Polarized-VAE: Proximity Based Disentangled Representation Learning for Text Generation

Learning disentangled representations of real-world data is a challenging\nopen problem. Most previous methods have focused on either supervised\napproaches which use attribute labels or unsupervised approaches that\nmanipulate the factorization in the latent space of models such as the\nvariational autoencoder (VAE) by training with task-specific losses. In this\nwork, we propose polarized-VAE, an approach that disentangles select attributes\nin the latent space based on proximity measures reflecting the similarity\nbetween data points with respect to these attributes. We apply our method to\ndisentangle the semantics and syntax of sentences and carry out transfer\nexperiments. Polarized-VAE outperforms the VAE baseline and is competitive with\nstate-of-the-art approaches, while being more a general framework that is\napplicable to other attribute disentanglement tasks.\n