Abstract
We propose a framework which makes a model predict fine-grained dimensionalemotions (valence-arousal-dominance, VAD) trained on corpus annotated withcoarse-grained categorical emotions. We train a model by minimizing EMDdistances between predicted VAD score distribution and \textit{sorted}categorical emotion distributions in terms of VAD, as a proxy of target VADscore distributions. With our model, we can simultaneously classify a givensentence to categorical emotions as well as predict VAD scores. We usepre-trained BERT-Large and fine-tune on SemEval dataset (11 categoricalemotions) and evaluate on EmoBank (VAD dimensional emotions), in order to showour approach reaches comparable performance to that of the state-of-the-artclassifiers in categorical emotion classification task and significant positivecorrelations with ground truth VAD scores. Also, if one continues training ourmodel with supervision of VAD labels, it outperforms state-of-the-art VADregression models. We further present examples showing our model can annotateemotional words suitable for a given text even those words are not seen ascategorical labels during training.