EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Design of an integrated training data generation pipeline for machine translation for less-resourced languages

D. Lysenko, O. Kobylin, O. Putiatina, O. Lysenko

Системи обробки інформації · 30.06.2026 · (2(185)) · 181-192

DOI: 10.30748/soi.2026.185.21 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Системи обробки інформації, online erschienen am 30.06.2026. Verfasst von D. Lysenko und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

This paper describes a five-stage workflow designed for automated parallel corpora creation for direct neural machine translation systems between low-resource languages. The solution’s architecture combines multilingual sentence representation information extraction, large language model synthetic generation, and unsupervised machine translation, alongside iterative scaling via self-learning mechanisms. Quality control is maintained through neural semantic proximity assessment and hierarchical clustering tools. The proposed toolkit can generate the training datasets needed for historical text and extinct language analysis.

Zitieren

APA

Lysenko, D., Kobylin, O., Putiatina, O., & Lysenko, O. (2026). Design of an integrated training data generation pipeline for machine translation for less-resourced languages. Системи обробки інформації, 181-192. https://doi.org/10.30748/soi.2026.185.21

Vancouver

Lysenko D, Kobylin O, Putiatina O, Lysenko O. Design of an integrated training data generation pipeline for machine translation for less-resourced languages. Системи обробки інформації. 2026:181-192. doi:10.30748/soi.2026.185.21

BibTeX
@article{Lysenko2026,
  author = {Lysenko, D. and Kobylin, O. and Putiatina, O. and Lysenko, O.},
  title = {Design of an integrated training data generation pipeline for machine translation for less-resourced languages},
  journal = {Системи обробки інформації},
  year = {2026},
  number = {2(185)},
  pages = {181--192},
  doi = {10.30748/soi.2026.185.21}
}
Kostenlos starten →