A Flow-Based Cross Projection Framework for Unsupervised Text Style Transfer

Unsupervised text style transfer aims at altering the style of the source text while maintaining its style-independent content. A common method is to disentangle the content and style representation in the latent space. However, the lack of a parallel corpus has led to the impracticality of disentangling the latent space, resulting in the loss of style-independent content information. To address this problem, this paper proposes a novel flow-based framework that explicitly models the distributions of different styles via normalizing flows. This approach enables us to perform style transfer without the need for disentangling content and style in the latent space. Furthermore, the invertible virtue of normalizing flows is utilized to establish cross-projection between the latent spaces of different styles, providing an effective method for transferring the style of a given text to a target style. Experimental results on two datasets demonstrate the effectiveness of the proposed model compared to several state-of-the-art baselines.