Improving Textual Emotion Recognition Based on Intra- and Inter-Class Variation

Textual Emotion Recognition (TER) is an important task in Natural Language Processing (NLP), due to its high impact in real-world applications. Prior research has tackled the automatic classification of emotion expressions in text by maximising the probability of the correct emotion class using cross-entropy loss. However, this approach does not account for intra- and inter-class variations within and between emotion classes. To overcome this problem, we introduce a variant of triplet centre loss as an auxiliary task to emotion classification. This allows TER models to learn compact and discriminative features. Furthermore, we introduce a method for evaluating the impact of intra- and inter-class variations on each emotion class. Experiments performed on three data sets demonstrate the effectiveness of our method when applied to each emotion class in comparison to previous approaches. Finally, we present analyses that illustrate the benefits of our method in terms of improving the prediction scores as well as producing discriminative features.