Emotion Detection with Deep Neural Network and Contrastive Learning
Ruochen Shi, Tong Chen · 2022
Researchers expended much effort identifying and classifying emotions in sentences, as they were critical in Natural Language Processing (NLP). Typically, existing algorithms approach this problem as a multi-label classification challenge, and some of them are sufficiently mature and perform reasonably well. However, upon investigation, it became clear that all of these methods had a common limitation: the model can hardly depict the relationship among sample data in the space. In this research, I propose to integrate the advantages of contrastive learning in data representation with these models in order to improve performance. The studies are divided into two stages: first, they transform the typical Cross Entropy loss function employed during the training stage to a contrastive loss function; and second, they employ a compound loss function to emphasize advantages and minimize bad impacts. The improved experiment results as compared to the previous methods demonstrate the practicality of my proposed improvement.