In the last ten years, several approaches for knowledge discovery in databases based upon the construction of a concept lattice have been developed. Most of them are dedicated to binary or propositional descriptions. This paper presents an approach to build a particular concept lattice, called the Generalization Space, for relational descriptions. It is based upon an iterative reformulation of the descriptions into a sequence of languages more and more expressive. The anytime property of the algorithm allows one to use it on large databases, and experiments show that its complexity grows linearly with the number of descriptions.