Neural Machine Translation is the latest flavor of data-driven Machine Translation (MT) approaches that allows the development of machine translation engines from collections of translated text. This approach offers elegant solutions to some of the fundamental challenges of statistical estimation, such as generalization of seen examples and inclusion of maximal context. Neural Machine Translation systems have shown superior performance in evaluation campaigns and in-house testing. However, in order to bring these systems to market for typical targeted applications, some problems still need to be addressed (such as domain adaptation, inclusion of diverse types of data, handling of specific problems such as tags or unit conversion, training and decoding speed, etc.), and there is active work on these issues which we discuss in this presentation.
