Self-attention-Based Deep Learning Approach for Machine Translation of Low Resource Languages: A Case of Sanskrit-Hindi

The automated translation of natural language text using software is known as machine translation (MT). Natural language’s intrinsic complexity and incompatibility have made machine translation (MT) a difficult process that is full of difficulties, especially in comparison to a physical conversion. The development of neural-based methods has brought machine translation outcomes closer to what people would expect. The seq-to-seq strategy, which uses encoder and decoder pairs, recurrent NN, complicated convolutions, and transformers, is one of the most modern deep learning techniques. This work proposes a self-attention deep learning architecture for MT that is built on a transformer with 8 layers of encoders and 8 layers of decoders and has all layers completely dedicated to multi-head attention. The important source and outcome of earlier layers that flow into the subsequent layer are weighted summarised to highlight the complete contribution of the proposed model. Comparing this process inputs that are not hybrid, the transformation produced is more exact. A parallel corpus for the Sanskrit-Hindi language pair is created in this study, and the model is assessed using the created corpus. The tests are carried out on Google Collab instances with GPUs, and the outcomes reveal an accuracy of 53.4 BLEU, which is an improvement over the earlier work.