Reviews: Content preserving text generation with attribute controls

The authors aim to achieve style transfer in text governed by controllable attributes. For this, authors propose a model which builds on prior works on unsupervised machine translation and style transfer. The proposed model uses reconstruction and back-translation losses. Authors propose to add denoising via dropping inputs and using an interpolated hidden representation. An additional adversarial loss is added to ensure the generated distribution of (hx,l) matches the input data distribution.