KAG-BERT: A KNOWLEDGE-AWARE GRAPH-BASED BERT FRAMEWORK FOR FAKE NEWS

Abstract

The rapid generation and dissemination of fake news pose serious challenges in the digital era, leading to misinformation, public deception, and erosion of trust in legitimate news sources. In the present paper, we propose a Knowledge-Aware Graph based BERT framework (KAG-BERT) for auto- matic fake news detection that uses contextual textual embeddings from BERT combined with relational reasoning through GNN. The model captures semantic information about news titles and structural relationships among news articles in the knowledge graph, hence providing robustness in the detection of misinformation. We evaluate our framework on the GossipCop dataset, showing an accuracy of 80.31%, with precision at 68.79%, recall at 33.46%, and an F1-score of 45.02%. These results confirm that a transformer-based embedding model combined with graph-based relational learning significantly outperforms the identification of fake news compared to text-only models. The proposed approach provides a scalable and interpretable solution to mitigate the spread of misinformation in real-world settings.