Enhancing fake news detection in sport: a deep learning approach to reliable sport information dissemination

The rapid growth of digital media platforms has fundamentally transformed how sport-related information is produced and consumed, while simultaneously increasing the risk of misinformation dissemination. Fake news about sports often blends factual reporting with speculation, emotional language, and rapidly evolving narratives, which makes reliable detection particularly challenging. This article proposes a multi-layer embedding fusion framework for detecting fake news in sports by integrating complementary representations from multiple transformer-based language models, including Robustly Optimized Bidirectional Encoder Representations from Transformers (RoBERTa), Decoding-Enhanced Bidirectional Encoder Representations from Transformers version 3 (DeBERTa-v3), and Large Language Model Meta AI 2 (LLaMA2). The proposed approach extracts embeddings from multiple intermediate layers of each model to capture linguistic information at different abstraction levels, combines the encoder-based representations, and further fuses them with embeddings from Large Language Model Meta AI 2 using a scaled dot-product attention mechanism. A lightweight neural classifier trained with a focal loss objective is used to mitigate the effects of class imbalance. Experiments conducted on the Fake News Football dataset show that the proposed method consistently outperforms strong recurrent and transformer-based baseline models, achieving 94.18% on the harmonic mean of precision and recall (F1-score) and 95.58% accuracy. These results indicate that the proposed framework is a robust solution for supporting the reliable dissemination of sport information in dynamic and complex media environments.