Elamur Research⌕
☰
home / papers / 2026 / Improving Machine Translation of Idioms: A Spanish–Galician Parallel Dataset and Synthetic Augmentation Approach

Improving Machine Translation of Idioms: A Spanish–Galician Parallel Dataset and Synthetic Augmentation Approach

Paper · 2026 · Lúa Santamaría Montesinos, Saúl Buján, Daniel Bardanca et al.

Summary

Neural MT systems, both seq2seq and LLM-based, often translate idioms word for word and miss their figurative meaning. This holds even between close languages such as Spanish and Galician. The authors began with about 250 Spanish–Galician idiom pairs drawn from dictionaries, phraseological resources and press texts, all checked by native speakers. They narrowed this list using corpus frequency counts and linguists' judgement. For each selected pair, DeepSeek V-2 produced Galician example sentences, which were then translated into Spanish and manually verified, giving a synthetic parallel idiom corpus. Two otherwise identical transformer models were trained with OpenNMT-py on roughly 70M ES–GL sentence pairs, one with the idiom data added and one without. Adding idioms sharply improved idiom translation and slightly raised general quality. On idiom data, the small model also beat the much larger SalamandraTA-7b-instruct. The gains did not carry over to idioms absent from training. The work shows that a modest, carefully curated resource can make compact, efficient MT models more idiom-aware for a low-resource language.

Key points

  • Idiom set: The final set holds 66 bilingual idiom pairs, selected by frequency in CORPES XXI, CORGA, ESLORA, TED2020 and OpenSubtitles2016 plus expert choice. DeepSeek V-2 generated about 100 sentences per idiom. The data is released as proxectonos/corpus_paralelo_idioms on HuggingFace.
  • Models: Both are OpenNMT-py transformers with 12+12 layers, 16 heads, hidden size 512 and a 30K vocabulary, totalling about 120M parameters. They were converted to CTranslate2 (about 500 MB). Full retraining took around 40 hours on one A100. Idiom-only fine-tuning was avoided because it harmed general quality.
  • Idiom test results: BLEU rose from 70.4 to 90.2 and COMET from 0.811 to 0.912. SalamandraTA-7b-instruct scored 51.9 BLEU and 0.798 COMET.
  • Human evaluation: The augmented model's output was preferred in 90.3% of cases. The original model's output was preferred in 0.9%, and the rest were judged equivalent.
  • General benchmarks: Gains were small, and the LLM did better on Flores (28.1 BLEU, 0.859 COMET). A new evaluation resource was released: a Galician and Portuguese version of Aya Red-Teaming, published as proxectonos/aya_nos.

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, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works

🧭 Topics 2Research subjects and languages the work is about.Machine translation · 1,087 worksIdioms & multiword expressions · 217 works

Cite

Cite as Lúa Santamaría Montesinos, Saúl Buján, Daniel Bardanca and Pablo Gamallo. 2026. Improving Machine Translation of Idioms: A Spanish–Galician Parallel Dataset and Synthetic Augmentation Approach. Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1.

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

@inproceedings{montesinos2026improving,
  title = {{Improving Machine Translation of Idioms: A Spanish–Galician Parallel Dataset and Synthetic Augmentation Approach}},
  author = {Lúa Santamaría Montesinos, Saúl Buján, Daniel Bardanca and Pablo Gamallo},
  booktitle = {Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1},
  year = {2026},
  url = {https://aclanthology.org/2026.propor-1.99/}
}

References

14 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Daniel Bardanca, Pablo Gamallo, Iria de Dios-Flores, and José Ramom Pichel Campos. 2024. Exploring the effects of vocabulary size in neural machine translation: Galician as a target language. In Proceedings of the 16th International Conference on Computational Processing of Portuguese - Vol. 1, pag… find
  2. Saúl Buján, Daniel Bardanca, Pablo Gamallo, Iria de Dios-Flores, and José Ramom Pichel. 2025. Machine translation for low-resource languages: Performance trade-offs between seq2seq and generative approaches. Procesamiento del Lenguaje Natural, 75:297–315. find
  3. Isaac Caswell, Julia Kreutzer, Lisa Wang, Ahsan Wahab, et al. 2021. Quality at a glance: An audit of web-crawled multilingual datasets. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9146–9162.
  4. Iria de Dios-Flores, Silvia Paniagua Suárez, Cristina Carbajal Pérez, Daniel Bardanca Out-eiriño, Marcos Garcia, and Pablo Gamallo. 2024. CorpusNÓS: A massive Galician corpus for training large language models. In Proceedings of the 16th International Conference on Computational Processing of Po… find
  5. Ona de Gibert, Ksenia Kharitonova, Blanca Calvo Figueras, Jordi Armengol-Estapé, and Maite Melero. 2022. Quality versus quantity: Building Catalan-English MT resources. In Proceedings of the 1st Annual Meeting of the ELRA/ISCA Special Interest Group on Under-Resourced Languages, pages 59–69, Marsei… find
  6. Marzieh Fadaee, Arianna Bisazza, and Christof Monz. 2018. Examining the tip of the iceberg: A data set for idiom translation. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), pages 925– 929, Miyazaki, Japan. European Language Resources Associa… find
  7. Javier Garcia Gilabert, Xixian Liao, Severino Da Dalt, Ella Bohman, Audrey Mash, Francesca De Luca Fornaciari, Irene Baucells, Joan Llop, Miguel Claramunt, Carlos Escolano, and Maite Melero. 2025. From salamandra to salamandrata: Bsc submission for wmt25 general machine translation shared task. In P… find
  8. Naman Goyal, Cynthia Gao, Vishrav Chaudhary, Peng-Jen Chen, Guillaume Wenzek, Da Ju, Sanjana Krishnan, Marc’Aurelio Ranzato, Francisco Guzmán, and Angela Fan. 2022. The Flores-101 evaluation benchmark for low-resource and multilingual machine translation. Transactions of the Association for Computa… find
  9. Vakul Goyle, Parvathy Krishnaswamy, Kannan Girija Ravikumar, Utsa Chattopadhyay, and Kartikay Goyle. 2023. Neural machine translation for low resource languages. arXiv preprint. find
  10. Pierre Lison and Jörg Tiedemann. 2016. OpenSubtitles2016: Extracting large parallel corpora from movie and TV subtitles. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC’16), pages 923–929, Portorož, Slovenia. European Language Resources Association (…
  11. Emmy Liu, Aditi Chaudhary, and Graham Neubig. 2023. Crossing the threshold: Idiomatic machine translation through retrieval augmentation and loss weighting. In Proc. of EMNLP, pages 15095– 15111.
  12. Linfeng Liu, Saptarshi Ghosh, and Tianyu Jiang. 2025. Evaluating the impact of verbal multiword expressions on machine translation. arXiv preprint. find
  13. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational…
  14. Wenhao Zhu, Delin Pang, Liang Wang, Hangbo Chen, et al. 2023. Multilingual machine translation with large language models: Empirical results and analysis. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 14266– 14281.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/montesinos-2026-improving-machine-translation-idioms-spanish/