Rough set theory based attribute selection clustering approaches for categorical data have attracted much attention in recent years. However, they have some limitations in the process of selecting clustering attribute. In this paper, we analyze the limitations of three rough set based approaches: total roughness (TR), min-min roughness (MMR) and maximum dependency attribute (MDA), and propose a mean mutual information (MMI) based approach for selecting clustering attribute. It is proved that the proposed approach is able to overcome the limitations of rough set based approaches. In addition, we define the concept of mean inter-class similarity to measure the accuracy of selecting clustering attribute. The experiment results show that the accuracy of selecting clustering attribute using our method is higher than that using TR, MMR and MDA methods.