Detecting Fine-Grained Emotions from COVID-19 Tweets Using Transformer-Based Architecture

With the proliferation of social media, the task of identifying emotions in textual content has become increasingly relevant. Automating the task of detecting emotions from social media content has a wide range of applications today. In this work, we investigate the efficacy of pre-trained transformer models to detect fine-grained emotions from real-world tweets. We conduct experiments to fine-tune the DistilBERT and XLNet transformers to recognize nine emotions including neutral, anger, anticipation, disgust, fear, joy, sadness, surprise and trust. Our DistilBERT and XLNet based models exhibit an F1-score of 0.6429 and 0.6679 respectively. The results are promising when compared with similar transformer-based models for fine-grained emotion recognition. We conclude from our study that fine-grained emotion detection from microblogs using transformers although promising is also challenging. The challenge is primarily on account of two reasons: a) difficulty in differentiating between intrinsically similar emotions and b) difficulty in building a large and balanced annotated corpus from real-world tweets.