<![CDATA[The Shopee Live streaming e-commerce feature generates a large volume of user reviews on the Google Play Store, yet the text is dominated by informal Indonesian, slang, and code-mixing, making it difficult to analyze using conventional sentiment analysis approaches. This study analyzes and compares the performance of IndoBERT and Multilingual BERT (mBERT) in classifying sentiment of Shopee Live feature reviews into three classes: Positive, Negative, and Neutral. The study follows the CRISP-DM framework, encompassing web scraping of reviews from January-June 2025, filtering, text preprocessing, manual labeling, dataset splitting (80:10:10), fine-tuning both models using HuggingFace Transformers on Google Colaboratory, and evaluation using confusion matrices. The cleaned dataset comprised 5,240 reviews with a class distribution of 44.96% Positive, 35.79% Negative, and 19.25% Neutral. Evaluation results on the test set show that IndoBERT achieved an Accuracy of 89.7%, Precision of 89.5%, Recall of 89.7%, and F1-Score of 89.6%, outperforming mBERT, which achieved an Accuracy of 84.3%, Precision of 84.5%, Recall of 84.3%, and F1-Score of 84.4%. Both models exhibited lower performance on the Neutral class compared to the Positive and Negative classes. These findings confirm that IndoBERT is a more optimal model for sentiment classification of informal Indonesian-language e-commerce reviews, and can be recommended as an AI-based analytical instrument for product development teams to automatically and continuously monitor user sentiment.]]>
