Propositionalization has recently received much attention in the ILP community as a mean to learn efficiently non-determinate concepts using adapted propositional algorithms. This paper proposes to extend such an approach to unsupervised learning from symbolic relational description. To help deal with the known combinatorial explosion of the number of possible clusters and the size of their descriptions, we suggest an approach that gradually increases the expressivity of the relational language used to describe the classes. At each level, only the initial object descriptions that could benefit from such an enriched generalization language are propositionalized. This latter representation allows us to use an efficient propositional clustering algorithm. This approach is implemented in the CAC system. Experiments on a large Chinese character database show the interest of using KIDS to cluster relational descriptions and pinpoint current problems for analyzing relational classifications.