FactGuard: Event-Centric and Commonsense-Guided Fake News Detection

TL;DR

FactGuard introduces an event-centric fake news detector that leverages LLMs to extract topic-content and commonsense rationale, reducing reliance on stylistic cues. A dual cross-attention interactor and a Rationale Usability Evaluator dynamically balance LLM-derived advice with ground-truth signals, enabling robust reasoning and improved accuracy. The framework is trained with a multi-term loss and is distilled into FactGuard-D to support cold-start and resource-constrained deployment, preserving much of the original performance. Experiments on GossipCop and Weibo21 demonstrate superior robustness and accuracy over baselines, with the distilled variant offering practical efficiency without significant accuracy loss.

Abstract

Fake news detection methods based on writing style have achieved remarkable progress. However, as adversaries increasingly imitate the style of authentic news, the effectiveness of such approaches is gradually diminishing. Recent research has explored incorporating large language models (LLMs) to enhance fake news detection. Yet, despite their transformative potential, LLMs remain an untapped goldmine for fake news detection, with their real-world adoption hampered by shallow functionality exploration, ambiguous usability, and prohibitive inference costs. In this paper, we propose a novel fake news detection framework, dubbed FactGuard, that leverages LLMs to extract event-centric content, thereby reducing the impact of writing style on detection performance. Furthermore, our approach introduces a dynamic usability mechanism that identifies contradictions and ambiguous cases in factual reasoning, adaptively incorporating LLM advice to improve decision reliability. To ensure efficiency and practical deployment, we employ knowledge distillation to derive FactGuard-D, enabling the framework to operate effectively in cold-start and resource-constrained scenarios. Comprehensive experiments on two benchmark datasets demonstrate that our approach consistently outperforms existing methods in both robustness and accuracy, effectively addressing the challenges of style sensitivity and LLM usability in fake news detection.

Figures (8)

• Figure 1: Overview of FactGuard and FactGuard -D. FactGuard main consists of two modules: (1) Feature Extraction, which identifies topic content and enables commonsense reasoning for each news article using an LLM. The resulting features and the original text are encoded for downstream processing. (2) Feature Concatenation, which adaptively integrates LLM-derived features with news content via a cross-attention mechanism and the Rationale Usability Evaluator, followed by MLP-based classification. After training, knowledge distillation yields a lightweight FactGuard-D without LLMs' advice.
• Figure 2: Data processing workflow. It consists of two parts: (1) extracting topics and content to mitigate the influence of text style; and (2) performing commonsense reasoning to identify contradictions in the news and generate LLM judgments.
• Figure 3: Similarity and Shannon entropy analysis on Weibo21 and GossipCop datasets.
• Figure 4: Sensitivity analysis of FactGuard model in Weibo21 and GossipCop datasets across four evaluation metrics: $\text{macF1}$, $\text{Accuracy}$, $\text{F1}_\text{real}$, and $\text{F1}_\text{fake}$.
• Figure 5: Sensitivity analysis of FactGuard -D model in Weibo21 and GossipCop datasets across four evaluation metrics: $\text{macF1}$, $\text{Acc.}$, $\text{F1}_\text{real}$, and $\text{F1}_\text{fake}$.