Improving multilabel classification of hate speech and abusive language in Indonesian using MAML

Abstract

This study investigates automated multi-label detection of hate speech and abusive language (HSAL) in Indonesian social media, addressing challenges of data imbalance, especially in minority labels. Two training approaches are compared: standard supervised learning and meta-learning using the model-agnostic meta-learning (MAML) algorithm. IndoBERTweet-BiGRU is adopted as the baseline model, while MAML is leveraged to enhance generalization and adaptability with limited training data. Both models are trained on a multilabel dataset with 13 HSAL categories exhibiting highly imbalanced distributions. The best supervised model achieved an F1-Micro of 84.02% and an F1-macro of 77.97%, whereas the best MAML-trained model reached 84.12% and 76.85%, respectively. Although the overall gap is small, MAML demonstrates notable improvements on minority classes such as hate speech (HS) physical, gender, and race, shown through higher F1-score and area under the receiver operating characteristic curve (AUROC) values. These results highlight its strength in low-resource classification settings. This study is limited to Indonesian language and YouTube transcript contexts, and MAML incurs higher training complexity. Cultural and linguistic nuances also present potential bias in real-world use. Despite these constraints, the proposed system offers practical benefits by enabling fine-grained HSAL classification and supporting earlier detection of harmful online content.