A concept learning framework for terminological representations is introduced. It is grounded on a method for inducing logic decision trees as an adaptation of the classic tree induction methods to the Description Logics representations adopted in the Semantic Web context. Differently from the original setting of logical trees based on clausal representations, tree-nodes contain terminological concept descriptions (corresponding to OWL-DL classes) which makes it appealing for the Semantic Web applications. The method has been implemented in a prototypical system which has been experimentally evaluated on real ontologies.