Triple-R: Automatic Reasoning for Fact Verification Using Language Models

Abstract

The rise of online social media platforms has made them a popular source of news. However, they are also prone to misinformation and fake news. To combat this, fact-checking is essential to verify the accuracy of claims made on these platforms. However, the existing methods in this field often lack the use of external sources and human-understandable explanations for system decisions. In this paper, we introduce a framework called Triple-R (Retriever, Ranker, Reasoner) that addresses these challenges. The framework uses the Web as an external knowledge source to retrieve relevant evidence for claims and includes a method to generate reasons based on the retrieved evidence for datasets lacking explanations. We then use this modified dataset to fine-tune a causal language model that generates natural language explanations and labels for pairs of retrieved evidence and claims. Our approach aims to improve the transparency and interpretability of fact-checking systems by providing understandable explanations for decision-making processes. We evaluated our method on a popular dataset and demonstrated its performance through an ablation study. The modified dataset is available on the Canadian Institute for Cybersecurity datasets webpage at https://www.unb.ca/cic/datasets/index.html.