A Fine-Tuned Distilled Zero-Shot Student Model for Emotion Detection in Academic-Related Responses

Abstract

Exploring emotions in academic settings, particularly in feedback on educational programs, is critical for understanding student experiences and improving educational policies. While the fine-tuning of pre-trained models has consistently delivered state-of-the-art results in emotion detection tasks, the potential of zero-shot learned model in this area remains largely unexplored. This paper presents a novel approach to emotion detection by fine-tuning a distilled zero-shot student model for classifying emotions in text, specifically focusing on feedback from beneficiaries of a free education program in the Philippines. Basic data cleaning and tokenization were performed, while retaining the stopwords in the corpus. Stopwords, in this work, contributes in understanding emotional expressions within academic-related texts. Our experiments highlight the superior performance of the distilled zero-shot student model achieving 84.21% accuracy and 84.23% F1-Score, notably outperforming the EmoRoberta model. The model exhibits excellent predictive ability in identifying emotions like desire, gratitude, and neutral, but encounters confusion in classifying optimism and relief. We deployed the model for automatic emotion labeling of feedback texts. Analysis revealed a predominantly positive reception on the program among its beneficiaries, with feelings of relief, approval, and gratitude being the most prominent. However, the presence of neutral and disappointment also highlights areas where the program needs improvement. These insights can be valuable for policymakers to understand the program impact of and to make data-driven decisions for its improvement and targeted interventions.