Dimensional Emotion Detection from Categorical Emotion

• DOI: 10.48550/arxiv.1911.02499
• Publication Date: 2019-01-01
• Authors: Sungjoon Park; Jiseon Kim; Seonghyeon Ye; Jaeyeol Jeon; Hee Young Park; Alice Oh
• Keywords: Categorical variable; Dominance (genetics)

We present a model to predict fine-grained emotions along the continuous dimensions of valence, arousal, and dominance (VAD) with corpus categorical emotion annotations. Our is trained by minimizing EMD (Earth Mover's Distance) loss between predicted VAD score distribution distributions sorted VAD, it can simultaneously classify categories scores for given sentence. use pre-trained RoBERTa-Large fine-tune on three different corpora labels evaluate EmoBank scores. show that our approach reaches comparable performance state-of-the-art classifiers in classification shows significant positive correlations ground truth Also, further training supervision leads improved especially when dataset small. also examples predictions appropriate words are not part original