The first line of approaches disentangle text into its content and attribute in the latent space, and apply generative modeling.

Hamilton College

With the success of deep learning in the last decade, a variety of neural methods have been recently proposed for TST. If parallel data are provided, standard sequence-to-sequence models are often directly applied. However, most use cases do not have parallel data, so TST on non-parallel corpora has
become a prolific research area

Approaches of TST

Jin, Di, et al. “Deep Learning for Text Style Transfer: A Survey.” ACL Anthology, https://aclanthology.org/2022.cl-1.6/. 

Deep Learning for Text Style Transfer: A Survey

Prototype editing extracts a sentence template and its attribute markers to generate the text. 

Prototype Editing

Disentanglement

Pseudo-parallel corpus construction trains the model as if in a supervised way with the pseudo-parallel data.

Learn Before

Related