mT5-Based Sentiment Classification and Insight Generation for Indonesian E-Commerce Reviews Toward Sustainable Digital Economic Growth (SDG 8)

This study develops an automated tool for extracting insights from unstructured Indonesian e-commerce reviews, which are published in large numbers daily and are laborious to analyze manually. The tool combines a fine-tuned mT5-Base sentiment classifier with a rule-based keyword mapping module that assigns review keywords to four product dimensions: delivery, packaging, product quality, and price. The classifier was fine-tuned on 12,000 resampled reviews and tested on a held-out, naturally imbalanced test set of 2,398 reviews. It obtained a macro F1-score of 0.55, comparable to a BERT-based reference (0.56), but performed poorly on the neutral class (F1 = 0.19). Human evaluation of the generated insights yielded a mean relevance score of 4.45 out of 5.00, above the 3.50 threshold. These findings highlight the potential of accessible, data-driven tools to support merchant in making more informed business decisions while contributing to sustainable growth in the digital economy.