Comparative Study of Different Models for Language Translation

As India is country with diverse culture, there are 22 official languages and many other regional languages. Because of this, it is difficult for people to exchange the ideas and communicate with each other. In Neural Machine Translation (NMT), one language is translated to other which helps to overcome the barrier of language difference. In this paper, we have studied NMT Models to translate English sentences to Marathi. The neural machine translation models on the basis of Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU) are implemented respectively considering the differences in the structure of their networks. Transformer is also implemented after we discover the limitations of the above three basic models which are Long Term Dependency of Words, Exploding Gradient and Vanishing Gradient problem. Common steps followed in all models are Preprocessing, encoding and decoding. After studying all the models, a comparative analysis of models is done based on the Bilingual Evaluation Understudy (BLEU) score where we calculated Unigram(1-gram) and Bigram(2-gram) scores using reference and translated sentences.