A Hybrid Technique for English to Urdu Machine Translation Using Transfer Learning

Machine Translation is a technique that helps in overcoming the linguistic barriers often encountered in the global world. Multiple techniques are being employed for automatic machine translation. Encoder-decoder models are frequently being implemented to improve the performance of automatic translation systems. Convolutional Neural Network (CNN) is a novel approach used for extracting textual features from the dataset. Merging recurrent neural networks (LSTM and GRU) with CNNs has presented fruitful results when working with low-resource language translation like English-to-Urdu translation. To reduce the workload of the model, transfer learning is implemented by using the Fasttext word embedding model. To evaluate the models, BLEU, TER, METEOR precision, recall, and F-measure are the parameters which are calculated. It can be perceived from the results that hybrid models (i.e. a combination of LSTM and GRU with CNN) demonstrated better results when compared to the LSTM and GRU alone.