Topic Modeling via Full Dependence Mixtures