DETECTION OF BULLYING CONTENT IN ONLINE NEWS USING A COMBINATION OF RoBERTa-BiLSTM

This research aims to build a bullying-themed online news classification system with a combined approach of RoBERTa embedding and BiLSTM. RoBERTa is used to generate context-rich text representations, while BiLSTM captures temporal relationships between words, thereby improving classification performance. The research dataset consisted of news from reputable portals such as Kompas.com, Detik.com, and iNews.com, labeled according to keywords relevant to the theme of bullying. The results of the experiment showed that the model achieved 95.2% accuracy, 98.2% precision, 93.6% recall, and 95.8% F1-score. Although there are few prediction errors (false positives and false negatives), this model shows excellent performance in detecting and classifying bullying-themed news. The main contribution of this research is the development of a new approach that combines RoBERTa and BiLSTM for the classification of complex bullying-themed news. This approach not only improves the accuracy of classification but can also be implemented in automated systems to detect negative content. Thus, this research has the potential to support the creation of a healthier digital space and encourage more responsible media practices.