Abstract
Style transfer is the task of rephrasing the text to contain specificstylistic properties without changing the intent or affect within the context.This paper introduces a new method for automatic style transfer. We first learna latent representation of the input sentence which is grounded in a languagetranslation model in order to better preserve the meaning of the sentence whilereducing stylistic properties. Then adversarial generation techniques are usedto make the output match the desired style. We evaluate this technique on threedifferent style transformations: sentiment, gender and political slant.Compared to two state-of-the-art style transfer modeling techniques we showimprovements both in automatic evaluation of style transfer and in manualevaluation of meaning preservation and fluency.