Learning Disentangled Textual Representations via Statistical Measures of Similarity