GANs for text are interesting, but also require a little bit of work.
I haven't used them, but my understanding is that they (might?) require a continuous vector representation to be learned, which could be an important step in the rest of the experimentation?