Deep Text Style Transfer

We propose to transfer the content of a text written in a certain style to an alternative text written in a different style, while maintaining as much as possible of the original meaning. Our work is inspired by recent progress of applying style transfer to images, as well as attempts to replicate the results to text. Our model is a deep neural network based on Generative Adversarial Networks (GAN). Our novelty is replacing the discrete next-word prediction with prediction in the embedding space, which provides two benefits (1) train the GAN without using gradient approximations and (2) provide semantically related results even for failure cases.