Real-world social networks are often hierarchical, reflecting the fact that some communities are composed of a few smaller, sub-communities. This paper describes a hierarchical Bayesian model based scheme, namely HSNPAM (Hierarchical Social Network-Pachinko Allocation Model), for discovering probabilistic, hierarchical communities in social networks. This scheme is powered by a previously developed hierarchical Bayesian model. In this scheme, communities are classified into two categories: super-communities and regular-communities. Two different network encoding approaches are explored to evaluate this scheme on research collaborative networks, including CiteSeer and NanoSCI. The experimental results demonstrate that HSN-PAM is effective for discovering hierarchical community structures in large-scale social networks.