Conditional generation of text with disentangled sentence representations

Abstract

Learning disentangled representation of a sentence aims to discover a representation which separates explanatory generative factors in a sentence. Existing methods extract a disentangled representation via independence constraints such as statistical independence between latent variables. We observed that previous approaches fail to encode enough information into low-dimensional latent variables and generator neglects that latent variables. In this paper, we propose two auxiliary losses to address this issue: the mutual information loss that encourages the encoder to maximize mutual information of a latent variable and data, and the Bag-of-Words similarity loss that controls and measures the influence of the permutation of single latent variable to the generator. Through the experiments on a sentiment transfer task, we prove the sentence representation can be disentangled and all latent variables involve in the sentence generation. We also show our framework can successfully learn disentangled jointly continuous and discrete representations in a semi-supervised manner.