Fake News Detection Based on Large Language Models

With the rapid advancement of social networks, the accessibility of information has significantly improved. However, alongside this progress, the proliferation of rumors and misinformation on social media platforms has become increasingly severe, substantially undermining the credibility of online information sources. Traditional methods for detecting fake news are often labor-intensive and time-consuming, with a high risk of misjudgment and erroneous classification. Against this backdrop, leveraging advanced large language models (LLMs) for automated fake news detection has become an area of considerable research interest. The rapid development of artificial intelligence, coupled with the vast and diverse news datasets available on the internet, provides favorable conditions for AI-driven misinformation detection. In this study, we utilize datasets collected from publicly available sources. After undergoing data cleaning and annotation enhancement, these datasets serve as the foundation for model training. For model selection, we employ a range of LLMs. To optimize the performance of smaller-scale models such as ChatGLM-6B, we apply low-rank adaptation (LoRA) for efficient fine-tuning. The dataset is divided into training and validation sets, and model performance is evaluated using F1 Score and Accuracy Score as key metrics.This study aims to compare the effectiveness of various large language models in detecting fake and misleading news, ultimately identifying the most suitable model for enhancing the reliability of information dissemination on social media platforms.