Performance comparison of deep learning approaches for Indonesian twitter hate speech detection using IndoBERTweet embedding (Procedia Computer Science)

Social media’s explosive expansion has accelerated the propagation of hate speech online, making it difficult to identify on Indonesian platforms because of the country’s complicated linguistic structure. LSTM, BiLSTM, CNN, CNN-LSTM, and CNN-BiLSTM are five deep learning architectures that are evaluated in this study for the detection of hate speech on Twitter/X using IndoBERTweet embeddings. The CNN-BiLSTM model performed the best, according to the results, with 86.75% accuracy and 84.39% F1-score. CNN-LSTM came in second with 86.28% accuracy and 83.5% F1-score. Additionally, CNN fared better than LSTM and BiLSTM alone. These results show how well IndoBERTweet embeddings work with CNN’s pattern recognition and LSTM/BiLSTM’s contextual modeling to detect hate speech in Indonesia.

https://www.sciencedirect.com/science/article/pii/S1877050925027796