Exploiting transitivity in probabilistic models for ontology learning