Comparative Analysis of Naive Bayes and Support Vector Machine Algorithms in Classifying Hate Speech and Abusive Indonesian Text

Abstract

This study examines the effectiveness of Naive Bayes and Support Vector Machine algorithms in classifying hate speech and abusive Indonesian text, aiming to understand the strengths and failure patterns of each model. The research method uses an annotated tweet dataset that undergoes specialized preprocessing and TF-IDF feature extraction. The performance of the Multinomial Naive Bayes (MNB) and Linear Support Vector Classifier (LinearSVC) models is rigorously evaluated using a holdout test set, cross-validation, and statistical significance testing. The results confirm the significant superiority of LinearSVC over MNB, with an accuracy difference of 8.7%. In-depth analysis reveals that while the model is highly accurate in identifying clean content, the main challenge lies in classifying abusive text, which is often confused with hate speech. This error pattern indicates a blurred boundary between the two categories, suggesting a continuous spectrum rather than discrete classes. The discussion affirms that LinearSVC's advantage stems from its ability to handle high-dimensional data and the linguistic complexity typical of Indonesian social media. These findings highlight the need for modeling approaches that are more sensitive to linguistic nuances and content severity gradations. The study's implications encourage the refinement of data annotation schemes, exploration of methods such as multi-label classification or ordinal regression, and the development of hybrid moderation systems that prioritize the gray area between abusive text and hate speech for greater fairness and accuracy.