Beyond Words: Modeling Inter-Word Relationships with Edge Graph Neural Networks for Fake News Detection

ACL ARR 2026 January Submission7246 Authors

Keywords: Fake news, Transformer, Edge Graph Neural Network, Graph

Abstract: Fake news detection remains challenging due to the subtle linguistic cues that differentiate fabricated content from factual reporting. Fake news exhibits distinctive patterns in how words relate to each other, such as unusual semantic associations between entities, inconsistent relationship chains, and anomalous co-occurrence patterns that differ from those found in authentic news. However, existing methods typically treat text as sequences of tokens rather than explicitly modeling these inter-word relationships.In this paper, we emphasize on identifying critical signals for fake news detection that are not adequately captured by current approaches. Essentially, our approach explicitly models word relationships through learnable edge embeddings. We present WR-EGNN (Word Relationship–based Edge Graph Neural Network), a framework for fake news detection that explicitly models atypical inter-word relationships by combining transformer-derived contextual representations with graph-based structural modeling.In addition, our approach emphasizes (1) interpretability, as the model explicitly learns the relational patterns that distinguish fake news from real news, and (2) robustness, since structural features are inherently less susceptible to adversarial style manipulations. Furthermore, the results demonstrate that WR-EGNN significantly outperforms three transformer-only models as well as four baseline fake news detection systems.\footnote{Exceutable code of our model is available at: {https://anonymous.4open.science/r/WRGNN-7604}}

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: Misinformation detection, Fake news detection

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Reproduction study

Languages Studied: English

Submission Number: 7246