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

Weilin Zhou, Zonghao Ying, Junjie Mu, Shengwei Tian, Quanchen Zou, Deyue Zhang, Dongdong Yang, Xiangzheng ZhangarXiv 2025 · DOI · arXiv

Introduction

Multimodal fake news—content combining text and imagery—presents a unique challenge for automated detection systems. Often, the "smoking gun" of a fabricated story is a subtle mismatch between what is written and what is shown. For instance, a text post might describe a violent riot while the accompanying image depicts a peaceful gathering. However, most current AI models are designed to find consistency between modalities, inadvertently smoothing over the very contradictions that signal deception.

The Dynamic Conflict-Consensus Framework (DCCF) addresses this by explicitly seeking out and amplifying cross-modal inconsistencies. By disentangling objective facts from subjective sentiment and employing a physics-inspired "tension field," DCCF achieves a significant performance boost over existing state-of-the-art methods.

The Challenge: The Trap of Over-Smoothing

Traditional multimodal fusion techniques typically aim to align features from different modalities into a shared representation space. While this works well for tasks like image captioning, it is counterproductive for fake news detection. By treating discrepancies as "noise" to be filtered out, these models suffer from "over-smoothing," effectively hiding the evidence of fabrication. Furthermore, many models fail to distinguish between objective facts (entities, locations) and subjective sentiment (emotional tone), making it difficult to pinpoint the exact nature of the cross-modal conflict.

The DCCF Framework

DCCF introduces an "inconsistency-seeking" paradigm structured across three distinct stages.

The three-stage architecture of the Dynamic Conflict-Consensus Framework (DCCF), illustrating feature evolution and metric extraction.

1. Fact-Sentiment Feature Extraction

To prevent the conflation of different types of information, DCCF projects inputs into two separate spaces:

  • Fact Space: Guided by a pretrained YOLO object detector, this space focuses on grounded, physical entities present in the news.
  • Sentiment Space: Guided by the SenticNet lexicon, this space captures the emotional "vibe" and polarity of the content.

2. Dynamic Feature Evolution (DARFU)

Inspired by physical field theory, the authors developed the Dynamic Feature Evolution Unit (DARFU). This component treats features as particles within a "tension field." Through an iterative process, semantically similar features are clustered together, while inconsistent features are polarized or "repelled." This active amplification of contrast ensures that conflicting signals are preserved and highlighted rather than averaged out.

3. Conflict-Consensus Metric Extraction

In the final stage, the model identifies the "Maximally Informative Conflict"—the pair of features exhibiting the highest tension. This local conflict is then standardized against the "Global Consensus" (the average tone of the entire post). By judging discrepancies relative to the specific context of the post, the model can more accurately distinguish between intentional deception and natural variance.

Experimental Results

The researchers evaluated DCCF on three major benchmarks: Weibo, Weibo-21, and GossipCop. The framework consistently outperformed 11 state-of-the-art baselines, including Large Language Model (LLM)-based systems.

  • Accuracy Improvements: DCCF achieved an average accuracy gain of 3.52% across all datasets.
  • State-of-the-Art Performance: On the Weibo dataset, it reached 95.1% accuracy, surpassing the strong LLM-based "LIFE" model.
  • Robustness: On the imbalanced GossipCop dataset (80% real news), DCCF maintained a high F1-Fake score of 0.723, demonstrating its ability to avoid the "crying wolf" problem common in other detectors.

Ablation studies confirmed that both the "tension weighting" mechanism and the fact/sentiment disentanglement are critical to the model's success.

Conclusion

The core insight of the DCCF framework is that for fake news detection, conflict is the signal, not the noise. By moving away from consistency-based fusion and toward a dynamic, physics-inspired model of feature tension, DCCF provides a more robust and interpretable way to catch multimodal lies. Future work will look into deeper integration with LLMs to provide even more nuanced semantic constraints for the fact and sentiment spaces.

Summary updated

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\%.
BibTeX
@misc{zhou2025disentangling,
  author = {Weilin Zhou and Zonghao Ying and Junjie Mu and Shengwei Tian and Quanchen Zou and Deyue Zhang and Dongdong Yang and Xiangzheng Zhang},
  title = {{Disentangling Fact from Sentiment: A Dynamic Conflict-Consensus Framework for Multimodal Fake News Detection}},
  year = {2025},
  eprint = {2512.20670},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.20670},
}
Figure