Adapting the CycleGAN Architecture for Text Style Transfer

Text Style Transfer, the process of transforming text from one style to another, has gained significant attention in recent years due to its potential applications in various Natural Language Processing (NLP) tasks. In this paper, we present a novel approach for Text Style Transfer using a Cycle Generative Adversarial Network (CycleGAN). Our method utilizes the adversarial training framework of CycleGAN to learn the mapping between different text styles in an unsupervised manner, without the need for paired data. By leveraging the cycle consistency loss, our model is able to simultaneously learn style transfer mappings in both directions, allowing for bidirectional style transfer. We conduct experiments on the Yelp dataset to evaluate the effectiveness of our approach. Our results illustrate that our proposed TextCycleGAN achieves reasonable performance in terms of style transfer accuracy and fluency considering the simple architecture adopted in both generators and discriminators, while also providing bidirectional transfer capabilities (negative-positive and positive-negative)