Emotion Detection in Conversation Using Class Weights

Emotion detection in conversations is a crucial research problem for developing a human-like intelligence system. The task is to predict an emotion label for all the utterances in a conversation. We observed that datasets generally have an imbalance in classes. This can cause inefficient testing results as the model will be trained biased towards the majority class. This paper proposes that using class weights to model the architectures will improve the state-of-an-art benchmark results. The proposed approach provides a way to achieve good classification results even when the datasets are skewed. This helps the model to achieve a better prediction score; especially by considering the minority class samples in a balanced manner along with majority class samples. We perform the prediction on two publicly available datasets, DailyDialog and IEMOCAP, unimodal and multimodal, respectively. An improvement in the F1 scores was observed for both datasets. However, DailyDialog showed significant improvement than IEMOCAP because a higher imbalance was seen in the former than the latter.