Rethinking Hate Speech Detection on Social Media: Can LLMs Replace Traditional Models? (arXiv)

Detecting hate speech on social media is complicated by linguistic diversity, informal expression, and challenges like code-mixing, transliteration, and cultural nuance. While fine-tuned models like BERT are standard, recent large language models (LLMs) outperform them and may redefine the field. To demonstrate this, the IndoHateMix dataset was introduced, capturing Hindi-English code-mixed content for testing robustness in multilingual settings. Experiments show LLMs like LLaMA-3.1 consistently beat BERT-based models, even with less data. Their adaptability and generalization suggest a promising shift in combating online hate—raising the debate over prioritizing model development versus expanding diverse datasets.

https://arxiv.org/abs/2506.12744