We propose a graphical model associated with categorical variables and study the problem of structure learning for this model. The model is naturally generalization of Ising model and the learning method is based on neighborhood selection by multinomial logistics regression. Group lasso was incorporated into the regression since parameters representing interaction between variables have group structure. The performance of proposed method is studies by simulations, and we apply this model to the music annotation dataset CAL500 to obtain the graphical model for music labels.