Bridging Dialects: Enhancing Neural Machine Dialect Translation for Gujarati Language Through Model Adaptation

Over the time, Natural Language Processing (NLP) has gained much prominence and also it has evolved in the way different approaches are used to design and implement the algorithms. One of the important tasks that can be performed with the help of NLP is of Machine Translation. While the Machine Translation systems are very much accurate in current scenario, there is still room for improvement in case of those languages which are low in resource. To solve the issue, this study is focused on development and assessment of a novel Neural Machine Translation (NMT) model which is specifically designed for the task of translating from Saurashtra Gujarati dialect to the Standard Gujarati dialect. While Gujarati is a low resource language, its dialectical variation of Saurashtra has even less available resources. So, to tackle this issue we undertake the creation of a self-generated supervised dataset with sentence pair of Saurashtra Gujarati sentence and standard Gujarati sentence, that helps in capturing the unique vocabulary and linguistic nuances of the Saurashtra Gujarati. For the NMT system, we are utilizing one of the recent developments by Meta which is NLLB (No Language Left Behind) model. It is used as the base architecture and then it is fine-tuned on the self-generated dataset that would help in capturing the complexities of Saurashtra Gujarati and enhancing the translation capabilities of the model. The results show that the NLLB model outperforms the MBART, GPT-4, and Gemini Advanced on evaluation metrics METEOR Score, BLEU Score, and ChrF Score by minimum 2% and maximum 14%. The results show that the NLLB is highly adaptive to low-resource language for dialect translation.