Exploiting likely-positive and unlabeled data to improve the identification of protein-protein interaction articles