<![CDATA[Political polarization on Indonesian social media intensified during the August 2025 House of Representatives (DPR) demonstrations, where provocative and sarcastic tweets helped amplify institutional criticism and widen public conflict. This study examines event-based automatic detection of provocative political discourse by comparing a feature-based Support Vector Machine (SVM) classifier with a transformer-based IndoBERT model on a large-scale Indonesian Twitter (X) corpus collected from 15 August to 15 September 2025. Tweets were preprocessed and labeled using a rule-based proxy lexicon to distinguish provocative from neutral content, then both models were trained and evaluated under the same experimental setting. Results show that SVM is highly effective for recognizing explicit provocation expressed through repetitive and lexically salient slogans, whereas IndoBERT provides more stable detection of implicit and context-dependent provocation, including irony and sarcasm that are common in Indonesian political talk online. In addition, temporal exploration indicates sharp spikes in tweet volume that align with key offline protest moments, suggesting a close coupling between street-level mobilization and digital discourse dynamics. Overall, the findings support the use of contextual NLP models within event-centered social media analysis to strengthen scalable monitoring of polarization and to inform early-warning approaches for escalating conflict in Indonesia’s digital public sphere.]]>
Copyrights © 2026
