Cross-Lingual Transfer in Turkic Machine Translation
A recent study published on arXiv (2607.29355) examines cross-lingual transfer in machine translation involving five Turkic languages: Turkish, Azerbaijani, Uzbek, Kazakh, and Kyrgyz. Utilizing pairwise transfer matrices, each model is fine-tuned on one source language and assessed on a different target, while keeping the translation target consistent. Experiments conducted with mT5 models indicate that transfer efficacy is highest between closely related languages, notably Turkish-Azerbaijani and Kazakh-Kyrgyz. The research also highlights the significance of transfer direction, as the same source-target pair may yield varied results with different translation targets. Latinization enhances BLEU and chrF scores in several script-mismatched contexts, though its impact varies across metrics. Further analysis reveals that transfer sources remain largely stable, aiding the understanding of cross-lingual transfer in low-resource machine translation.
Key facts
- Study examines cross-lingual transfer among five Turkic languages: Turkish, Azerbaijani, Uzbek, Kazakh, and Kyrgyz.
- Uses pairwise transfer matrices, fine-tuning on one transfer source and evaluating on a different transfer target.
- mT5 experiments show strongest transfer between Turkish-Azerbaijani and Kazakh-Kyrgyz pairs.
- Transfer direction matters; same source-target pair can behave differently when translation target changes.
- Latinization improves BLEU and chrF in script-mismatched settings, but not uniformly across metrics.
- Transfer sources are mostly stable in additional analyses.
- Research is from arXiv paper 2607.29355.
- Study addresses low-resource machine translation within closely related language families.
Entities
Institutions
- arXiv