Learning Topic Models and Latent Bayesian Networks Under Expansion Constraints