Translation for Hindi-English: A Comparative Performance Analysis of Multilingual and Code-Mixed Text

The importance of code-mixed text is significant in light of India's rich diverse languages and cultures. Its significance is particularly notable in this modern era, as people are spending more and more time on social media platforms. Here, people use code-mixed text for convenience in conversations. Therefore, it's crucial to focus research efforts on these important topics. In this study, we aim to investigate the comparative performance of machine translation from Hindi to English and Hindi-English code-mixed text to English. We used a publicly available dataset HinGE introduced by Srivastava et al. Due to the relatively small size of our dataset, we opted for the MarianMT model, a transformer-based architecture that has been recognized for its ability to perform machine translation tasks, even when dealing with limited data. The BLEU scores of 0.91 for multilingual translation and a remarkable 0.97 for code-mixed Hindi-English translation indicate that our choice performed well. These findings highlight the model's excellent results and ability to perform in situations with limited training data.