Enhancing Hate Speech Detection in Low-Resource Code-Mixed Indonesian Tweets via GPT-Based Data Augmentation

Abstrak

Automatic hate speech detection in low-resource, code-mixed languages, such as Indonesian social media environments, presents significant challenges due to the scarcity of annotated data and the linguistic variability introduced by code-mixing. However, due to the growing prevalence of hate speech on social media, there is a need for robust hate speech detection systems. This study investigates the effectiveness of data augmentation strategies, specifically Generative Pretrained Transformer (GPT)-based paraphrasing and aggressive text transformation, in enhancing the performance of hate speech detection models for Indonesian code-mixed tweets. To achieve that, we employed traditional machine learning models, Recurrent Neural Network (RNN)-based models, and transformer-based models to assess the impact of these augmentation strategies. Our findings reveal that GPT-generated data improve model performance, with transformer-based models, including Indonesian Bidirectional Encoder Representations from Transformers (IndoBERT) and the Cross-lingual Language Model Robustly Optimized BERT Pretraining approach (XLM-RoBERTa).

Identifier

• OpenAlex: https://openalex.org/W7111017995
• DOI: https://doi.org/10.48084/etasr.14342