Hate Speech Detection using CoT and Post-hoc Explanation through Instruction-based Fine Tuning in Large Language Models

In recent times, both social media platforms and researchers have been working to identify hateful language using large language models. However, none of these efforts have focused on using instruction-based fine-tuning with explanations and additional context information in the detection process. This research study has conducted experiments with LaMini and Google Flan-T5 models to analyze the impact of CoT prompts and explanations in zero shot and few shot settings to enhance model interpretability. The analysis reveals that on average including target information in pipeline significantly improves model performance. There is also a considerable effect of adding the explanations into the pipeline, boosting accuracy. The research findings highlight the importance of context and structured prompts in improving model accuracy, highlighting the need for responsible development and deployment of hate speech detection systems in online environments.