Multi-label Classification for Hate Speech and Abusive Language in Indonesian-Local Languages

Each hate speech has a target, category, and level that needs to be detected to help the authorities prioritize hate speech cases that need to be solved first. Various studies have been conducted in Indonesia on abusive speech and hate speech and their targets, categories, and levels, but only in Indonesian and English. On the other hand, various local languages in Indonesia open up opportunities for hate speech to occur using the local language. This study aims to compare some of the best machine learning algorithms, transformation methods, and feature extraction techniques in classifying abusive language and hate speech and their targets, categories, and levels using Twitter data in Indonesian and local languages. This study uses five local languages in Indonesia with the most speakers: Javanese, Sundanese, Madurese, Minangkabau, and Musi (Palembang). The algorithms used are Support Vector Machine (SVM), Multinomial Naive Bayes (MNB), and Random Forest Decision Tree (RFDT) with Binary Relevance (BR), Classifier Chains (CC), and Label Powerset (LP) as transformation methods. The term weighting used in this study is TF-IDF with word n-gram and char n-gram features. The results showed that the SVM algorithm with the CC transformation method and unigram feature extraction gave the highest F1-score results, 66.33% for Javanese and 65.68% for Sundanese. In Madurese, Minangkabau, and Musi language data, the best F1-score was obtained using the RFDT algorithm with the CC transformation method and unigram feature extraction with F1-score 76.37%, 80.75%, and 77.34%.