Сопоставительный Анализ Дообучения Мультиязычных Нейросетевых Моделей Машинного Перевода Для Узбекско-Русской Языковой Пары

Key Points

• This research aims to evaluate the performance of various multilingual neural machine translation models for the Uzbek-Russian language pair after fine-tuning.
• Conducted a comparative analysis of six multilingual machine translation models: NLLB-200, M2M-100, mBART-50, OPUS-MT, Google Neural MT, and SeamlessM4T.
• Utilized a parallel corpus of 185,000 aligned segments for fine-tuning the models.
• Translation quality was assessed using BLEU, chrF, and TER metrics.
• NLLB-200-3.3B model showed the highest improvement in translation quality with a BLEU score of 31.6 for Uzbek to Russian.
• Identified specific error types for each model, providing insights into performance weaknesses.
• Recommendations for model architecture and fine-tuning strategies were made for low-resource Turkic languages.