<![CDATA[Aspect-based sentiment analysis (ABSA) on Indonesian social media has steadily improved in accuracy, yet the field still lacks evidence on which ABSA subtasks are learnable on imbalanced, event-driven text and how much of a reported score reflects the protocol rather than the model. This study benchmarks relevance, aspect, and aspect-level polarity classification on 7,131 tweets from the August 2025 Indonesian demonstrations, against a documented gold standard of 6,075 relevant tweets. Majority-class prediction, four lexical models, a fine-tuned IndoBERT, and zero-shot GPT-4o-mini are compared under identical five-fold stratified cross-validation, with macro-F1 as the primary metric, with 95% paired-bootstrap and McNemar tests on identical folds, and with a repeated benchmark under near-duplicate-grouped and day-held-out splits. Headline accuracy is misleading for relevance: models reach 85-89% against an 85.19% majority baseline, yet lexical models detect only one in ten off-topic tweets (recall 0.099) and the fine-tuned encoder no more than half (0.495). Aspect classification is the most learnable subtask, at 0.772 macro-F1 for the fine-tuned encoder, and zero-shot GPT-4o-mini reaches 0.808 on the 78.9% of records it chose to label, against 0.711 when its omissions are counted as errors. Positive polarity is the bottleneck for every family: lexical models reach 0.306 F1 with 0.190 recall, and although contextual and zero-shot models recover positive recall to roughly 0.45, F1 stays in the 0.51-0.57 range on the records they label. Sentiment toward the people aspect is the hardest cell at 57.5% accuracy, against 86.6% for the government. The zero-shot comparator labels 84.2% of the relevant tweets for polarity, and its errors are systematic: 68.2% over-predict the negative class, conflating event valence with evaluative polarity. The study contributes a task-level, coverage-aware benchmark for Indonesian ABSA under class imbalance]]>
Copyrights © 2026
