<![CDATA[The Pertamina Mandalika International Circuit is a major tourism destination in Lombok. This study develops an aspect-based sentiment analysis (ABSA) system extracting sentiment from 1,772 valid Google Maps reviews filtered from 4,671 raw reviews, covering nine aspects defined via topic modeling and keyword validation. A two-stage IndoBERT pipeline is applied with Focal Loss, mean pooling, a differential learning rate, Back-Translation augmentation, and 5-fold cross-validation with per-aspect threshold tuning. The final model achieves a macro-F1 of 0.882 for aspect detection (accuracy 0.976; exact match accuracy 82.58%; hamming loss 0.024) and a macro-F1 of 0.87 for sentiment classification. Against classical TF-IDF baselines, IndoBERT (macro-F1 0.882) outperforms a linear Support Vector Machine (0.727) and Logistic Regression (0.711) on aspect detection, though a McNemar test shows the gap versus SVM is not statistically significant (p = 0.104), while the gap between SVM and Logistic Regression is significant (p = 0.001). Because labels are produced via weak supervision, additional validation is conducted using human annotation on 313 reviews (17.7% of the data); inter-annotator agreement (Cohen's Kappa) for aspect detection is very low (κ = 0.159), indicating that weak-supervision metrics may not directly reflect ground-truth sentiment accuracy. Venue (41.5%), Scenery (29.9%), and Organization (18.4%) are the most frequently discussed aspects in the reviews. The findings are translated into an actionable promotion strategy matrix, though larger-scale human validation remains necessary before direct application.]]>
Copyrights © 2026
