Disentangling Fact from Sentiment: A Dynamic Conflict-Consensus Framework for Multimodal Fake News Detection

TL;DR

This work addresses multimodal fake news detection by challenging the prevailing consistency-based fusion paradigm and treating cross-modal inconsistencies as primary evidence. It introduces the Dynamic Conflict-Consensus Framework (DCCF), which disentangles inputs into fact and sentiment spaces, uses a tension field network to dynamically amplify conflicts, and standardizes them against global consensus for robust predictions. Across three real-world benchmarks, DCCF achieves state-of-the-art accuracy improvements, with an average gain of $3.52\%$ over strong baselines, and demonstrates robustness to class imbalance and varying data distributions. The approach also offers interpretable reasoning through maximally informative conflicts and tone-reference standardization, though it relies on auxiliary pseudo-labels and could be enhanced by integrating Large Language Models for stronger semantic constraints.

Abstract

Prevalent multimodal fake news detection relies on consistency-based fusion, yet this paradigm fundamentally misinterprets critical cross-modal discrepancies as noise, leading to over-smoothing, which dilutes critical evidence of fabrication. Mainstream consistency-based fusion inherently minimizes feature discrepancies to align modalities, yet this approach fundamentally fails because it inadvertently smoothes out the subtle cross-modal contradictions that serve as the primary evidence of fabrication. To address this, we propose the Dynamic Conflict-Consensus Framework (DCCF), an inconsistency-seeking paradigm designed to amplify rather than suppress contradictions. First, DCCF decouples inputs into independent Fact and Sentiment spaces to distinguish objective mismatches from emotional dissonance. Second, we employ physics-inspired feature dynamics to iteratively polarize these representations, actively extracting maximally informative conflicts. Finally, a conflict-consensus mechanism standardizes these local discrepancies against the global context for robust deliberative judgment.Extensive experiments conducted on three real world datasets demonstrate that DCCF consistently outperforms state-of-the-art baselines, achieving an average accuracy improvement of 3.52\%.

Figures (4)

• Figure 1: Schematic of inconsistency distortion, where $f_1$ and $f_2$ denote a pair of inconsistent features (e.g., conflicting text and image). (a) Graph-based methods blur this specific conflict by averaging neighbor features via edges. (b) Attention-based methods dilute the inconsistency by aggregating global features via weights, resulting in homogenized outputs.
• Figure 2: The DCCF framework: (a) Fact-Sentiment Feature Extraction projects features into Fact ($S^F$) and Sentiment ($S^E$) spaces; (b) Feature Dynamics Evolution refines features through DARFU blocks to compute conflict/consensus; (c) Multi-View Deliberative Judgment fuses metrics for the final decision.
• Figure 3: Analysis of hyperparameter sensitivity. This figure shows the impact of four different hyperparameters on the model's F1-real score across three datasets. (a) Effects of the number of DARFU unit iterations (m) and the Temperature Coefficient ($\tau$). (b) Effects of the fact auxiliary loss coefficient ($\mathcal{L}_F$) and the sentiment auxiliary loss coefficient ($\mathcal{L}_E$).
• Figure 4: T-SNE visualization of test set features. Same color dots indicate the same label.