NORMALIZING KHOREZM DIALECT SPEECH TRANSCRIPTS INTO STANDARD UZBEK: A FINE-TUNED NLLB200 APPROACH

The Khorezm dialect differs significantly from the Uzbek literary language, making the direct application of standard linguistic rules insufficient for its accurate interpretation and analysis. This study addresses the task of translating Khorezm dialect transcripts into standard Uzbek by adapting the NLLB-200 multilingual translation model, which already includes Uzbek among its supported languages. A manually aligned corpus consisting of 2,932 dialect–standard Uzbek sentence pairs collected from the Khorezm region was developed and used for model refinement. The baseline zero-shot model achieved 16.6 BLEU and 51.7 ChrF scores on the test set. After fine-tuning on the prepared corpus, the model performance improved substantially, reaching 50.3 BLEU and 76.8 ChrF. The improvement of nearly 34 BLEU points demonstrates that a relatively small parallel corpus, combined with a pretrained model containing knowledge of the target language, is sufficient to develop an effective dialect-to-literary language translation system. The resulting corpus and optimized model provide a foundation for further research on automatic processing and translation of other Uzbek regional dialects.

Hali tarjima qilinmagan