In computer-assisted structure elucidation, a large number of candidate structures for the unknown compound can be generated. A computer program designed to aid in the recognition of significant differences between these structures is described. The goal of the program is to group the candidates into meaningful classes with a minimum of input from the user; each class is distinguished by a substructure its members have in common. A computational approach is developed based on the combinatorial problem of set covering. The program evaluates its results using an information-theoretical criterion. An application of the program to a real-word structure problem is presented.