Cycle Generative Adversarial Networks for seq2seq

Sequence-to-sequence transduction is a wide class of problems. The one of the most famous such problem is neural machine translation. The common goal of problems in this class is to build the mapping between source and target sequence spaces, where the matched objects are consistent in some way. Supervised deep learning techniques, trained on the large amount of paired data, show the state-of-the-art results on the most of problems of this class. But in many cases, it's not possible to obtain a lot of pairs of sequences. On the other side, usually it is significantly cheaper to acquire unpaired sequences. In this paper in order to reduce the number of paired sequences I attempt to build semi-supervised sequence-to-sequence learning method by applying the ideas of cycle generative adversarial networks which are very successful on the image-to-image translation problem.