Generative Adversarial Neural Machine Translation for Phonetic Languages via Reinforcement Learning

Neural Machine Translation (NMT) heavily depends on the context vectors generated via attention network for the target word prediction. Existing works primarily focus on generating context vectors from words or subwords of sentences, limiting NMT models' ability to learn sufficient information about the source sentence representations. These situations are even worse when languages belong to extremely low-resource categories due to rare word problem. To improve the learning of source sentence representations and handle the rare word problem of Low Resource Languages (LRLs), we propose a novel improvement in Generative Adversarial Networks (GAN)-NMT by incorporating deep reinforcement learning-based optimised attention in generator and convolutional neural network in discriminator. We also create the novel joint embedding of subwords and sub-phonetic representation of sentences as input to GAN that helps models to learn the better representations and generate suitable context vectors compared to existing traditional approaches for LRLs. To show the effectiveness of our method, we demonstrate experiments on LRLs pairs, e.g., Gujarati ↔ Hindi, Nepali ↔ Hindi, Punjabi ↔ Hindi, Maithili ↔ Hindi and Urdu ↔ Hindi. Our proposed novel approach suppress the existing state-of-the-art techniques with considerable improvement.