Analysis of Provocative Speech During the 2025 DPR Demonstration on X Using the IndoBERTweet Method

Abstract

Social media platforms have become important channels for public discussion during political events. During the DPR demonstrations in August 2025, online discussions on X (formerly Twitter) contained various forms of expressions, including provocative speech that may influence public opinion and collective behavior. Detecting such content automatically is challenging due to the informal language, slang, and contextual nuances commonly found in social media texts. This study aims to analyze provocative speech on the social media platform X using text classification techniques and transformer-based models. A total of 8,899 Indonesian tweets related to the demonstration period from August 25 to August 31, 2025 was collected using the Tweet Harvest crawling tool. The dataset was manually labeled into two categories, namely provocative and non-provocative, using a majority voting approach by three annotators. Several preprocessing steps were applied, including cleaning, normalization, stemming, tokenization, and stopword removal. Several models were evaluated, including Multinomial Naïve Bayes, Linear Support Vector Machine, BiLSTM, IndoBERT, and IndoBERTweet. Experimental results show that transformer-based models outperform traditional machine learning approaches. The best performance was achieved by the IndoBERTweet model with a learning rate of 3×10⁻⁵, achieving an accuracy of 93.07% and an F1-score of 91.56%. These findings indicate that domain-specific language models are effective for detecting provocative speech in Indonesian social media discussions related to political events.