Advancing Hate Speech Detection with Transformers: Insights from the MetaHate (arXiv)

Individuals and communities are at significant social, psychological, and bodily risk from hate speech, which includes insults and libelous posts. Platforms like X, Facebook, Instagram, and Reddit facilitate mass communication, but they also disseminate hate speech, which is increasingly connected to hate crimes in the real world. Effective automated detection techniques in a variety of social media scenarios are necessary to address this problem. Despite their potential, deep learning models such as CNNs, LSTMs, and RNNs have trouble with parallelization and long-term dependencies. The current study uses the MetaHate dataset, which consists of 36 datasets with 1.2 million samples, to investigate transformer-based models for hate speech recognition. Models such as BERT, RoBERTa, GPT-2, and ELECTRA are evaluated; the best results are obtained with fine-tuned ELECTRA (F1 score: 0.8980). Error analysis identifies three recurring issues: label noise, coded language, and sarcasm.

https://arxiv.org/abs/2508.04913