Research map / Video / Video-language understanding
Text-Guided Fine-grained Counterfactual Inference for Short Video Fake News Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TGFC-SVFN, which uses LLM causality-based reasoning prompts to build a text teacher that distills knowledge into video and audio students, with diffusion-based fine-grained counterfactual debiasing, evaluated on the FakeSV dataset.
Problem
Short video fake news detection suffers from modality heterogeneity, modality bias amplification via inter-modal distillation, and teacher model inadequacy when the full text sequence is used as teacher.
Approach
Uses ChatGPT-3.5-generated causality-based reasoning prompts to guide a text teacher, distills to video/audio students with Pearson-correlation inter/intra-class losses, and removes bias via diffusion-based noise-injected counterfactual scenarios and total indirect effect debiasing at each module.
Key results
- On FakeSV time-partitioned split, TGFC-SVFN reaches 91.99% accuracy and 91.50% F1, versus 85.98% accuracy for the best baseline MMCAN.
- On the event-partitioned split, TGFC-SVFN achieves 81.73% accuracy and 81.66% F1, above MultiEMO (79.11% accuracy).
- Adding the components to SV-FEND raises its F1 from 81.69 to 89.05 on the time split (SV-FEND+All).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TGFC-SVFN instead of MMCAN | FakeSV short-video fake news detection · FakeSV time split | F1: 85.32 -> 91.50 | Helped |
| Use TGFC-SVFN instead of MMCAN | FakeSV short-video fake news detection · FakeSV event split | F1: 79.67 -> 81.66 | Helped |
| Add GPT-2 temporal feature analysis (ablation) | TGFC-SVFN · FakeSV time split | F1: 86.70 -> 91.50 | Helped |
| Add Cross-modal Transformer alignment (ablation) | TGFC-SVFN · FakeSV time split | F1: 85.77 -> 91.50 | Helped |
| Add counterfactual inference (ablation) | TGFC-SVFN · FakeSV time split | F1: 85.19 -> 91.50 | Helped |
| Add causality-based reasoning prompt guidance (ablation) | TGFC-SVFN · FakeSV time split | F1: 82.01 -> 91.50 | Helped |
| Add knowledge distillation (ablation) | TGFC-SVFN · FakeSV time split | F1: 88.21 -> 91.50 | Helped |
| Add counterfactual inference (CI) (ablation) | MultiEMO · FakeSV time split | F1: 82.05 -> 82.42 | Helped |
| Add causality-based reasoning prompt guidance (CR) (ablation) | MultiEMO · FakeSV time split | F1: 82.05 -> 83.05 | Helped |
| Add knowledge distillation (KD) (ablation) | MultiEMO · FakeSV time split | F1: 82.05 -> 81.57 | Hurt |
| Add CI, CR, and KD (ablation) | MultiEMO · FakeSV time split | F1: 82.05 -> 84.22 | Helped |
| Add counterfactual inference (CI) (ablation) | SV-FEND · FakeSV time split | F1: 81.69 -> 87.60 | Helped |
Surprising findings
- Causality-based reasoning prompts had the largest ablation impact, larger than counterfactual inference, as removing CR dropped F1 to 82.01.
- ChatGPT-3.5 with direct prompting performed poorly (45.94% accuracy on the time split), below all trained multimodal baselines.
Limitations
- Evaluated on a single benchmark (FakeSV) only.
- Relies on ChatGPT-3.5 generated reasoning prompts, which adds dependence on a proprietary LLM.
- Each module is trained separately, and the paper gives little detail on the counterfactual hyperparameters and diffusion schedule.
Experimental setup
Datasets: FakeSV
Benchmarks: FakeSV
Metrics: F1 macro recall macro precision accuracy
Baselines: FANVM CAFE MultiEMO SV-FEND SV-FEND-SNEED MMCAN ChatGPT-3.5-direct ChatGPT-4-direct ChatGPT-3.5-CR
Builds on: FakeSV Hinton knowledge distillation Cross-modal Transformer GPT-2 denoising diffusion probabilistic models
Closest papers by meaning
- Fact-R1: Towards Explainable Video Misinformation Detection with Deep Reasoning (NeurIPS 2025 · similarity 0.88)
- DAMMFND: Domain-Aware Multimodal Multi-view Fake News Detection (AAAI 2025 · similarity 0.87)
- Pioneering Explainable Video Fact-Checking with a New Dataset and Multi-role Multimodal Model Approach (AAAI 2025 · similarity 0.86)
- Unveiling Implicit Deceptive Patterns in Multi-Modal Fake News via Neuro-Symbolic Reasoning (AAAI 2024 · similarity 0.86)
- Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection (AAAI 2024 · similarity 0.86)
- Reinforced Adaptive Knowledge Learning for Multimodal Fake News Detection (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.