The large number of SNS users brings marketers and managers huge opportunities and tough challenges simultaneously to extract managerial implications from SNS user behaviors. To gain insight into user behaviors, researchers divide users into roles (i.e. user groups) to analyze the difference of user behaviors between distinct roles. In traditional role discovery algorithms, the number of roles is intractable and predefined. User features implied in unstructured text data have been rarely used. In this paper, we propose a Dirichlet Process Mixture Model to automatically optimize the number of roles and integrate features mined from text data to analyze user roles.