LEVERAGING CONTINUAL FINE-TUNING FOR EMOTION CLASSIFICATION IN PRODUCT REVIEWS ON MSME SUSTAINABILITY SUPPORT

Abstract

Automatic analysis of consumer product reviews is essential for understanding granular customer perceptions beyond basic sentiment. While transformer-based models are prevalent in Indonesian sentiment analysis, their adaptation for multi-emotion classification shifting from broad polarities to specific affective states remains underexplored. This study addresses this gap by proposing a Continual Fine-Tuning (CFT) approach to adapt a pre-trained IndoBERTweet model from three sentiment categories into five distinct emotion classes: Happiness, Sadness, Fear, Love, and Anger. The novelty lies in the strategic repurposing of sentiment-oriented weights to capture nuanced emotional representations in Indonesian e-commerce discourse. Experimental results on the PRDECT-ID dataset demonstrate that the proposed CFT model achieves an accuracy of 0.8157 and a weighted F1-score of 0.8118, significantly outperforming traditional neural networks and multilingual baselines. The CFT model demonstrates a 2.13% improvement in accuracy compared to the base IndoBERTweet without continual tuning and a substantial 59.54% lead over the multilingual BERT (mBERT) baseline. Despite limitations concerning the dataset scale (5,400 samples) and inherent subjectivity in emotion labeling, this research provides a robust conceptual framework for model adaptation in the Indonesian NLP ecosystem. These findings suggest that CFT is an efficient strategy for enhancing the emotional intelligence of transformer models, especially in domain-specific tasks where high-quality labeled data is constrained.