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

Valerio Lorini, Paula Vlaic, Ulascan Akbulut, Daniele Marcoaldi


Abstract
EU legislation is equally authentic and legally binding in all 24 official languages, rendering high-quality translation a legal obligation rather than a mere choice. Therefore, high-quality language technology supporting translation processes in all EU languages are essential for language professionals at the European Parliament (EP). This paper investigates whether domain-specific fine-tuning of an open-weight Large Language Model (LLM) yields consistently larger quality gains on legislative text compared to generic text, in all 23 EU target languages from English. We evaluate ten experimental conditions: base model, in-domain and cross-domain fine-tuning, sequential generic-then-legislative fine-tuning, and zero-shot Claude Sonnet 4.6 as a proprietary reference. We analyse BLEU, chrF, TER, and COMET metrics on nearly 700,000 segments. Results confirm the hypothesis for all 23 languages: legislative fine-tuning enhances BLEU by +12.30 compared to +7.10 for generic fine-tuning, demonstrating a consistent advantage of +5.20 BLEU in all the metrics. The fine-tuned EuroLLM-22B decisively outperforms Claude Sonnet 4.6, Anthropic’s latest frontier model, on both domains, highlighting that targeted adaptation of a smaller open-weight model can surpass a state-of-the-art proprietary system. Cross-domain transfer within institutional domain is positive for all languages, with no catastrophic forgetting. Low-resource languages such as Irish and Maltese benefit the most from fine-tuning, while a divergence between BLEU and COMET rankings for some languages underlines the need of evaluation metrics alongside traditional measures.
Anthology ID:
2026.eamt-1.20
Volume:
Proceedings of the 26th Annual Conference of the European Association for Machine Translation (Volume 1)
Month:
June
Year:
2026
Address:
Tilburg, The Netherlands
Editors:
Dimitar Shterionov, Eva Vanmassenhove, Mirella De Sisto, Fred Blain, Javad Pourmostafa Roshan Sharami, Lisa Lepp, Chiara Manna, Argentina Anna Rescigno, Alina Karakanta, Ayla Rigouts Terryn, Manuel Lardelli, Natalia Resende, Elena Murgolo, Janiça Hackenbuchner, Anna Zaretskaya, Miquel Esplà-Gomis, Thierry Etchegoyhen, Dagmar Gromann, Rachel Bawden, Barry Haddow, Sara Szoc, Mikel Forcada, Helena Moniz
Venue:
EAMT
SIG:
Publisher:
European Association for Machine Translation
Note:
Pages:
304–320
Language:
URL:
https://aclanthology.org/2026.eamt-1.20/
DOI:
Bibkey:
Cite (ACL):
Valerio Lorini, Paula Vlaic, Ulascan Akbulut, and Daniele Marcoaldi. 2026. One Size Does Not Fit All: Why EU Legislative Translation Demands Domain-Specific Fine-Tuning of LLMs. In Proceedings of the 26th Annual Conference of the European Association for Machine Translation (Volume 1), pages 304–320, Tilburg, The Netherlands. European Association for Machine Translation.
Cite (Informal):
One Size Does Not Fit All: Why EU Legislative Translation Demands Domain-Specific Fine-Tuning of LLMs (Lorini et al., EAMT 2026)
Copy Citation:
PDF:
https://aclanthology.org/2026.eamt-1.20.pdf