Performance comparison of deep learning approaches for Indonesian twitter hate speech detection using IndoBERTweet embedding

The growth of social media has enabled the spread of online hate speech, while detecting hate speech in Indonesian social media, particularly Twitter/X, remains a challenge due to the language’s complexity. This study compares five deep learning models: LSTM, BiLSTM, CNN, CNN-LSTM, and CNN-BiLSTM, leveraging IndoBERTweet embedding to detect Indonesian Twitter hate speech. Our results show that combining CNN and BiLSTM worked best, achieving 86.75% accuracy and 84.39% F1-score. The CNN-LSTM combination also performed well with 86.28% accuracy and 83.5% F1-score. On the other side, CNN outperformed LSTM and BiLSTM, achieving 85.2% accuracy while LSTM and BiLSTM achieved 84.16% and 84.12% accuracy, respectively. These findings highlight the effectiveness of combining IndoBERTweet word embedding with CNN’s ability to capture pivotal patterns and LSTM/BiLSTM’s ability to capture contextual understanding.