Multilabel Classification Abusive Language and Hate Speech on Indonesian Twitter Using Transformer Model: IndoBERTweet & IndoRoBERTa

Recently, the rampant hate speech and abusive language on Indonesian Twitter has become a concern. The overlap between the two makes it difficult to distinguish clearly. Legal action can be taken against those who spread hate speech due to its serious impact. Thus, multi-label classification in NLP, which utilizes a context-aware model, becomes very important. This study proposes the use of transformer models such as IndoBERTweet, IndoBERT, Indonesia RoBERTa Base, IndoRoBERTa Small and will be compared with the accuracy of previous studies [1], [2]. Various data balancing methods such as SMOTE, Random Undersampling, and Random Oversampling are applied to overcome the problem of imbalanced data in this dataset [1], which are then evaluated for their impact on the model. The results show that IndoBERTweet, with random oversampling and optimal hyperparameters (learning rate 1e-4, batch size 64, 3 epochs), outperforms other models with an accuracy of 0.86, an average precision of 0.85, a recall of 0.86, and an F1 score of 0.85. This study improved the accuracy by almost 10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> over previous research. Furthermore, data balancing can be useful in improving accuracy in some scenarios such as random oversampling in the IndoBERTweet model, but its effectiveness is inconsistent across models and configurations. Therefore, it is important to carefully consider the choice of data balancing technique and its impact on model performance in the context of a particular task and dataset.