Label Smoothing for Emotion Detection (Student Abstract)

Automatically detecting emotions from text has countless applications, ranging large scale opinion mining to social robots in healthcare and education. However, are subjective nature often expressed ambiguous ways. At the same time, can also require implicit reasoning, which may not be available as surface- level, lexical information. In this work, we conjecture that overconfidence of pre-trained language models such BERT is a critical problem emotion detection show alleviating considerably improve generalization performance. We carry out comprehensive experiments on four benchmark datasets calibrating our model predictions leads an average improvement 1.35% weighted F1 score.