Unpaired Abstract-to-Conclusion Text Style Transfer using CycleGANs

The availability of paired examples greatly facilitates the task of style transfer by allowing the use of supervised learning. However, our scenario does not enjoy such a condition. We focus on style transfer for academic writing, and examine the possibility of performing style transfer between sentences from the abstract and conclusion sections of a scientific article in the Natural Language Processing field, in both directions. We assume a latent correlation between the abstract and conclusion styles, and construct an unpaired data set. We propose the use of a version of CydeGAN based on transformers to perform the task. Our approach is shown to realize differences in tense or word usage which are characteristic of the different sections.