With the development of web 2.0, users are becoming more and more deeply involved in Internet, not only as readers, but also as authors. Wording preference is a well-known phenomenon that different people probably use different words even when they talk about the same topic. We think this phenomenon has a great impact on modeling texts by different authors, especially on topic modeling. This paper proposes a way to model user's preference by Dirichlet process (DP) in a topic model frame. Experiments show that our model outperforms the hierarchical Dirichlet process mixture model (DPMM) on a corpus of social tagging data from del.icio.us. Combination of user's preference can not only bring better performance on normal topic modeling task, but also discover the user's preference.