Inferring ground truth from multi-annotator ordinal data: a probabilistic approach