Group decision making problems with linguistic preference information are studied. Two measures of group consensus are defined to reflect respectively the level of agreement among all the experts and the difference degree between the experts’ individual opinions and the collective one. Based on the consensus measure indicators, an optimization model of linguistic scales granulation is constructed. An improved PSO algorithm is established to obtain the optimal cutoff points for the linguistic scales. A group decision-making method is developed, and a ranking order of alternatives and the best alternative are obtained. Finally, a numerical example illustrates the feasibility and effectiveness of the proposed method.