Transformers Unveiled: A Comprehensive Evaluation of Emotion Detection in Text Transcription
Sara Ali, Bushra Naz, Sanam Narejo, Sahil Ali, Jitander Kumar Pabani · 2024
The fast growth of digital communication via smartphones and social media has made textual data as one of the most essential forms of people's expressions. Accurate emotion discovery from text is thus important for applications in customer sentiment analysis and interactive technologies, be it a web, desktop, or mobile application, an intelligent robot, or minor connected devices. This paper is intended to evaluate the performance of five state-of-the-art transformer models, including BERT, ALBERT, DistilBERT, Transformer XL, and RoBERTa, for multiclass emotion detection using the IEMOCAP dataset. From the results, it is observed that DistilBERT has the highest accuracy of 79.2%, implying an effective approach in balancing model size and performance. These research show that smaller models like DistilBERT often outperform larger models by reducing their overfitting and computational demands. The findings also highlights the need for task-specific fine-tuning and optimization in emotion detection.