Elamur Research⌕
☰
home / papers / 2026 / One Size Does Not Fit All: Why EU Legislative Translation Demands Domain-Specific Fine-Tuning of LLMs

One Size Does Not Fit All: Why EU Legislative Translation Demands Domain-Specific Fine-Tuning of LLMs

Summary

The authors ask whether adapting an open-weight LLM to EU legislation helps more than adapting it to ordinary institutional text. This matters because every language version of EU law is legally authentic, so translation errors have legal consequences. Using translation memories from the European Parliament's DG TRAD (extracted from EUramis), they built English-source parallel corpora in all 24 official languages: one of consolidated legislative documents and one of generic parliamentary texts. They fine-tuned EuroLLM-22B with LoRA in single-domain, cross-domain and sequential (generic, then legislative) settings, and compared it with zero-shot Claude Sonnet 4.6. Ten conditions were scored with BLEU, chrF, TER and COMET on roughly 700,000 segments. Across all 23 target languages, legislative tuning produced clearly larger gains than generic tuning. Cross-domain transfer was positive, and the adapted model beat the proprietary system. The practical conclusion is that institutions can get top quality from in-house data without sending sensitive drafts to external providers.

Key points

  • Legislative fine-tuning raised BLEU by +12.30, compared with +7.10 for generic fine-tuning on generic text. That gap of +5.20 BLEU is matched by +3.23 chrF, −4.45 TER and +0.49 COMET.
  • Sequential tuning (C8) achieved the best legislative result at 63.17 BLEU (+13.51 over base) and kept 79% of the generic-domain gain.
  • Fine-tuned EuroLLM-22B scored 8.06 BLEU higher than Claude Sonnet 4.6 for legislative content and 5.30 higher for generic content.
  • Each domain-specific model improved the other domain by about 3.5 BLEU, with no catastrophic forgetting. Irish and Maltese gained the most.
  • The data come from 112 legislative and 3,113 generic EP documents. The evaluation datasets are released on Zenodo (doi:10.5281/zenodo.20072253) for the AI4TRAD project.

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 · Machine translation · Law

Topics: Machine translation

Cite

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{lorini2026one,
  title = {{One Size Does Not Fit All: Why EU Legislative Translation Demands Domain-Specific Fine-Tuning of LLMs}},
  author = {Lorini, Valerio and Vlaic, Paula and Akbulut, Ulascan and Marcoaldi, Daniele},
  booktitle = {Proceedings of the 26th Annual Conference of the European Association for Machine Translation (Volume 1)},
  year = {2026},
  doi = {10.5281/zenodo.20072252},
  url = {https://doi.org/10.5281/zenodo.20072252}
}

