Elamur Research⌕
☰
home / papers / 2026 / LoRA Fine-Tuning of English–Norwegian NMT for the Oil & Gas Industry

LoRA Fine-Tuning of English–Norwegian NMT for the Oil & Gas Industry

Summary

Translating petroleum-industry texts is hard: terminology and regulatory phrasing are specialised, in-domain parallel data is scarce, and full fine-tuning is expensive. The authors propose a three-stage recipe for adapting facebook/nllb-200-distilled-600M to English→Norwegian Bokmål with LoRA. They clean the Norwegian Petroleum Directorate (NPD) translation memory (ELRC 559) to 17,414 sentence pairs. They then (1) measure how quality grows with training-set size, (2) tune LoRA settings with both a grid search and Optuna with ASHA pruning, and (3) compare the result against commercial engines, open-source models and full fine-tuning. The tuned adapter trains under 0.4% of the weights and reaches 61.48 BLEU, about 24.6 points above the base model. On BLEU it beats all four commercial systems tested, and its COMET (0.9298) is close to the strongest one. A manual review finds rare but serious errors in dates, directions and terms that BLEU overlooks. The study offers a cheap, reproducible template for specialised, low-resource MT.

Key points

  • Data: About 26k raw EN–NB pairs from ELRC 559 were normalised, filtered and deduplicated (30.9% removed, mostly template repetition). This left 17,414 pairs, split 13,935 / 1,737 / 1,742 for train/dev/test. The raw corpus scored 0.976 on the composite OQS 2.0 quality metric.
  • Data scaling: Training subsets ranged from 100 to 13,935 samples, each run with 3 seeds. About 8,000 samples already yield roughly 96% of the maximum BLEU.
  • Hyperparameter search: A 27-configuration grid and 50 Optuna+ASHA trials converged on r=8, α=64 and dropout=0. α was by far the most influential parameter (importance 0.97).
  • Benchmarks: The tuned LoRA model scored 61.48 BLEU and 0.9298 COMET. Microsoft Translator reached 57.88 BLEU and 0.9313 COMET, and the untuned base model 36.86 BLEU. With default settings, full fine-tuning beat LoRA by at most 0.85 BLEU, and LoRA varied less across seeds.
  • Human evaluation: On 50 test sentences (Cohen's κ 0.71 for error type, 0.61 for severity), annotators found 62 errors.
    • 83% were minor and 12% critical.
    • The most common categories were Word Choice, accounting for 40%, and, at 19%, Variant Mixing, which is attributed to Nynorsk forms present in the data.

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

Topics: Machine translation

Cite

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

@inproceedings{yang2026lora,
  title = {{LoRA Fine-Tuning of English–Norwegian NMT for the Oil \& Gas Industry}},
  author = {Yang, Xiaojing and Li, Zhihan and Sun, Gege and Li, Mengyue and Beloucif, Meriem},
  booktitle = {Proceedings of the 26th Annual Conference of the European Association for Machine Translation (Volume 1)},
  year = {2026},
  url = {https://aclanthology.org/2026.eamt-1.25/}
}

References

5 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 optimization framework. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 2623–2631. find our card
  2. Li, Xiang Lisa and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582… find
  3. Artetxe, Mikel and Holger [Li et al.2018] Li, Lisha, Kevin Jamieson, Giulia De-Salvo, Afshin Rostamizadeh, and Ameet Talwalkar. Schwenk. 2019. Massively multilingual sentence 2018. Hyperband: A novel bandit-based approach embeddings for zero-shot cross-lingual transfer and to hyperparameter optimiza… — [arXiv:2512.09910](https://arxiv.org/abs/2512.09910) find
  4. Wolf, Thomas, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lh… find our card
  5. Zheng, Jiawei, Hanghai Hong, Feiyan Liu, Xiaoli Wang, Jingsong Su, Yonggui Liang, and Shikai Wu. 2024. Fine-tuning large language models for domain-specific machine translation. https://arxiv.org/abs/2402.15061. arXiv preprint arXiv:2402.15061. — [arXiv:2402.15061](https://arxiv.org/abs/2402.15061) ↗

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yang-2026-lora-fine-tuning-english-norwegian/