Tackling Online Hate Speech Using Large Language Models

Online Hate speech poses significant challenge in online space through methods like polarization, misinformation and psychological harm. It fosters a myriad of other problems such as polarization, misinformation, and mental trauma. Most moderation policies that exist deal with content removal. Active engagement strategies are often absent. Automated counternarratives are an interventionary solution by responding to hateful messages with context-sensitive non-violent replies that transform them into positive, constructive dialogues. This study analyzes the counter-speech generation model leveraging large language models (LLMs) and the CROWDCOUNTER dataset composed of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{3, 4 2 5}$</tex> annotated hate speech-counterspeech. Our strategy uses Few shot prompting, Chain of Thought (CoT) prompting to improve reasoning and applies LoRA, a low resource fine-tuning approach, to facilitate extensive training with constrained resources. The workflow includes prompt design, data preparation, tokenization, model training and fine tuning, and performance measurement using benchmark scoring methods such as as BLEURT, ROUGE, METEOR. Models like Flan-T5-base, Flant-T5-large, GODEL-small, GODEL-base, Mistral-7B-Instruct are evaluated here in there capability to generate counter speech. We observe that flan-t5 performs best with fewshot prompts and Mistral-7B-Instruct performs best with CoT prompts.