Enhancing Indonesian Abusive Language Detection on Imbalanced News Comment Dataset Using Support Vector Machine with Oversampling

The growth of the internet facilitates ease of interaction and communication, including in the online news platforms with their comment sections. However, this phenomenon triggers the emergence of the online disinhibition effect, leading individuals to exhibit more aggressive behavior on the internet (toxic disinhibition). This has resulted in the prevalence of abusive language in online news comment sections, posing risks to the other readers and the subjects targeted by the abusive comments. In the context of abusive language detection, comments can be categorized into three classes, i.e., not abusive, abusive but not offensive, and abusive and offensive. Therefore, this study conducts a multiclass classification of Indonesian online news comments with those three categories. Experimental results indicate that SVM with word unigram TF-IDF serves as the best baseline classification model with a macro-average F1-score of 45.37%. We compared Easy Data Augmentation (EDA) and Synthetic Minority Oversampling Technique (SMOTE) to address the issue of extreme imbalance in the dataset. The results show that SMOTE outperforms EDA with a macro-average F1-score of 51.45% (an increase of 6.08% from baseline classification), while EDA achieves a macro-average F1-score of 49.83%. The random deletion (RD) operation was chosen for EDA, as other operations tended to decrease its performance. One drawback of EDA is the lack of a corpus for finding similar words. On the other hand, the classification results using SMOTE improve the generalization of the model more effectively. However, some issues persist, such as the abundance of Out of Vocabulary (OOV) occurrences leading to misclassification.