On the role of human and machine metadata in relevance judgment tasks