Combating false information in Emergencies: Leveraging LLMs for Targeted False Information and Clarification Generation

Abstract

The rapid dissemination of false information during emergencies has become a significant challenge in managing online public opinion. While large language models (LLMs) have enhanced the speed and efficiency of information generation, they have also exacerbated the complexity of public sentiment by facilitating the spread of both false and clarifying information. This paper addresses the critical need for targeted clarification information to counteract false narratives during emergencies. We propose a novel approach by fine-tuning open-source LLMs to generate both false information and corresponding clarification texts, tailored to specific emergency scenarios and public opinion dynamics. By constructing a high-quality dataset of 1,715 paired false and clarification information samples from authoritative platforms, we employ a task-separated fine-tuning strategy using LoRA (Low-Rank Adaptation) to optimize model performance. Our evaluation metrics, including text fluency (BLEU), novelty (NOV), and diversity (DIV), demonstrate that fine-tuned models, particularly LLaMA3.1, excel in generating coherent and relevant texts. The results highlight the potential of LLMs in both generating and debunking false information, offering a robust framework for improving public opinion management during emergencies. This research contributes to the growing body of work on false information mitigation and provides practical insights for leveraging LLMs in crisis communication.