<![CDATA[Social media has become one of the primary communication channels for Indonesians, but it has also contributed to the increasing circulation of hate speech and hoaxes, which can polarize communities and undermine public trust. This situation highlights the urgent need for an automated system capable of identifying harmful content quickly and accurately. This study aims to develop a multi-label classification model to detect hate speech and hoax-related content in Indonesian social media text using a Transformer-based architecture. The model employed is IndoBERT, a variant of the Bidirectional Encoder Representations from Transformers (BERT) that is specifically trained on a large Indonesian corpus, making it more contextually relevant and effective than multilingual models for this task. The dataset consists of 1020 Indonesian social media texts annotated with binary labels for two categories: hate speech (0 = non-hate, 1 = hate) and hoax (0 = non-hoax, 1 = hoax). The research process includes text cleaning, tokenization, handling class imbalance, fine-tuning the model, and evaluating its performance using exact match accuracy, precision, recall, and F1-score at both micro and macro levels. To address class imbalance in the hoax label, the RandomOverSampler (ROS) technique was applied to enhance model stability and generalization. The experimental results show that IndoBERT achieved an accuracy of 97.55%, indicating that the model performs effectively in detecting harmful content and provides meaningful support for digital content moderation within the Indonesian online ecosystem.]]>
Copyrights © 2026
