Towards Efficient and Explainable Hate Speech Detection via Model Distillation (arXiv)

To stop hatred and nasty language from spreading online, automatic detection is crucial. By identifying and elucidating hate speech, we can raise awareness of its detrimental impacts. The majority of detection models, however, function as opaque black boxes that are difficult to understand and analyze. LLMs, or large language models, have shown promise in detecting hate speech and improving interpretability. However, they are computationally expensive to operate. The authors suggest employing Chain-of-Thought to extract explanations that assist the classification objective in order to condense large language models. It will be easier to employ these activities in operational situations if they have compact language models. The researchers show that distilled models outperform bigger models in classification performance while providing explanations of the same caliber, making hate speech detection more accessible, intelligible, and useful.

https://arxiv.org/abs/2412.13698