Text Style Transfer with Confounders
Existing methods for style transfer operate either with paired sentences or distributionally matched corpora which differ only in the desired style. In this paper, we relax this restriction and consider data sources with additional confounding differences, from which the desired style needs to be inferred. Specifically, we first learn an invariant style classifier that takes out nuisance variation, and then introduce an orthogonal classifier that highlights the confounding cues. The resulting pair of classifiers guide us to transfer text in the specified direction, creating sentences of the type not seen during training. Experiments show that using positive and negative review datasets from different categories, we can successfully transfer the sentiment without changing the category.
PDF Abstract
