TL;DR
This paper introduces an "annotation-free" automatic fact-checking (AFC) framework designed to assist journalists. It goes beyond simply classifying claims by providing detailed, human-readable explanations for its verdicts, aiming to build trust and integrate seamlessly into journalistic workflows.
What Problem Does This Paper Address?
The sheer volume of online information and the rapid spread of misinformation make manual fact-checking an increasingly daunting task for journalists. While automated fact-checking (AFC) systems offer a potential solution, existing approaches often fall short due to a lack of explainability, reliance on human-annotated datasets (which are costly and can become outdated), and a disconnect from the practical needs of journalists. This disconnect hinders adoption, as journalists are hesitant to cede control or trust systems they don't understand. The paper aims to bridge this gap by developing an AFC system that is not only accurate but also transparent and explainable, empowering journalists rather than replacing them.
Methods
The core of the proposed framework is an "annotation-free" pipeline that leverages Large Language Models (LLMs) for several key tasks:
- Claim Decomposition: Instead of relying on pre-annotated decompositions, the system uses an LLM (GPT-4o) to break down a claim into a series of yes/no questions. This is done by prompting the LLM to consider the claim as both true and false, generating questions and justifications for each scenario.
- Evidence Search: The system uses the Bing search engine API to find relevant web pages for each decomposed question, filtering out potentially biased or unreliable domains.
- Evidence Summarization: For each piece of evidence, the system generates a summary using sentence embeddings and cosine similarity.
- Explanation Generation: An LLM then merges these evidence summaries into a single, coherent explanation for each decomposed question.
- Verdict Generation: Another LLM answers each decomposed question with a "Yes," "No," or "Unverified" based on the generated explanations. These answers form a numeric dataset.
- Verdict Prediction: A numeric classifier (trained on the generated numeric dataset) predicts the overall verdict for the claim. The authors propose a "Soft Macro-F1" metric for evaluating these multi-label classifiers, accounting for the inherent ambiguity in human fact-checking labels.
- Article Generation: Finally, the system uses an LLM to generate a full explanation article and a summary of the verdict.
A key innovation is the annotation-free approach, which eliminates the need for costly human annotation for training data, allowing for the use of more current claims.
Datasets
- Training Data: The system generates its own training datasets using LLMs, inspired by existing claim decomposition work. For the verdict classifier, they generated a dataset of 539 samples from the CLAIMDECOMP dataset and later expanded it with claims from the Washington Post Fact Checker and SNOPES datasets, creating a generic test dataset of 414 claims.
- Evaluation Data: A political claims test dataset was auto-generated using claims from Chen et al.'s test dataset. For data leakage verification, they created a dataset of 37 false claims published after the LLM's knowledge cutoff date.
Key Insights and Ideas
- Annotation-free is feasible and beneficial: The paper demonstrates that it's possible to build a robust AFC system without relying on human-annotated datasets, which is a significant barrier in current research. This allows for more up-to-date and scalable fact-checking.
- Journalists value explainability and control: The user study highlights that journalists are more likely to adopt AFC systems if they are transparent and allow them to retain authority over the fact-checking process. The framework's focus on generating human-readable explanations directly addresses this.
- "Soft Accuracy" is a more appropriate metric: The authors argue that traditional accuracy metrics are insufficient for multi-label fact-checking due to the inherent ambiguity in human judgments. Their proposed "Soft Macro-F1" metric better captures this nuance.
- Generalization is key: The framework shows strong generalization across different claim subjects and even different fact-checking organizations without retraining, a critical feature for real-world deployment.
- Human augmentation, not replacement: The findings strongly suggest that AFC systems are best suited to augment journalists, providing them with tools and insights, rather than fully automating the process.
Key Results
- Classification Performance: The proposed framework's verdict prediction module achieved a Soft Accuracy of 0.78/0.77 and Soft Macro-F1 of 0.70/0.65, outperforming the state-of-the-art in multi-label classification by 12% on the CLAIMDECOMP dataset.
- Generalization: The system demonstrated robust generalization across different fact-checkers and subjects without retraining.
- User Study Validation (N=103 professional journalists):
- Explainability: 92.1% of participants rated the process as adequately or highly explainable.
- Transparency: 90.3% of journalists rated the process as exceptionally, highly, or adequately transparent.
- Trustworthiness: 85.9% of participants found the framework to have similar or superior credibility to existing fact-checking processes.
- Artifact Quality: Participants rated the generated summaries and explanations highly for coverage, support of verdicts, and factual alignment with human-generated summaries.
Limitations and Future Work
- Data Leakage Verification: While efforts were made to prevent LLM hallucinations and the use of future information, the verification dataset for this was limited to "false" claims. Future work will expand this to include other labels.
- Source Transparency and Bias: Journalists highlighted the need for clearer criteria for source selection and bias evaluation. Future iterations may integrate tools like NewsGuard to address this.
- Language Support: The current framework is limited to English. Expanding to other languages is a significant future direction.
- Processing Time: The current processing time per claim is around 15-16 minutes, which is reasonable but could be improved with more optimized hardware.
- Claim Subcategories: The analysis revealed that "Not Verifiable" and "Imagery" claims are particularly challenging and might benefit from pre-filtering or multimodal AFCs. "Politics" claims also showed vulnerability, suggesting a need for specialized models.
- Dataset Size and Balance: The auto-generated policy classifier training dataset was small and unbalanced, impacting its performance. Future work aims to improve this.
- Personalization: Integrating the framework with personalized explanation techniques could further enhance its impact.
Summary updated