Evaluating BiLSTM  Performance with BERT, RoBERTa, and DistilBERT in Online Bullying News Detection

This study examines the performance of BiLSTM combined with three transformer-based word embeddings—BERT, RoBERTa, and DistilBERT—in classifying bullying news in online media. BiLSTM was chosen for its significant advantages in processing text sequences compared to traditional RNN and LSTM models. The study used a dataset of 2,800 articles from three major Indonesian news portals, with 2,000 articles for training and 800 for testing, labeled using the lexicon method. The testing results showed that the combination of BiLSTM and RoBERTa achieved the best performance, with an accuracy of 94% and a near-perfect precision of 99%. Statistical significance tests confirmed that BiLSTM with RoBERTa performs significantly better than with BERT or DistilBERT. These findings suggest that the BiLSTM and RoBERTa combination is the most effective for classifying bullying news, especially for new or unseen data. This research contributes to the development of automatic bullying content detection systems to enhance content moderation on news platforms.