Building a shared and widely accessible repository, in order for scientists and end users to exploit it easily, results in tackling a variety of issues. Among others, the need for automatic labelling of available resources arises. We present an architecture in which machine learning techniques are exploited for resources classification and understanding. Furthermore, we show how learning tasks can be carried out more effectively if training sets and learned theories are expressed by means of Resource Description Framework (RDF) formalism and the storage/retrieval/query operations are managed by an ad hoc component.