Few-shot short-text classification with language representations and centroid similarity.