Elamur Research⌕
☰
home / papers / 2026 / Script Correction and Synthetic Pivoting: Adapting Tencent HY-MT for Low-Resource Turkic Translation

Script Correction and Synthetic Pivoting: Adapting Tencent HY-MT for Low-Resource Turkic Translation

Summary

This LoResMT 2026 shared-task submission covers three Turkic tracks: Russian-Kazakh, Russian-Bashkir and English-Chuvash. The backbone, Tencent HY-MT1.5-7B, is a strong multilingual translator, but it writes Kazakh and Bashkir in Arabic script rather than the Cyrillic used today. The author applied LoRA adapters to the attention layers and fine-tuned on Cyrillic parallel corpora. This both adapted the model to the new language pairs and removed the script problem, with no rule-based converter needed. Chuvash has very little English-paired data, so two options were compared. The first mixes the small book-derived En-Cv set with a large Ru-Cv corpus. The second builds a fully synthetic En-Cv corpus by machine-translating the Russian side of Ru-Cv into English and trains only on that. The adapted models clearly beat the NLLB and Gemma baselines on chrF++, and synthetic pivoting edged out mixed training. For Ukrainian NLP work, the paper shows a cheap way to repair script errors and to bootstrap rare language pairs through a better-resourced pivot language.

Key points

  • Setup: HY-MT1.5-7B with LoRA (rank 16 on q/k/v/o, about 13.6M trainable parameters) was trained for 2 epochs on 4 A100 GPUs. After adaptation, manual checks found no Arabic characters in the output.
  • Data: Kazakh used ISSAI KazParc plus extra En-Kk data. Bashkir used the AigizK Bashkir-Russian corpus. Chuvash used alexantonov's chuvash_english_parallel (200k pairs) and chuvash_russian_parallel (1.4M pairs).
  • Pivoting: The synthetic English side was produced with facebook/wmt19-ru-en using beam size 3. Final test outputs used beam 5.
  • Results (chrF++):
    • Ru-Kk rose from 16.0 to 48.0 against NLLB 1.3B distilled.
    • Ru-Ba rose from 28.0 to 52.0 against the same baseline.
    • En-Cv reached 35.8 with synthetic data and 34.4 with mixed data, against 23.0 for Gemma 2 3B.
  • Error analysis: A review of about 50 samples showed the mixed model absorbing the literary register of the book data (for example, rendering ginger as "red plant"). The pivot model handled everyday and technical terms better. Listed limitations are dependence on pivot quality and not combining synthetic data with the Ru-Cv corpus.

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

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.loresmt-1.20

@inproceedings{maxim2026script,
  title = {{Script Correction and Synthetic Pivoting: Adapting Tencent HY-MT for Low-Resource Turkic Translation}},
  author = {Maxim, Bolgov},
  booktitle = {Proceedings for the Ninth Workshop on Technologies for Machine Translation of Low Resource Languages (LoResMT 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.loresmt-1.20},
  url = {https://doi.org/10.18653/v1/2026.loresmt-1.20}
}

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. Isaac Caswell and Ankur Bapna. 2022. Unlocking zero-resource machine translation to support new languages in google translate. Google Research Blog. find
  2. Khalid N. Elmadani and Jan Buys. 2024. Neural machine translation between low-resource languages with synthetic pivoting. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 12144–12158, Torino, Italia. E… find
  3. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. Preprint, arXiv:2106.09685. find
  4. Jamshidbek Mirzakhalov, Anoop Babu, Duygu Ataman, Sherzod Kariev, Francis Tyers, Otabek Abduraufov, Mammad Hajili, Sardana Ivanova, Abror Khaytbaev, Antonio Laverghetta Jr., Bekhzodbek Moydinboyev, Esra Onal, Shaxnoza Pulatova, Ahsan Wahab, Orhan Firat, and Sriram Chellappan. 2021. A large-scale stu… find
  5. Nathan Ng, Kyra Yee, Alexei Baevski, Myle Ott, Michael Auli, and Sergey Edunov. 2019. Facebook FAIR’s WMT19 news translation task submission. In Proceedings of the Fourth Conference on Machine Translation (Volume 2: Shared Task Papers, Day 1), pages 314–319, Florence, Italy. Association for Computat… find our card
  6. Toan Q. Nguyen and David Chiang. 2017. Transfer learning across low-resource, related languages for neural machine translation. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 296–301, Taipei, Taiwan. Asian Federation of Natu… find our card
  7. Nikolay Plotnikov and Alexander Antonov. 2024. Open the data! chuvash datasets. Preprint, arXiv:2407.11982. — [arXiv:2407.11982](https://arxiv.org/abs/2407.11982) find
  8. Iskander Shakirov and Aigiz Kunafin. 2023. Bashkir-russian parallel corpora. Hugging Face dataset. find
  9. Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, Johan Ferret, Peter Liu, Pouya Tafti, Abe Friesen, Michelle Casbon, Sabela Ramos, Ravin Kumar, Charline Le Lan, Sammy Jerome, and 179… find
  10. NLLB Team, Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, and 20 o… find
  11. Rustem Yeshpanov, Alina Polonskaya, and Huseyin Atakan Varol. 2024. KazParC: Kazakh parallel corpus for machine translation. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 9633–9644, Torino, Italia.… find
  12. Mao Zheng, Zheng Li, Tao Chen, Mingyang Song, and Di Wang. 2025. Hy-mt1.5 technical report. Preprint, arXiv:2512.24092. — [arXiv:2512.24092](https://arxiv.org/abs/2512.24092) find
  13. Barret Zoph, Deniz Yuret, Jonathan May, and Kevin Knight. 2016. Transfer learning for low-resource neural machine translation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1568–1575, Austin, Texas. Association for Computational Linguistics. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/maxim-2026-script-correction-synthetic-pivoting-adapting/