seuret, Jean Hennebert, and Rolf Ingold. 2017. Convolutional neural networks for page segmentation of historical document images. In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), volume 1, pages 965–970. IEEE. find
  11. Kyunghyun Cho. 2014. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259.
  12. 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 Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 13094–13102. find
  13. Marten Düring, Estelle Bunout, and Daniele Guido. 2024. Transparent generosity. introducing the impresso interface for the exploration of semantically enriched historical newspapers. Historical Methods: A Journal of Quantitative and Interdisciplinary History, 57(1):20–40. find
  14. Bernhard Liebl and Manuel Burghardt. 2021. An evaluation of dnn architectures for page segmentation of historical newspapers. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 5153–5160. IEEE. find
  15. David Fleischhacker, Roman Kern, and Wolfgang Göderle. 2025. Enhancing ocr in historical documents with complex layouts through machine learning. International Journal on Digital Libraries, 26(1):3. find
  16. Jaione Macicior Mitxelena. 2025. Del papel al pixel: Experimentos para la digitalización de documentos históricos españoles. Master’s thesis, UNED: Universidad Nacional de Educación a Distancia, Madrid, Spain, September. Directed by Ana Garcia-Serrano. Máster en Tecnologías del Lenguaje. Grade: Sobr… find
  17. Enrique Garcia-Arias and Ana Garcia-Serrano. 2025. Creación de un modelo de descripciones de imágenes especializado en arqueología griega (pending edit). Procesamiento del Lenguaje Natural, 75(0). find
  18. M. Miguez Lamanuzzi and A. García Serrano. 2026. Annotation of historical texts for automatic processing in the gresel-uned project. In Congreso Internacional de Lingüística de Corpus (CILC 2026), Madrid. UAM. Aceptada. find
  19. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30.
  20. M. Miguez Lamanuzzi, J. Macicior Mitxelena, Y. Torterolo, R. Ortuño Casanova, and A. García Serrano. 2026. Guía de transcripción y anotación para prensa histórica. https://doi.org/10.5281/zenodo.19187624. — [doi:10.5281/zenodo.19187624](https://doi.org/10.5281/zenodo.19187624) ↗
  21. Zhiyuan Zhao, Hengrui Kang, Bin Wang, and Conghui He. 2024. Doclayout-yolo: Enhancing document layout analysis through diverse synthetic data and global-to-local adaptive perception. arXiv preprint arXiv:2410.12628. — [arXiv:2410.12628](https://arxiv.org/abs/2410.12628) find
  22. Antonio Moreno-Sandoval, Leonardo Campillos-Llanos, and Ana García-Serrano. 2024. Language resources in spanish for automatic text simplification across domains. find
  23. Wenzhen Zhu, Negin Sokhandan, Guang Yang, Sujitha Martin, and Suchitra Sathyanarayana. 2022. Docbed: A multi-stage ocr solution for documents with complex layouts. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 12643–12649. find
  24. Vahid Rezanezhad, Konstantin Baierer, Mike Gerber, Kai Labusch, and Clemens Neudecker. 2023. Document layout analysis with deep learning and heuristics. In Proceedings of the 7th International Workshop on Historical Document Imaging and Processing, pages 73–78. 9. find
  25. Eva Sánchez-Salido, Antonio Menta, and Ana García-Serrano. 2023. Seeking information in spanish historical newspapers: The case of diario de madrid (18th and 19th centuries). DHQ: Digital Humanities Quarterly, (4). find
  26. Language Resource References M. Miguez Lamanuzzi, A. García Serrano, Jaione Macicior Mitxelena, and Yanco Torterolo. 2025. Los101. [Data set]. find
  27. A. Montejo-Ráez, E. Sánchez Nogales, G. Expósito Álvarez, A. Ureña López, M. T. Martín-Valdivia, J. Collado-Montañez, I. Cabrera de Castro, M. V. Cantero Romero, A. García Serrano, R. Ortuño Casanova, and Y. A. Torterolo Orta. 2025. find
  28. Pastreader 2025. https://doi.org/ 10.5281/zenodo.15084265. [Data set]. — [doi:10.5281/zenodo.15084265](https://doi.org/10.5281/zenodo.15084265) find
  29. Eder Silva dos Santos Júnior, Thuanne Paixão, and Ana Beatriz Alvarez. 2025. Comparative performance of yolov8, yolov9, yolov10, and yolov11 for layout analysis of historical documents images. Applied Sciences, 15(6):3164. find
  30. Alexander Sergeev, Valeriya Goloviznina, Mikhail Melnichenko, and Evgeny Kotelnikov. 2025. Talking to data: Designing smart assistants for humanities databases. find
  31. F. Obispo, R. Ortuño Casanova, L. Garçon, Y. Seyeux, E. Sinardet, D. Villanueva Romero, and E. Vivó Capdevila. 2026. Corpus de prensa histórica con el layout marcado en pagexml. https://doi.org/10.5281/zenodo. 18774961. [Data set]. — [doi:10.5281/zenodo](https://doi.org/10.5281/zenodo) ↗
  32. Zejiang Shen, Ruochen Zhang, Melissa Dell, Benjamin Charles Germain Lee, Jacob Carlson, and Weining Li. 2021. Layoutparser: A unified toolkit for deep learning based document image analysis. arXiv preprint arXiv:2103.15348. — [arXiv:2103.15348](https://arxiv.org/abs/2103.15348) find
  33. Arno Simons, Michael Zichert, and Adrian Wüthrich. 2025. Large language models for history, philosophy, and sociology of science: Interpretive uses, methodological challenges, and critical perspectives. find
  34. Yanco Amor Torterolo-Orta, Jaione Macicior-Mitxelena, Marina Miguez-Lamanuzzi, and Ana García-Serrano. 2025. Transcribing spanish texts from the past: Experiments with transkribus, tesseract and granite. find
  35. Heidi JS Tworek. 2024. Digitized newspapers and the hidden transformation of history. The American Historical Review, 129(1):143–147. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mitxelena-2026-data-matters-looking-high-quality/