Controlled text style transfer via noise enhancement of deep learning transformer

Abstract

Text style transfer is one of the controllable text generation tasks, which can convert text style attributes. The mainstream method is to separate the content and style of the text, and then combine the content vectors with other style vectors to generate. However, implicit expression cannot completely separate meaning and style, separation and recombination may also lead to a decrease in the naturalness and fluency of the text. Therefore, we propose a new idea, which first encodes the text into a latent representation, and iteratively optimizes the latent representation with the smallest changes to achieve style transfer. Introducing noise enhancement before generating results improves the robustness of the generated system and reduces the occurrence of individual results with large errors. Experiments show that our optimization method based on noise enhancement performs well on two public datasets, Yelp and Amazon. Result has excellent performance in three indicators: content preservation, transfer strength, and fluency.