Research on data integration has provided languages and systems able to guarantee an integrated intensionalrepresentation of a given set of data sources. A significant limitation common to most proposals is that only intensional knowledge is considered, with little or no consideration for extensional knowledge.In this paper we propose a technique to enrich the intension of an attribute with a new sort of metadata: the “relevant values”, extracted from the attribute values. Relevant values enrich schemata with domain knowledge; moreover they can be exploited by a user in the interactive process of creating/refining a query. The technique, fully implemented in a prototype, is automatic, independent of the attribute domain and it is basedon data mining clustering techniques and emerging semantics from data values. It is parametrized with various metrics for similarity measures and is a viable tool for dealing with frequently changing sources, as in the Semantic Web context.