Exploratory Research on Harmful Speech Detection based on Large Language Models

With the rapid expansion of online platforms, harmful speech has emerged as a critical issue, challenging mental health and social order. As the amount of online review and content posting generated daily continues to increase, there is a pressing need for developing automated systems/models to detect and address harmful speech. Despite the achievement of harmful speech detection models based on machine learning and deep learning, the challenges of accurately identifying harmful content still remain. This study thus investigates the application of Large Language Models (LLMs) for harmful speech detection. We fine-tuned three LLMs, Llama3.1, Qwen2, and Qwen2.5 using a meticulously curated dataset of harmful speech examples. The experimental results revealed that the improved models attained higher performance levels, achieving accuracy rates above 80 %. This highlights the potential of LLMs to identify complex language patterns and context effectively when detecting harmful content in the context of Chinese language. Our fine-tuned models contribute to fostering safer online environments, and this research will serve as a reference for future advancements in harmful speech detection systems/models.