Discovering Latent Patterns with Hierarchical Bayesian Mixed-Membership Models