Neural Network Based Machine Translation Systems for Low Resource Languages: A Review

Machine translation of documents into regional languages has important role nowadays. Deep neural networks are used in neural machine translation (NMT), which is the process of converting a set of words from one source language to some other. It is a neural network-based, fully automated translation technique. Instead of just translating a word on its own, NMT takes into account the context in which a word is used to produce more accurate translation. Instead of starting with a set of established rules, the neural network in neural machine translation is in charge of encoding and decoding the source text. MT has several advantages as compared to the traditional translation techniques and approaches. Critical analysis of different approaches used for machine translation of low resource languages were done here. Deep learning based machine translation systems, transformer learning, transfer learning techniques are some of them. After the study it is concluded that nowadays NMTs developed by taking the advantages of Deep neural networks and transfer learning approaches. Gives better accuracy than other systems. Though it is a tedious task to convert one or multiple languages to another language with 100% of accuracy as manual translation, the machine translation systems developed with these techniques can score a remarkable accuracy. As there is a lack of large parallel corpora for most of the Indian languages, the translation process become more tedious. The role of transfer learning comes in this point. Transfer learning can improve translation of low resource languages, as it can use prior knowledge in translation of a separate language pair in machine translation. This is a work done for developing a translation system for low resource language pair like Sanskrit and Malayalam. There’re very less research works done in Sanskrit and Malayalam machine translation.