Elamur Research⌕
☰
home / papers / 2026 / Automatic Generation of Graded Texts in Old Church Slavonic

Automatic Generation of Graded Texts in Old Church Slavonic

Paper · 2026 · Iglika Nikolova-Stoupak, Gaël Lejeune, Aliona Shestakova-Stukun et al.

Summary

Graded readers grounded in comprehensible input have been developed for a few classical languages, notably Latin and Ancient Greek. No comparable material exists for Old Church Slavonic (OCS), a low-resourced language with considerable dialectal and diachronic variation. The authors explore whether such readers can be produced (semi-)automatically.

First, they profile measurable textual features of six established readers in Latin, Greek and Biblical Hebrew to obtain level-specific reference values. Next, they adapt two OCS source texts: 'Genesis: 1' toward A1 and 'Saint George and the Dragon' toward B1. For this they use mT5, trained on multilingual professional graded readers, and GPT-5 with one-shot prompting. The outputs are judged by expert reading and by comparison with the reference profiles.

mT5 failed at the task. GPT-5 produced pedagogically sensible text whose level differences follow the trends of professional readers. Its output still contained case errors, questionable word choices and interference from later Church Slavonic and Russian. The manually corrected texts are released, giving teachers a faster route to OCS reading material.

Key points

  • The reference corpus consists of six readers: Lingua latina per se illustrata and Fabulae Faciles (Latin), Athenaze and Logos (Greek), and the Graded Reader of Biblical Hebrew and Biblical Hebrew Easy Stories (Hebrew).
  • The adapted texts are 'Genesis: 1' (A1 target, early-OCS reconstruction by Tomáš Spevák) and 'Saint George and the Dragon' (B1 target, later normalised OCS from Rystenko's monograph).
  • mT5 could not perform CEFR-guided adaptation, even when prompted in a higher-resourced language. GPT-5 in a one-shot setting yielded short sentences, simple grammar and dialogue.
  • The B1 output contained fewer verbs than the professional benchmark, giving it a "lack of action". A few A1 verses were too difficult despite the overall simplicity.
  • Manual post-editing took about one hour per text and remains the recommended step. The model's heavy reliance on input spelling conventions makes rule-based correction impractical. The edited GPT-5 texts are shared via GitHub.

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

🎓Sorbonne UniversitéParis · France32 works · 7,270 citations🔬Sens, Texte, Informatique, HistoireParis · France4 works · 14 citations

 Fields, topics and concepts

📚 Shelves 4The 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 works

🧭 Topics 2Research subjects and languages the work is about.Semitic languages · 2,908 worksHebrew & Aramaic · 496 works

Cite

Cite as Iglika Nikolova-Stoupak, Gaël Lejeune, Aliona Shestakova-Stukun and Eva Schaeffer-Lacroix. 2026. Automatic Generation of Graded Texts in Old Church Slavonic. Proceedings of the 2nd Workshop on Evaluating Text Difficulty in a Multilingual Context (DeTermIt! 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/3mgy6qi4bkoq

@inproceedings{nikolovastoupak2026automatic,
  title = {{Automatic Generation of Graded Texts in Old Church Slavonic}},
  author = {Iglika Nikolova-Stoupak, Gaël Lejeune, Aliona Shestakova-Stukun and Eva Schaeffer-Lacroix},
  booktitle = {Proceedings of the 2nd Workshop on Evaluating Text Difficulty in a Multilingual Context (DeTermIt! 2026)},
  year = {2026},
  doi = {10.63317/3mgy6qi4bkoq},
  url = {https://doi.org/10.63317/3mgy6qi4bkoq}
}

References

8 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Mikhail Arkhipov and Maria Trofimova. 2021. Bertislav: Pre-trained language model for slavic languages. In Proceedings of the International Conference on Analysis of Images, Social Networks and Texts (AIST), Cham. Springer. find
  2. Maurice Balme and Gilbert Lawall. 2003. Athenaze: An Introduction to Ancient Greek. Book I, second edition. Oxford University Press, Oxford; New York. find
  3. Christopher Blackwell and Neel Smith. 2009. Alpheios: Open source tools for reading ancient languages. In Proceedings of the ACM Symposium on Document Engineering. ACM. find
  4. Srinivasa Rao Bogireddy and Nagaraju Dasari. 2024. Comparative analysis of chatgpt-4 and llama: Performance evaluation on text summarization, data analysis, and question answering. In 2024 15th International Conference on Computing Communication and Networking Technologies (ICCCNT), pages 1–7. find
  5. Gregory Crane. 1996. The perseus digital library. In Proceedings of the ACM/IEEE Joint Conference on Digital Libraries, New York, USA. ACM. find
  6. Karl C. Diller and Thomas M. Walsh. 1978. “living” and “dead” languages: A neurolinguistic distinction. In Actes du 5e Congrès de l’Association Internationale de Linguistique Appliquée, Montréal. Les Presses de l’Université Laval. find
  7. William H. DuBay. 2007. The Classic Readability Studies. ERIC Clearinghouse. find
  8. Björn Engelmann, Fabian Haak, Christin Katharina Kreutz, Narjes Nikzad Khasmakhi, and Philipp Schaer. 2023. Text simplification of scientific texts for non-expert readers. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/stoupak-2026-automatic-generation-graded-texts-old/