Contrastive learning from label distribution: A case study on text classification

over the strong RoBERTa-Large baseline on ten datasets. Particularly, our contrastive learning in the label space can effectively capture label correlations better than that in the embedding space, achieving an improvement of