References

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

  1. Akiba, Takuya, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. 2019. Optuna: A next-generation hyperparameter [Hu et al.2022] Hu, Edward J., Yelong Shen, Phillip optimization framework. In Proceedings of the 25th Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, ACM SIGKDD international conference on knowl-Lu Wang, and Weizhu Chen. 2022. LoRA: Low-edge discovery & data mining, pages 2623–2631. rank adaptation of large language models. arXiv preprint arXiv:2106.09685. ↗
  2. Alves, Duarte M, José Pombal, Nuno M Guerreiro, Pedro H Martins, João [Huang and others2024] Huang, Jie et al. 2024. Miti-Alves, Amin Farajian, Ben Peters, Ricardo Rei, gating catastrophic forgetting in large language mod-Patrick Fernandes, Sweta Agrawal, et al. 2024. els with self-synthesized rehearsal. In Proceedings Tower: An open multilingual large language of the 62nd Annual Meeting of the Association for model for translation-related tasks. arXiv preprint Computational Linguistics (ACL 2024), volume 1, arXiv:2402.17733. pages 1416–1428, Bangkok, Thailand. ↗ our card
  3. Barman, Amit, Atanu Mandal, and [Jaccard1912] Jaccard, Paul. 1912. The distribution Sudip Kumar Naskar. 2025. From scratch to fine-of the flora in the alpine zone. 1. New phytologist, tuned: A comparative study of transformer training 11(2):37–50. strategies for legal machine translation. In Proceedings of the 1st Workshop on NLP for Empowering [Levenshtein and others1966] Levenshtein, Vladimir I Justice (JUST-NLP 2025). Association for Compu-et al. 1966. Binary codes capable of correcting tational Linguistics. deletions, insertions, and reversals. In Soviet physics doklady, volume 10, pages 707–710. Soviet Union. find
  4. Bié, Laurent, Aleix Cerdà-i Cucó, Hans Degroote, Amando Estela, Mercedes Garcı́a- [Lewis et al.2020] Lewis, Patrick, Ethan Perez, Alek-Martı́nez, Manuel Herranz, Alejandro Kohan, Maite sandra Piktus, Fabio Petroni, Vladimir Karpukhin, Melero, Tony O’Dowd, Sinéad O’Gorman, Mārcis Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-Pinnis, Roberts Rozis, Riccardo Superbo, and Artūrs tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-Vasiļevskis. 2020. Neural translation for the Euro-augmented generation for knowledge-intensive nlp pean Union (NTEU) project. In Martins, André, He-tasks. Advances in neural information processing lena Moniz, Sara Fumega, Bruno Martins, Fernando systems, 33:9459–9474. Batista, Luisa Coheur, Carla Parra, Isabel Trancoso, Marco Turchi, Arianna Bisazza, Joss Moorkens, [Luo et al.2023] Luo, Yun, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. 2023. An empiri-Ana Guerberof, Mary Nurminen, Lena Marg, and cal study of catastrophic forgetting in large language Mikel L. Forcada, editors, Proceedings of the 22nd models during continual fine-tuning. In IEEE Trans-Annual Conference of the European Association for actions on Audio, Speech and Language Processing. Machine Translation, pages 477–478, Lisboa, Portugal, November. European Association for Machine find
  5. Broder, Andrei Z. 1997. On the resem-Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, blance and containment of documents. In Proceed-Skyler Wang, Guillaume Wenzek, Al Youngblood, ings. Compression and Complexity of SEQUENCES Bapi Akula, Loic Barrault, Gabriel Mejia Gonza-1997 (Cat. No. 97TB100171), pages 21–29. IEEE. lez, Prangthip Hansanti, John Hoffman, Semarley find
  6. Chaplot, Devendra Singh. 2023. Al-Jarrett, Kaushik Ram Sadagopan, Dirk Rowe, Shan-bert q. jiang, alexandre sablayrolles, arthur mensch, non Spruit, Chau Tran, Pierre Andrews, Necip Fazil chris bamford, devendra singh chaplot, diego de las Ayan, Shruti Bhosale, Sergey Edunov, Angela Fan, casas, florian bressand, gianna lengyel, guillaume Cynthia Gao, Vedanuj Goswami, Francisco Guzmán, lample, lucile saulnier, lélio renard lavaud, marie-Philipp Koehn, Alexandre Mourachko, Christophe anne lachaux, pierre stock, teven le scao, thibaut Ropers, Safiyyah Saleem, Holger Schwenk, and Jeff lavril, thomas wang, timothée lacroix, william el Wang. 2024. Scaling neural machine translation to sayed. arXiv preprint arXiv:2310.06825, 3. 200 languages. Nature, 630(8018):841–846. ↗ our card
  7. Dettmers, Tim, Artidoro Pagnoni, [Ramos et al.2026] Ramos, Miguel Moura, Duarte M. Aman Srivastava, and Ari Holtzman. 2023. Alves, et al. 2026. EuroLLM-22B: Technical report. QLoRA: Efficient finetuning of quantized language arXiv preprint arXiv:2602.05879. models. In Advances in Neural Information Process- ↗
  8. Rei, Ricardo, Nuno M Guerreiro, José ing Systems (NeurIPS 2023). Pombal, João Alves, Pedro Teixeirinha, Amin Farajian, and André FT Martins. 2025. Tower+: Bridg- find
  9. Grattafiori, Aaron, Abhimanyu ing generality and translation specialization in mul-Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek tilingual llms. arXiv preprint arXiv:2506.17080. Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783. ↗
  10. Saunders, Danielle. 2023. Domain adaptation and multi-domain adaptation for neural machine translation: A survey. Journal of Artificial Intelligence Research, 75:1–67. ↗
  11. Seracini, Francesca L. 2021. Phraseology in multilingual EU legislation: a corpus-based study of translated multi-word terms. Perspectives, 29(2). ↗
  12. Shahnazaryan, Lia et al. 2024. Defining boundaries: The impact of domain specification on cross-language and cross-domain transfer in machine translation. arXiv preprint arXiv:2408.11926. ↗
  13. Touvron, Hugo, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971. ↗

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/papers/lorini-2026-one-size-does-not-fit/