Incongruent News Detection Using LLM
Abstract
Misinformation and clickbait have proliferated in the modern digital age, frequently taking the shape of contradictory news stories with dramatic headlines that don't fairly convey the content. In order to reduce such deceptive methods, this work suggests an automated approach for title-body consistency detection utilizing Large Language Models (LLMs). Our method assesses whether the title of a news story accurately represents its content by combining abstractive summarization with semantic similarity analysis. In particular, we employ SentenceBERT (SBERT) to calculate semantic similarity scores between titles and summaries using a refined BART model to produce succinct summaries of article bodies. Furthermore, by differentiating between consistent and inconsistent news, a DistilBERT-based classifier is integrated to enhance decision robustness. The suggested method successfully identifies false or inflated headlines and achieves high accuracy on real-world datasets. This work's primary contribution is the unification of categorization, embedding-based similarity, and summarization into a single pipeline that offers a scalable and comprehensible method for automated media integrity verification. This study contributes to advancing AI-assisted journalism by providing a practical, data-driven tool to promote trustworthy and transparent news dissemination.