Rethinking Collapsed Variational Bayes Inference for LDA