Latent Dirichlet Allocation (LDA) is an attractive topic model research because LDA is so flexible for solving different problems. Because of its different core dependencies, it can be applied to many topics, such as emotion detection, information systems or image clustering. In recent works, researchers have focused on novel dependency to obtain perfect fitting to datasets. However, real world data is too diverse and abundant to be fitted with one single dependency. A single dependency model can only concentrate on the overall characteristic of datasets and thus ignores small details in the data. In addition, model selection for different situation is always difficult. As a result, we propose Multi-dependent Latent Dirichlet Allocation (MD-LDA). MD-LDA can be applied various dependencies into the model. We don't need to select a specific model. For each piece of data, MD-LDA can pick up the most optimal fitting dependency from the dependency set and therefore obtain the best dependencies for the dataset. We also apply some previous works into MD-LDA as a basis for comparison. In our experiments, MD-LDA exhibits the best performance in various cases and is an improvement compared to the other models under consideration.