Neural machine translation: past, present, and future

Deep neural networks (DNN) have achieved great success in several research areas like information retrieval, image processing, and speech recognition. In the field of machine translation, neural machine translation (NMT) has been able to overcome the statistical machine translation (SMT), which has been the dominant technology for a long-term span of time. The recent machine translation approach, which consists of two sub networks named an encoder and a decoder, has gained state-of-the-art performance on different benchmarks and for several language pairs. The increasing interest of researchers in NMT is due to its simplicity compared to SMT which consists of several components tuned separately. This paper describes the evolution of NMT. The different attention mechanism architectures are discussed and the purpose of each. The paper also presents some toolkits that are developed specifically for research and production of NMT systems. The superiority of NMT over SMT is discussed, as well as the problems facing NMT.