Bengali-English Neural Machine Translation Using Deep Learning Techniques

Abstract

Bengali is one of the most widely spoken languages and one of the hardest to translate due to its extensive vocabulary. Earlier, it was fairly difficult to translate from Bengali to English. Using neural machine translation (NMT), it is now possible to translate from Bengali to English quite flawlessly. In order to carry out the task of neural machine translation, four different Seq2Seq models - Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU), Bidirectional LSTM (BiLSTM), and Bidirectional GRU (BiGRU)—are studied in this work. We combined four distinct datasets and used the resultant dataset in association with the four models. Our study shows that BiLSTM is the most effective model for the Bengali to English NMT task. Here, the resemblance between the generated output and the real one is assessed using two frequently used performance metrics termed BLEU and ROUGE. We have achieved scores of 47.4, 35.8, 32.0 and 22.8 for BLEU-1, 2, 3 and 4 respectively on BiLSTM. Last but not least, our outcomes are among the finest of other studies performed earlier.