Chat bots, image to caption, and question answering are some more realistic use cases of the generator side. In those cases there is some input (previous chat message, an image, a question) which is encoded into a vector sometimes referred to as a context or thought vector. The decoder/generator unfolds that vector into a series of words.