This paper addresses the problem of unsupervised group discovery in social networks. We adopt a nonparametric Bayesian framework that extends previous models to networks where the interacting objects can simultaneously belong to several groups (i.e., mixed membership). For this purpose, a hierarchical nonparametric prior is utilized and inference is performed using Gibbs sampling. The resulting mixed-membership model combines the usual advantages of nonparametric models, such as inference of the total number of groups from the data, and provides a more flexible modeling environment by quantifying the degrees of membership to the various groups. Such models are useful for social information processing because they can capture a user's multiple interests and hobbies.