A Study of Word Embedding Models for Machine Translation of North Eastern Languages

Neural Machine Translation (NMT) has experienced significant growth in recent years and is now a well-established field. Despite being the most popular machine translation solution, NMT’s performance on low-resource language pairs is inferior to that of high-resource language pairs because of the lack of large parallel corpora for these languages.This paper explores how pre-trained word embedding models, specifically FastText and Word2Vec, can enhance the translation performance of NMT (neural machine translation) systems for low resource English-Assamese, English-Manipuri, and English-Nepali language pairs. We have compared the performance of NMT systems trained with the use of these pre-trained word embedding models and evaluated the results on test sets of English-Assamese, English-Manipuri, and English-Nepali translations. The resukts indicate that the use of FastText significantly improves the translation performance of NMT systems for both language pairs, with an average increase in BLEU score of 5.6, 7.3, and 4.4 points for English-Assamese, English-Manipuri, and English-Nepali, respectively. We have also discussed the potential challenges and limitations of using pre-trained word embedding models for NMT in low-resource languages and suggested directions for future work in this area.