Elamur Research⌕
☰
home / papers / 2026 / Challenges in Machine Translation of Interactive Multimodal Exercises

Challenges in Machine Translation of Interactive Multimodal Exercises

Paper · 2026 · Lucie Polakova, Miroslav Hrabal, Věra Kloudová et al.

Summary

The authors report on an applied project that localizes the Czech e-learning portal Škola s nadhledem into Ukrainian, English and German. The portal is used by many Ukrainian pupils in Czech schools. Its content is difficult for MT: exercises are short, heavy in terminology, interactive, multimodal, and wrapped in complex XML/HTML. The team translates plain text first and then projects the markup back, rather than translating text and tags together. They fine-tuned a locally hosted EuroLLM-9B-Instruct with LoRA, using in-domain parallel and back-translated data, and compared SFT, CPO and SimPO variants. Automatic metrics, two rounds of human evaluation and in-portal testing identified SFT+CPO as the best system. For Czech-Ukrainian it outperformed the earlier Transformer-based system. The paper also gives a linguistic classification of errors. It separates classic MT mistakes from newer LLM-specific ones, such as "fixing" deliberately false statements or completing unfinished sentences. These findings are useful for EdTech teams weighing open-source LLMs for confidential, domain-specific translation.

Key points

  • Data: The portal has 9,000+ exercises, of which about 1,000 suitable ones were preselected. Professionals translated 396 biology, chemistry and geography exercises into all three languages, split into 190 dev and 206 test exercises.
  • Adaptation data:
    • For Czech-English: CzEng 2.0 and back-translated Wikipedia.
    • For Czech-German: 57,000 sentence pairs crawled with Bitextor and 1.2M monolingual sentences from oebv.at.
    • Preference pairs were chosen using MetricX and Gemma-3-27B-it.
  • Main result: Compared with the pre-LLM Czech-Ukrainian system, SFT+CPO gained +1.5 chrF2, −0.4 MetricX24, and +0.61 in human scores on a 0–10 scale.
  • In-image text detection: In a preliminary test, a small VLM detected images needing translation with 84% recall (F1 = 0.64) on 1,000 images. Under 7% of these images actually required translation.
  • Releases:
    • The models (Hrabal et al., 2025) and the translation software (Popel et al., 2025) are available in LINDAT/CLARIAH. The software includes a web frontend, an API server and a tagged-document translator.
    • The translated exercises are due to become public by the end of 2026.

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

🎓Charles UniversityPrague · Czechia133 works · 20,686 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.Machine translation · 1,087 worksVision & language · 657 works

Cite

Cite as Lucie Polakova, Miroslav Hrabal, Věra Kloudová, Michal Novák, Mariia Anisimova and Martin Popel. 2026. Challenges in Machine Translation of Interactive Multimodal Exercises. Proceedings of the 21st Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.bea-1.11

@inproceedings{polakova2026challenges,
  title = {{Challenges in Machine Translation of Interactive Multimodal Exercises}},
  author = {Lucie Polakova, Miroslav Hrabal, Věra Kloudová, Michal Novák, Mariia Anisimova and Martin Popel},
  booktitle = {Proceedings of the 21st Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.bea-1.11},
  url = {https://doi.org/10.18653/v1/2026.bea-1.11}
}

