Multilabel Hate Speech Classification in Indonesian Political Discourse on X using Combined Deep Learning Models with Considering Sentence Length

Abstract

Hate speech, as public expression of hatred or offensive discourse targeting race, religion, gender, or sexual orientation, is widespread on social media. This study assesses BERT-based models for multi-label hate speech detection, emphasizing how text length impacts model performance. Models tested include BERT, BERT-CNN, BERT-LSTM, BERT-BiLSTM, and BERT with two LSTM layers. Overall, BERT-BiLSTM achieved the highest (82.00%) and best performance on longer texts (83.20% ) with high and, highlighting its ability to capture nuanced context. BERT-CNN excelled in shorter texts, achieving the highest (79.80%) and an of 79.10%, indicating its effectiveness in extracting features in brief content. BERT-LSTM showed balanced and across text lengths, while BERT-BiLSTM, although high in r, had slightly lower on short texts due to its reliance on broader context. These results highlight the importance of model selection based on text characteristics: BERT-BiLSTM is ideal for nuanced analysis in longer texts, while BERT-CNN better captures key features in shorter content.