A proposed model for neural machine translation of Sanskrit into English

Abstract

Sanskrit, one of the oldest of human languages, has an algorithmic, calculus like grammar which resembles a computer program. In this paper, we propose a model for machine translation of Sanskrit Sentences into English using recurrent neural networks. We have trained our Recurrent Neural Network by sequence to sequence examples, to account for infrequent cases like extra-long sentences and unusual words. Special weight factors are used during training to account for unusual words. We have employed parallel analysis of all words in source sentence to speed up the translation. Novelty in this work is use of a two pronged approach to find the most suitable word in target language for a word in source language. For simple words, consisting of a single words or a conjunction of at most two words, we use a partial bilingual dictionary and for words formed by conjunction of three or more words we use the output of machine learning classifier module as target word. In case of simple sentences, with up to 5 words, use of a combination of partial dictionary and a classifier improves the translation speed by 30% as compared to by using full dictionary only and accuracy by 10% as compared to using only the output of classifier for translation. We used Support Vector Machine classifier to find English word for a Sanskrit word in case of sentences with more than 5 words and it resulted in 10% more accurate translations as compared to that by using Naive Bayes Classifier.