TL;DR
This paper introduces TrendFact, a novel benchmark for evaluating Automatic Fact-Checking (AFC) systems, specifically focusing on their ability to perceive and prioritize "hotspot" claims with high social impact. They also propose FactISR, a framework that enhances AFC systems by dynamically allocating reasoning resources based on claim influence, leading to improved accuracy and efficiency.
What Problem Does This Paper Address?
The proliferation of misinformation poses significant social risks. While Large Language Models (LLMs) and Reasoning Language Models (RLMs) are increasingly used for Automatic Fact-Checking (AFC), they face a critical challenge: risk asymmetry. This means high-influence claims, if fact-checked incorrectly, can cause disproportionately more harm than low-influence claims. Existing AFC systems and benchmarks often fail to account for this social impact, lacking the necessary data and evaluation frameworks to assess a system's ability to dynamically allocate reasoning resources based on a claim's potential harm. This "Hotspot Perception Ability" (HPA) is crucial for effective and responsible fact-checking in real-world, resource-constrained environments.
Methods
The paper introduces two main contributions:
- TrendFact Benchmark: This is a new dataset designed to evaluate HPA. It includes 7,643 curated samples from trending platforms and professional datasets, with an evidence library of 366,634 entries. Crucially, it incorporates social metadata (views, discussions, engagements, posts) and a GPT-evaluated risk score to quantify claim influence.
- FactISR Framework: This is an influence-aware enhancement framework for RLMs-served AFC systems. It features two key components:
- Dynamic Evidence Augmentation (DEA): Unlike traditional Retrieval-Augmented Generation (RAG) that loads all evidence at once, DEA dynamically retrieves and adds relevant evidence during the reasoning process, ensuring more refined information.
- Iterative Self-Reflection (ISR): This module shifts the reflection trigger from internal model uncertainty to external social risk. It proactively initiates deeper reasoning for high-influence claims, ensuring rigorous verification where it's most needed, thus optimizing resource allocation.
Datasets
- TrendFact: A newly constructed benchmark dataset comprising 7,643 samples. It includes claims from trending platforms (Weibo, DouYin, Baidu) and existing fact-checking datasets (like CHEF) to capture real-world dissemination dynamics and complex verification logic.
- CHEF Dataset: Used as a supplement to TrendFact, providing additional claims from fact-checking websites.
- Existing Fact-Checking Datasets: Mentioned as sources for claims, including Hover, FEVEROUS, CLAIMDECOMP, and QUANTEMP, though TrendFact aims to go beyond their limitations.
Key Insights and Ideas
- Risk Asymmetry is Real: High-influence claims pose a significantly greater risk if misverified, demanding specialized attention.
- HPA is Crucial: The ability of an AFC system to dynamically allocate reasoning resources based on social impact is essential for real-world effectiveness.
- Existing Benchmarks Fall Short: Current datasets lack the social metadata and evaluation frameworks needed to assess HPA, hindering progress in AFC.
- LLMs/RLMs Struggle with HPA: Even advanced LLMs and RLMs show limited performance on TrendFact, highlighting the need for specialized frameworks.
- FactISR's Influence-Awareness: Shifting reflection triggers from internal model uncertainty to external social risk is a key innovation for efficient and effective fact-checking.
- DEA Improves Reasoning: Dynamically augmenting evidence during the reasoning process leads to more refined information and better fact-checking performance.
Key Results
- TrendFact Performance: Existing AFC systems exhibit limited performance on the TrendFact benchmark, underscoring its challenging nature.
- FactISR Enhancements:
- FactISR significantly improves HPA and computational efficiency for RLMs-served AFC systems.
- It enhances verification accuracy by 5-10 percentage points for some models.
- DEA module reduces reasoning time by up to 30-50% while maintaining accuracy.
- FactISR improves the HCPI (Hotspot Claim Perception Index) score by up to nearly 4% for RLMs.
- Explanation Quality: While RLMs may produce less fluent explanations than LLMs on surface-level metrics, FactISR improves their consistency and expressiveness.
Limitations and Future Work
- Real-time Relevance: Converting trending statements into complex reasoning claims requires significant human effort.
- Labor Costs: Manually gathering and summarizing evidence and explanations is labor-intensive. The authors suggest exploring more powerful LLMs for summarization.
- Single Language: The current benchmark is limited to Chinese. Future work aims to extend it to a multilingual setting.
Summary updated