Optimizing Hyper-Parameters for Sanskrit to English Neural Machine Translation with LSTM and GRU Models

This paper describes and evaluates different approaches to a simple Sanskrit-English translation using a very low-resource language dataset. Long-Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) techniques form the main component of the neural network architecture. The models use Encoder-decoder based sequence-to-sequence framework. The Bhagavad Gita dataset comprising of 700 verses from Sanskrit to English form the bilingual parallel corpus. The performances of the developed models are evaluated with different hyper-parameters like GRU with bidirectional network, LSTM with attention, GRU with embedding dimensions, etc. in order to improve the accuracy and quality of the translation output. The source being Sanskrit language and target being English language. The models are also evaluated on other standard parameters like number of epochs, more number of layers for LSTM or GRU with a combination of other parameters. Finally we have proposed the hyper-parameters suitable for a simple LSTM and GRU variant for producing effective translation. The standard metrics such as BLEU score, METEOR score, Word Error Rate (WER) and human translations on benchmark datasets have been adopted for better evaluation of the predicted results. Experimental results demonstrate that the LSTM-based NMT model with certain standard hyper-parameters outperforms baseline systems and achieves competitive translation accuracy for Sanskrit-English language pairs. The reader can easily find out the hyper-parameters suitable for developing a simple Sanskrit-English machine translation leveraging the LSTM and GRU techniques from the paper. The NMT models after the fine-tuning of the hyper-parameters have achieved a METEOR score of 0.675 in case of GRU-based translation and 0.648 in case of LSTM-based translation.