Neural-Symbolic LogicReasoning for FakeNewsVideo Explanation
Lizhi Chen , Zhong Qian , Peifeng Li , Qiaoming Zhu
The task of fake news video explanation generation requires models to logically expose misleading elements in news videos, producing text that combines factual accuracy with argumentative coherence. Existing methods face dual bottlenecks: shallow semantic mining struggles to identify deep multimodal contradictions, while dense semantic unit extraction lacks a unified logical framework, leading to redundant, disorganized, and unconvincing generated content. To address these challenges, this paper proposes a neural-symbolic logical reasoning explanation framework (LogicExP), which reconstructs the explanation generation process as a symbolic reasoning paradigm from multimodal evidence to logical conclusions. This framework first extracts fine-grained semantic units from videos through a multimodal encoder; then designs a meta-predicate system covering typical forgery techniques and combines it with a graph attention network to dynamically instantiate logical constants; subsequently introduces a differentiable logic rule generator to adaptively learn interpretable detection rules as the backbone of structured arguments; and finally guides the decoder to generate explanations that are well-structured and logically rigorous. Experiments on benchmark datasets show that LogicExP significantly outperforms existing methods in automatic evaluation and provides an effective paradigm for building a highly reliable explanation system for fake news videos.
Fake news video explanation / neuro-symbolic / multimodal / logical reasoning
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(4522KB)