References

22 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Nuaman Chen, Li Zhou, and Zilong Wang. 2023. Frustratingly easy label projection for cross-lingual transfer. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 4500–4515, Singapore. Association for Computational Linguistics. find
  2. Miquel Esplà-Gomis. 2009. Bitextor: a free/open-source software to harvest translation memories from multilingual websites. In Beyond Translation Memories: New Tools for Translators Workshop, Ottawa, Canada. find
  3. Miroslav Hrabal, Martin Popel, Lucie Poláková, Michal Novák, Věra Kloudová, and Mariia Anisimova. 2025. EdUKate translation models 2025. find
  4. LINDAT/CLARIAH-CZ digital library at the Institute of Formal and Applied Linguistics (ÚFAL), Faculty of Mathematics and Physics, Charles University. find
  5. Juraj Juraska, Daniel Deutsch, Mara Finkelstein, and Markus Freitag. 2024. MetricX-24: The Google Submission to the WMT 2024 Metrics Shared Task. In Proceedings of the Workshop on Machine Translation (WMT 2024).
  6. Anna-Lena Kempe and Åke Grönlund. 2019. Collaborative digital textbooks–a comparison of five different designs shaping teaching and learning. Education and Information Technologies, 24(5):2909–2941. find
  7. Věra Kloudová, Ondřej Bojar, and Martin Popel. 2021. Detecting post-edited references and their effect on human evaluation. In Proceedings of the Workshop on Human Evaluation of NLP Systems (HumEval), pages 114–119, Online. Association for Computational Linguistics. find
  8. Tom Kocmi, Ekaterina Artemova, Eleftherios Avramidis, Rachel Bawden, Ondřej Bojar, Konstantin Dranch, Anton Dvorkovich, Sergey Dukanov, Mark Fishel, Markus Freitag, Thamme Gowda, Roman Grundkiewicz, Barry Haddow, Marzena Karpinska, Philipp Koehn, Howard Lakougna, Jessica Lundin, Christof Monz, Kent… — Findings of the WMT25 General Machine Translation Shared Task: Time to Stop Evaluating on Easy Test Sets
  9. Tom Kocmi, Martin Popel, and Ondřej Bojar. 2020. Announcing CzEng 2.0 Parallel Corpus with over 2 Gigawords. Preprint, arXiv:2007.03006. — [arXiv:2007.03006](https://arxiv.org/abs/2007.03006) find
  10. Lieve Macken, Laura Van Brussel, and Joke Daems. 2019. NMT’s wonderland where people turn into rabbits. A study on the comprehensibility of newly invented words in NMT output. Computational Linguistics in the Netherlands Journal, 9:67–80. find
  11. Yu Meng, Mengzhou Xia, and Danqi Chen. 2024. Simpo: Simple preference optimization with a reference-free reward. In Advances in Neural Information Processing Systems (NeurIPS). find
  12. Maslahatun Nadziro, Aufia Aisa, and Rina Dian Rahmawati. 2023. The development of QR code-based Arabic textbooks for non-arabic education students. Scaffolding: Jurnal Pendidikan Islam dan Multikul-turalisme, 5(2):960–981. find
  13. NLLB Team, Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, and 1 others. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672. find
  14. Tanmay Parekh, I-Hung Hsu, Kuan-Hao Huang, Kai-Wei Chang, and Nanyun Peng. 2024. Contextual label projection for cross-lingual structured prediction. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Vol… find
  15. Lucie Poláková, Martin Popel, Věra Kloudová, Michal Novák, Mariia Anisimova, and Jiří Balhar. 2025. Mitigating language barriers in education: Developing multilingual digital learning materials with machine translation. In EDULEARN25 Proceedings, pages 8754–8760, Valencia, Spain. IATED. find
  16. Martin Popel, Michal Novák, Jiří Balhar, Ondřej Košarko, Jiří Mayer, Lucie Poláková, Věra Kloudová, and Mariia Anisimova. 2025. EdUKate translation software 2. LINDAT/CLARIAH-CZ digital library at the Institute of Formal and Applied Linguistics (ÚFAL), Faculty of Mathematics and Physics, Charles… find
  17. Martin Popel, Lucie Poláková, Michal Novák, Jindřich Helcl, Jindřich Libovický, Pavel Straňák, Tomáš Krabač, Jaroslava Hlaváčová, Mariia Anisimova, and Tereza Chlaňová. 2024. Charles Translator: A Machine Translation System between Ukrainian and Czech. In Proceedings of the 2024 Joint Internat… find
  18. Maja Popović. 2015. chrF: character n-gram F-score for automatic MT evaluation. In Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 392–395, Lisbon, Portugal. Association for Computational Linguistics.
  19. D. K. Thennal, Chris Biemann, and Hans Ole Hatzel. 2026. Just Use XML: Revisiting Joint Translation and Label Projection. Preprint, arXiv:2603.12021. find
  20. 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.
  21. Haoran Xu, Amr Sharaf, Yunmo Chen, Weiting Tan, Lingfeng Shen, Benjamin Van Durme, Kenton Murray, and Young Jin Kim. 2024. Contrastive preference optimization: Pushing the boundaries of LLM-based machine translation. In Proceedings of the 41st International Conference on Machine Learning, Vienna, Au… find
  22. Thomas Zenkel, Joern Wuebker, and John DeNero. 2021. Automatic bilingual markup transfer. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3524–3533, Punta Cana, Dominican Republic. Association for Computational Linguistics. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/polakova-2026-challenges-machine-translation-interactive-multimodal/