An Enhanced Context-based Emotion Detection Model using RoBERTa

Emotions are integral in conveying information in a particular context. For example, the most basic questions can have multiple answers, and the only way to pinpoint the correct answer for a particular question is by understanding the proper context behind it. Therefore, emotion detection plays a crucial part in a deeper understanding of the subject, whether in conversation or text. In this paper, the proposed approach EmoRoBERTa is an attempt to build a more robust emotion detection model, which can be implemented in various NLP tasks such as semantic and propaganda analysis, that involve the heavy usage of emotions. The approach involves combining a pretrained, fine-tuned RoBERTa model, and fitting it on a GoEmotions dataset, which includes twenty-seven different emotions and a neutral one, the most detailed taxonomy currently present. The model was then tested on three different emotion taxonomies and rendered desirable results, a higher MARCO-F1 score than the state-of-the-art model currently being used.