Abstract
Language style transfer is the problem of migrating the content of a sourcesentence to a target style. In many of its applications, parallel training dataare not available and source sentences to be transferred may have arbitrary andunknown styles. First, each sentence is encoded into its content and stylelatent representations. Then, by recombining the content with the target style,we decode a sentence aligned in the target domain. To adequately constrain theencoding and decoding functions, we couple them with two loss functions. Thefirst is a style discrepancy loss, enforcing that the style representationaccurately encodes the style information guided by the discrepancy between thesentence style and the target style. The second is a cycle consistency loss,which ensures that the transferred sentence should preserve the content of theoriginal sentence disentangled from its style. We validate the effectiveness ofour model in three tasks: sentiment modification of restaurant reviews, dialogresponse revision with a romantic style, and sentence rewriting with aShakespearean style.