NMT for English-Marathi using RNNs & Attention

Neural Machine Translation (NMT) has led to significant changes in the world of machine translation. It is a huge improvement over Statistical Machine Translation (SMT). With the advent of the transformer model introduced in the paper “Attention is all you need” [1], (which introduces mechanism of self attention) the industry has seen significant improvements in translation, but the transformer model is resource & data intensive hence RNNs seem a better option in terms of data and computing constraints. While there have been numerous implementations of NMT for western to western language translations, there have been comparatively fewer implementations for Western to other languages especially Indic languages apart from Hindi. Indian languages also have various features like the use of honorifics, gendered objects, differences in usage like formal/informal setting, etc thus differing from most of the Western languages in a fundamental way where a lot of mentioned features are unavailable & language structure is rigid in nature. Our paper aims to find the best as well as give a comparative analysis of various NMT models for English to Marathi translations. Our models include - Plain LSTM, Plain GRU, Plain Bidirectional LSTM & GRU, LSTM with Attention, GRU with Attention, Bidirectional LSTM with Attention & Bidirectional GRU with Attention.