TL;DR
This paper introduces TRUE, the first comprehensive dataset for explainable video fact-checking, featuring detailed rationales and evidence. They also present 3MFact, a novel multi-role multimodal framework that leverages this dataset to not only verify video claims but also provide clear, evidence-backed explanations, significantly improving upon existing methods.
What Problem Does This Paper Address?
The proliferation of misinformation, especially in multimodal formats like videos, poses a significant challenge. Existing video fact-checking methods often fall short because they lack detailed explanations and supporting evidence, making their outputs less persuasive and harder to trust. Current datasets are also criticized for their limited annotations, reliance on machine-generated evidence, and poor organization. This makes it difficult to build robust and interpretable fact-checking systems that can effectively combat sophisticated misinformation.
Methods
The core of this work is the 3MFact framework, a multi-role, collaborative system designed for explainable video fact-checking. It operates in several stages:
- Video Descriptor: Converts video content into textual descriptions using Large Multimodal Models (LMMs), capturing both visual and temporal information.
- Claim Verifier: Assesses the sufficiency of available information (claim, video description, background info) to verify a claim. It uses a Chain-of-Thought (CoT) strategy and outputs a confidence score.
- Question Manager: If information is insufficient, this module formulates targeted questions to gather more data, either from the video itself or through online retrieval.
- Information Retriever: This module searches for relevant and credible online evidence based on the questions generated. It scores evidence based on website quality, recency, and relevance, selecting the top pieces.
- Reasoner: The final module synthesizes all gathered information (claim, video description, QA pairs, evidence) to produce the final truthfulness label, a detailed rationale, and supporting evidence.
The framework is designed to iteratively gather and synthesize information, ensuring a robust and well-supported decision.
Datasets
The paper introduces the TRUE (Truthfulness and Rationale with Underlying Evidence) dataset. This is a novel benchmark for explainable video fact-checking, designed to address the limitations of existing datasets. Key features of TRUE include:
- Comprehensive Annotations: Includes veracity labels, detailed rationales (both human-extracted "Expert-Crafted Rationale" - ECR, and LLM-synthesized "LLM-Summary Rationale" - LSR), and supporting evidence for each claim.
- Multimodal Focus: Specifically targets video-based claims.
- Diverse Sources: Claims are sourced from Snopes, covering a range of official and social media platforms.
- Temporal Coverage: Includes claims from 2016 to 2024.
The paper also compares their approach against existing datasets like CHECKED, Mocheg, and FakeSV, highlighting TRUE's superiority in providing explainability and evidence.
Key Insights and Ideas
- The Importance of Explainability: Simply labeling content as "fake" is insufficient. Providing clear, evidence-backed rationales is crucial for persuasive and effective fact-checking, especially against sophisticated AI-generated misinformation.
- Dual Rationales (ECR vs. LSR): The TRUE dataset offers both human-extracted (ECR) and LLM-synthesized (LSR) rationales. LSR generally shows higher accuracy and comprehensiveness, suggesting LLMs can effectively synthesize complex information, while ECR retains more of the original human reasoning nuance.
- Low Information Richness in False Claims: Analysis of claims in the TRUE dataset reveals that false claims tend to be shorter and have lower information richness (entropy), suggesting they often rely on superficial or incomplete information, making them harder to verify without external context.
- Video Context is Key, but Challenging: Incomplete video footage and scene transitions can be indicators of manipulation. However, accurately extracting and utilizing this visual context remains a significant challenge for current LMMs.
- Modular Framework for Robustness: The multi-role architecture of 3MFact allows for specialized tasks (description, verification, retrieval, reasoning), leading to a more robust and adaptable system than monolithic approaches.
Key Results
- Accuracy: The 3MFact framework (specifically the 3MFact(M+M) configuration) achieved state-of-the-art performance, reaching 79.63% accuracy on the TRUE dataset, significantly outperforming traditional methods (e.g., SV-FEND at 62.80%) and standalone LMMs (e.g., VideoLLaVa at 44.11%).
- Explanation Quality: 3MFact demonstrated superior explanation quality across various metrics (e.g., Logical Consistency, Strength of Evidence, Faithfulness) compared to other models, as evaluated by GPT-4.
- Dataset Performance: The LLM-Summary Rationale (LSR) in the TRUE dataset achieved significantly higher accuracy (92.3%) and comprehensiveness (98.4%) compared to Expert-Crafted Rationale (ECR) (70.9% and 33.7% respectively) in quality assessment.
- Ablation Study: The Information Retriever module was found to have the most significant impact on overall accuracy, followed by validation steps and the Question Answering module.
Limitations and Future Work
- Video Descriptor Limitations: While crucial, the Video Descriptor module's ability to extract detailed textual descriptions from videos is still a bottleneck. The paper notes that improvements in LMMs are expected to enhance this component's contribution.
- Human-Crafted Rationales: The ECR rationales in the TRUE dataset require further refinement for completeness and accuracy.
- LLM Challenges: LMMs still face challenges in specific fact-checking subtasks, such as answering complex fact-related questions and generating highly accurate video descriptions.
- Future Work: The authors suggest continued refinement of the dataset, further improvements in LMM capabilities for video understanding, and exploring more advanced reasoning strategies within the 3MFact framework. They also highlight the potential for the dataset to facilitate comparisons between human and AI-generated fact-checking.
Summary updated