<![CDATA[Emotion analysis in product reviews plays a crucial role in Indonesia’s e-commerce landscape, yet training advanced transformer models demands high computational resources, posing a significant constraint for small businesses. This research evaluates the Low-Rank Adaptation (LoRA) technique for optimizing emotion recognition in Indonesian product reviews, with a focus on resource efficiency. Indobert-large-p2, Indobert-base-uncased, and Indobertweet-base-uncased models were trained on the PRDECT-ID dataset, comparing the performance of full fine-tuning and LoRA. Results show LoRA provided competitive performance; on the Indobert-large-p2 model, the F1-score reached 67.81%, surpassing full fine-tuning’s 67.74%, despite training only about 1% of the total parameters. LoRA significantly reduced VRAM consumption by up to 17% and accelerated training duration by up to four minutes. Furthermore, the transformer models fine-tuned in this study, using both LoRA and full fine-tuning, consistently outperformed the results of previous research that employed complex architectures like CNN and BiLSTM, with the highest F1-score reaching 69.72% compared to the previous best of 66.13%. LoRA proves to be a practical and efficient solution for limited computational resources, enabling effective emotion analysis without expensive infrastructure.]]>
Copyrights © 2026
