Research map / Video / Video-language understanding
Pioneering Explainable Video Fact-Checking with a New Dataset and Multi-role Multimodal Model Approach
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.
Introduces the TRUE dataset of Snopes-sourced video claims with veracity labels, expert and LLM rationales, and evidence, and a zero-shot multi-role LMM/LLM framework (3MFact) that iteratively retrieves online evidence to output label, rationale, and evidence.
Problem
Existing video fact-checking datasets lack detailed evidence and explanations, and fact-checking methods lack interpretability and reliable multimodal reasoning for video claims.
Approach
A multi-role framework where a video descriptor, claim verifier, question manager, information retriever and reasoner iteratively gather and score online evidence via question-guided loops until enough information is available to decide veracity with rationale; paired with a dataset annotating both expert-crafted and LLM-summarized rationales linked to evidence.
Key results
- 3MFact(M+M) reaches 79.63% accuracy and 73.85% F1 on the 2-class TRUE test set (433 claims), versus 62.73% for MiniCPM-V 2.6 and 44.11% for VideoLLaVa.
- Removing the Information Retriever drops accuracy from 79.63% to 69.98%, the largest ablation drop.
- Human annotators rated LSR at 92.3% accuracy and 98.4% comprehensiveness versus 70.9% and 33.7% for ECR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the 3MFact (M+M) framework instead of MiniCPM-V 2.6 alone | TRUE video fact-checking · TRUE test set | accuracy: 62.73% -> 79.63% | Helped |
| Use 3MFact (M+M) instead of MiniCPM-V 2.6 alone | TRUE video fact-checking · TRUE test set | ROUGE-L: 1.43 -> 2.37 | Helped |
| Add the Information Retriever module (ablation) | 3MFact (M+M) · TRUE test set | accuracy: 69.98% -> 79.63% | Helped |
| Add Question Answering with VideoLMM (ablation) | 3MFact (M+M) · TRUE test set | accuracy: 76.18% -> 79.63% | Helped |
| Add the Video Descriptor module (ablation) | 3MFact (M+M) · TRUE test set | accuracy: 79.42% -> 79.63% | No clear effect |
| Add validation of QA&Evidence pairs and retrieved evidence (ablation) | 3MFact (M+M) · TRUE test set | accuracy: 72.09% -> 79.63% | Helped |
| Add Question Answering with VideoLMM (ablation) | 3MFact (M+M) · TRUE test set | Strength of Evidence: 4.24 -> 3.96 | Mixed |
| Add Question Answering with VideoLMM (ablation) | 3MFact (M+M) · TRUE test set | Comprehensiveness: 4.07 -> 3.94 | Mixed |
| Add the Information Retriever module (ablation) | 3MFact (M+M) · TRUE test set | Comprehensiveness: 3.28 -> 3.94 | Helped |
Surprising findings
- Removing the VideoLMM from the Question Manager unexpectedly improved some evidence-related explanation metrics.
- Video Descriptor contributes little to accuracy despite VideoLMM advances.
Limitations
- Video detail analysis errors (e.g., miscaptioning) cause misjudgments of accurate claims.
- Human-crafted rationales (ECR) have accuracy and completeness issues.
- Video fact-checking evaluated on a single dataset of 433 test claims drawn from Snopes.
Experimental setup
Datasets: TRUE
Benchmarks: TRUE
Metrics: Accuracy Recall Precision F1 BLEU ROUGE-L G-Eval scores (content credibility)
Baselines: SV-FEND VideoLLaVA MiniCPM-V 2.6 3MFact(C+V)
Builds on: Question-guided Multi-hop fact-checking (QACheck) G-Eval Mocheg
Closest papers by meaning
- DEFAME: Dynamic Evidence-based FAct-checking with Multimodal Experts (ICML 2025 · similarity 0.87)
- Fact-R1: Towards Explainable Video Misinformation Detection with Deep Reasoning (NeurIPS 2025 · similarity 0.87)
- Text-Guided Fine-grained Counterfactual Inference for Short Video Fake News Detection (AAAI 2025 · similarity 0.86)
- MAFT: Multimodal Automated Fact-Checking via Textualization (AAAI 2025 · similarity 0.85)
- AVerImaTeC: A Dataset for Automatic Verification of Image-Text Claims with Evidence from the Web (NeurIPS 2025 · similarity 0.84)
- ViFactCheck: A New Benchmark Dataset and Methods for Multi-Domain News Fact-Checking In Vietnamese (AAAI 2025 · similarity 0.84)
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.