Reviews: Unsupervised Text Style Transfer using Language Models as Discriminators

This paper proposes using the language models' log-likelihood score as a discriminator in order to train style transfer generation models. In particular, given a text x, the goal is to train an encoder and a decoder. The decoder should be able to restore x from (z_x, v_x), but generate a stylistically different sentence from z_x and a different vector v_y. The paper proposes to judge the style difference by a language model through its log-likelihood score. Despite the simple idea, the authors show that it works well on three style transfer tasks, and achieves comparable or better performances than the state-of-the-art adversarially trained models.