Abstract
Automatic detection of hate speech and abusive language is crucial for combating online toxicity. This study explores Gaussian Naive Bayes for multi-label classification of hate speech on Indonesian Twitter, including target, category, and level. We combined TF-IDF features with contextual BERT embeddings. The model achieved balanced performance for general hate speech and good non-abusive language detection. However, it exhibited limitations with imbalanced data and specific hate speech types. The classifier consistently favored the majority class (non-hateful/non-abusive) across labels, particularly struggling with HS_Gender, HS_Physical, etc. This suggests difficulty detecting less frequent but potentially severe hate speech, likely due to limited training data. Overall accuracy and F1-scores confirm that while Gaussian Naive Bayes is efficient, it lacks robustness for nuanced multi-label classification with imbalanced datasets. This necessitates exploring alternative approaches for effectively detecting specific and less frequent hate speech.