In this paper we propose a general framework for computing similarity between concepts. This framework generalizes the Tversky’s model of similarity and uses a non-negative matrix factorization procedure to estimate abstract classes on which similarity between concepts may be computed. We applied the framework to semantic features used to describe concepts. Experimental results suggest that the general framework is feasible and this method is applicable across different concepts. This framework may be considered as a valuable measurement method to test hypotheses about category-specific disorders.