Cross-Lingual Transfer for Machine Translation in Turkic Languages

2026-07-31

AI Summary

This study addresses the lack of systematic analysis on cross-lingual transfer in machine translation among low-resource Turkic languages. The authors construct a pairwise transfer matrix for Turkish, Azerbaijani, Uzbek, Kazakh, and Kyrgyz, fine-tuning the mT5 model and evaluating performance across source–target language pairs using BLEU and chrF metrics. Experiments incorporate Latin-script preprocessing and multi-dataset stability analysis. Results indicate that linguistic relatedness strongly influences transfer effectiveness, with Turkish↔Azerbaijani and Kazakh↔Kyrgyz achieving the highest scores. Translation direction and the choice of target language significantly modulate performance. While Latinization improves output quality in certain script-mismatch scenarios, its benefits are not consistent across all settings. Finally, the source language demonstrates robustness across varying experimental configurations.

Abstract

Cross-lingual transfer is central to low-resource machine translation, but its behavior within closely related language families remains insufficiently characterized. We study transfer among five Turkic languages; Turkish, Azerbaijani, Uzbek, Kazakh, and Kyrgyz; using pairwise transfer matrices. In this setting, each model is fine-tuned with one transfer source and evaluated on a different transfer target while the translation target remains the same. Across mT5 experiments, we find that transfer is strongest between closely related Turkic pairs, especially Turkish-Azerbaijani and Kazakh-Kyrgyz. We also show that transfer direction matters, and that the same transfer source-transfer target pair can behave differently when the translation target changes. Latinization improves BLEU and chrF in several script-mismatched settings, but its effect is not uniform across metrics. Additional analyses show that transfer sources are mostly stable across different datasets and model settings.

Problem

• cross-lingual transfer
• low-resource machine translation
• Turkic languages
• language relatedness
• script mismatch

Innovation

• cross-lingual transfer
• Turkic languages
• pairwise transfer matrices
• script normalization
• low-resource machine translation