Generating Synthetic Data for Mixed Russian-Kazakh Code-Switching Translation

Code-switching between Russian and Kazakh is common in Kazakhstan, yet there is almost no training data for machine translation models. Researchers from MWS AI and partner universities created a synthetic dataset by taking existing Kazakh-Russian parallel corpora (legal documents, press releases) and artificially inserting Russian words into Kazakh sentences using SimAlign, which selects replacements based on context via embeddings. They fine-tuned several open models (mBART, M2M100, NLLB-600, NLLB-3.3B) on this synthetic data. Although commercial systems scored higher on automated metrics (BLEU, ChrF++, COMET), a human evaluation using a Likert scale showed the fine-tuned NLLB-3.3B model achieved 3.09/5, beating two commercial APIs (2.80 and 3.49). The approach also improved performance when adding Twitter data translated to Kazakh. However, the synthetic data was only rated 2.62/5 for naturalness, the test set is small (618 sentences), and transferability to other language pairs remains uncertain.