This repository is a summery of the project: "The Stylish Autoencoder" by galshachaf and RoyHirsch {Tel Aviv University}
Textual style transfer is a developing area of research. The majority of the works on the task try to disentangle the content and the style in the text using generative models and adversarial losses. In this work we propose a new method for trainingan attentive stylish text generation model without parallel data. Our model is easy to train, computationally efficient and scalable. Our model is a self-attention-based encoder decoder, trained in back translation manner with the supervision ofpretrained style classifier. The model with the train methodology as well as the train and test data can be found here.
Some sample results:
