Residual Information Flow for Neural Machine Translation

Automatic machine translation plays an important role in reducing language barriers between people speaking different languages. Deep neural networks (DNN) have attained major success in diverse research fields such as computer vision, information retrieval, language modelling, and recently machine translation. Neural sequence-to-sequence networks have accomplished noteworthy progress for machine translation. Inspired by the success achieved by residual connections in different applications, in this work, we introduce a novel NMT model that adopts residual connections to achieve better performing automatic translation. Evaluation of the proposed model has shown an improvement in translation accuracy by 0.3 BLEU compared to the original model, using an ensemble of 5 LSTMs. Regarding training time complexity, the proposed model saves about 33% of the time needed by the original model to train datasets of short sentences. Deeper neural networks of the proposed model have shown a good performance in dealing with the vanishing/exploding problems. All experiments have been performed over NVIDIA Tesla V100 32G Passive GPU and using the WMT14 English-German translation task.