Skip to content
Review Open access

Sentiment Analysis of Indonesian E-Commerce Product Reviews Using Indobert Fine-Tuning on A Multi-Source Combined Corpus

Oct 2026 · Journal of Artificial Intelligence and Engineering Applications (JAIEA) · 0 citations · 34 references

Abstract

Consumer reviews on Indonesian e-commerce platforms now arrive faster than manual analysis can process them, and IndoBERT has become a common choice for classifying them automatically. Many studies, however, train on a single data source and report one training run, with no documented hyperparameter search and no test of stability. This study addressed both shortcomings. The training corpus combined three public datasets that differ in writing style and domain: SmSA, PRDECT-ID, and Review Product Shopee, totalling 20,608 reviews labelled Positive, Negative, or Neutral. Evaluation ran in three stages: a focused random search over seven hyperparameter combinations, 5-fold cross-validation, and an A/B ablation study. Class weighting compensated for a 51.9% imbalance spread between classes, and the macro F1-score served as the primary metric. The best configuration (learning rate 3×10⁻⁵, batch size 8, three epochs) reached 94.19% accuracy and a macro F1-score of 90.31% on 2,064 test reviews, above the initial target of 0.85. Across five folds the mean macro F1 was 90.16%, with a standard deviation of 1.0%. Back-translation augmentation of the minority Neutral class raised recall but lowered precision, so the model without augmentation was kept as final.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.