On Variational Learning of Controllable Representations for Text without Supervision

The variational autoencoder (VAE) has found success in modelling the manifold of natural images on certain datasets, allowing meaningful images to be generated while interpolating or extrapolating in the latent code space, but it is unclear whether similar capabilities are feasible for text considering its discrete nature. In this work, we investigate the reason why unsupervised learning of controllable representations fails for text. We find that traditional sequence VAEs can learn disentangled representations through their latent codes to some extent, but they often fail to properly decode when the latent factor is being manipulated, because the manipulated codes often land in holes or vacant regions in the aggregated posterior latent space, which the decoding network is not trained to process. Both as a validation of the explanation and as a fix to the problem, we propose to constrain the posterior mean to a learned probability simplex, and performs manipulation within this simplex. Our proposed method mitigates the latent vacancy problem and achieves the first success in unsupervised learning of controllable representations for text. Empirically, our method significantly outperforms unsupervised baselines and is competitive with strong supervised approaches on text style transfer. Furthermore, when switching the latent factor (e.g., topic) during a long sentence generation, our proposed framework can often complete the sentence in a seemingly natural way -- a capability that has never been attempted by previous methods.