RL-ST: Reinforcing Style, Fluency and Content Preservation for Unsupervised Text Style Transfer

Bhargav Upadhyay, Akhilesh Sudhakar, Arjun Maheswaran

Keywords: style transfer, text generation, reinforcement learning, sentiment transfer, RL

TL;DR: A reinforcement learning approach to text style transfer

Abstract: Unsupervised text style transfer is the task of re-writing text of a given style into a target style without using a parallel corpus of source style and target style sentences for training. Style transfer systems are evaluated on their ability to generate sentences that 1) possess the target style, 2) are fluent and natural sounding, and 3) preserve the non-stylistic parts (content) of the source sentence. We train a reinforcement learning (RL) based unsupervised style transfer system that incorporates rewards for the above measures, and describe novel rewards shaping methods for the same. Our approach does not attempt to disentangle style and content, and leverages the power of massively pre-trained language models as well as the Transformer. Our system significantly outperforms existing state-of-art systems based on human as well as automatic evaluations on target style, fluency and content preservation as well as on overall success of style transfer, on a variety of datasets.

Code: https://anonymous.4open.science/r/d2c45647-1f9d-4f6c-a51c-acccafbf7cf4/