Transfer Learning Techniques for Detecting Emotions in Text
Shubhendu Banerjee, Pushpita Roy, Soumalya Roy, Rupsha Singha Ray, Panchatapa Mondal, Saptarshi Halder, Sanjeev Prakashrao Kaulgud · 2025
Many attempts have been made to use different classic deep learning models, including LSTM, GRU, and BiLSTM, to overcome the difficulties of automating textual emotion recognition. Large datasets, significant computer power, and a significant amount of training time are necessary for these models, however. They also have trouble doing effectively on short datasets and are prone to forgetting. Our goal in this study is to demonstrate how transfer learning approaches, even with little data and training time, may better capture the contextual meaning of text and enhance emotion recognition. We accomplish this by comparing the performance of a pre-trained model, Emotional-BERT, based on bidirectional encoder representations from transformers (BERT), with RNN-based models on two benchmark datasets. We pay particular attention to how training data volume affects model performance.