With the rapid development of information technology, large amounts of text data have been produced, collected and stored. Topic modeling is one of the important tools in text analysis, and is widely used for large text collection analysis. However, the topic model usually cannot be combined with users' domain knowledge intuitively and effectively during the topic modeling process. In order to solve this problem, this paper proposes an interactive visual analysis system to help users refine generated topic models. First, the hierarchical Dirichlet process is modified to support the word constraints. Then, the generated topic models is displayed via a matrix view to visually reveal the underlying relationship between words and topics, and semantic-preserving word clouds is used to help users find word constraints effectively. User can interactively refine the topic models by adding word constraints. Finally, the applicability of this new system is demonstrated via case studies and user studies.