Likelihood estimation of sparse topic distributions in topic models and its applications to Wasserstein document distance calculations