Our study registers in the framework of the automatic construction of classifications. We tackle an issue which has been less explored, that of the discovery of classifications. To tackle this problem we have chosen to pursue and develop the works of Mineau in the domain of the organization of knowledge bases using generalization [20]. We propose an original approach, called COING, to the discovery of classifications of structured objects represented using conceptual graphs. This approach consists in building a space of concepts which generalize the objects descriptions, called the Generalization Space, and then exploring this space so as to iteratively extract one or several conceptual classifications [2]. In this paper, we describe the method of construction of the Generalization Space that has been implemented in COING. This method is an extension of the MSG proposed by Mineau [19], which enables to account for knowledge about the types during the construction of the Generalization Space. However, we propose a formaliuzation of the use of constraints in the construction of the Generalization Space. We present empirical results of the method proposed that show its effiency.