Let classifiers filter unlabeled documents during training

Only labeled (with 'class' field) documents should be used during training and therefore each classifier should filter (and not fail when run against) them.