Comparative Study of IndoBERT and Few-Shot SentenceTransformer-LR for Sentiment Analysis of Indonesian Grok AI Reviews

Grok AI is a Large Language Model-based artificial intelligence assistant that attracted significant user interest in Indonesia, particularly following its transition from freemium to paid access in early 2026. This study compared full fine-tuning of Indonesian Bidirectional Encoder Representations from Transformers (IndoBERT) and few-shot SentenceTransformer with Logistic Regression (ST-LR) for three-class sentiment classification of Indonesian Grok AI reviews on Google Play Store, using a TF-IDF + Logistic Regression model as a classical baseline. A dataset of 2,419 clean reviews was constructed through text normalization and deduplication, then divided using a stratified 80:10:10 split. To account for the small neutral class, IndoBERT was trained over five random seeds and reported as mean with a 95% confidence interval (CI): 81.98% accuracy (95% CI [78.09, 85.87]) and 64.86% macro-averaged F1-score (95% CI [61.26, 68.45]). The TF-IDF + LR baseline reached 83.06% accuracy and 67.05% macro-F1, which falls within the IndoBERT confidence interval; a McNemar test confirmed the two models are not significantly different (p = 0.238). The best few-shot configuration, ST-LR 32-shot with 96 training samples, reached 67.36% accuracy and 72.64% weighted F1-score averaged over ten seeds. The neutral class was the principal limitation across all models due to underrepresentation and mixed-sentiment three-star reviews. Negative sentiment was primarily driven by the freemium-to-paid transition, while positive sentiment reflected response quality and ease of use. The dataset and source code are publicly available in a GitHub repository.