Architecture of a machine translation system

The paper describes architecture of a Neural Machine Translation (NMT) system. The subject is brought up since NMT, i.e., translation using artificial neural networks, is now a leading Machine Translation paradigm. The NMT systems manage to deliver much better quality of output than the machine translators of the previous generation (statistical translation systems) do. Yet, the translation they produce still may contain various errors and it is relatively inaccurate compared with human translations. Therefore, to improve its quality, it is important to see more clearly how an NMT system is built and works. Commonly, its architecture consists of two recurrent neural networks, one to get the input text sequence and the other to generate translated output (text sequence). The NMT system often has an attention mechanism helping it cope with long input sequences. As an example, Google's NMT system is taken as the Google Translate service is one of the most highly demanded today, it processes around 143 billion words in more than 100 languages per day. The paper concludes with some perspectives for future research.