TLDR: Researchers introduce Politi-Fact-Only (PFO), a new benchmark dataset for fact-checking that removes post-claim analysis to provide more realistic evaluations for Large Language Models (LLMs). They also propose RA V (Recon-Answer-Verify), an agentic framework with Question, Answer, and Label Generator agents that iteratively verify claims. RA V outperforms existing methods and demonstrates greater robustness on the PFO dataset, highlighting the importance of realistic data and iterative reasoning in automated fact-checking.
Automated fact-checking using large language models (LLMs) offers a promising way to combat the rapid spread of misinformation, especially on digital platforms like social media. However, a significant challenge in evaluating these AI systems has been the realism of existing benchmark datasets.
Many current datasets, often derived from fact-checking websites, include what researchers call ‘leakage’ – information added after a claim was made, such as detailed analyses or explicit verdicts from annotators. This post-claim analysis can inadvertently guide AI models, making them appear more accurate than they would be in real-world scenarios where claims need to be verified immediately.
To address this, researchers Satyam Shukla, Himanshu Dutta, and Pushpak Bhattacharyya from the Indian Institute of Technology Bombay have introduced a new benchmark dataset called Politi-Fact-Only (PFO). This dataset comprises 2,982 political claims from politifact.com, where all post-claim analysis and annotator cues have been meticulously removed. This ensures that models are evaluated using only the information that would have been available before the claim’s verification. When LLMs were tested on PFO, they showed an average performance drop of 22% compared to the unfiltered version, highlighting their reliance on these hidden cues.
Based on the identified challenges, the researchers also propose a novel agentic framework called RA V (Recon-Answer-Verify). This system mimics the human fact-checking process by employing three specialized AI agents:
Also Read:
- OpenFActScore: Advancing Open-Source Evaluation for AI Text Factuality
- Combating Misinformation: A Multi-Agent System for Verifying Multimedia Content
The RA V Framework:
- Question Generator (QGagent): This agent iteratively generates sub-questions based on the original claim and the history of previous questions and answers. It aims to break down the claim into verifiable components, asking both true/false and inquiry-based questions.
- Answer Generator (AGagent): This agent takes a generated question and uses the provided evidence to formulate an answer. This step connects the verification process to the factual context.
- Label Generator (LGagent): Once the claim has been sufficiently explored through the question-and-answer process, this agent synthesizes all the information to predict the final veracity label (e.g., true, mostly-true, half-true, mostly-false, false) and provides reasoning for its decision.
The RA V pipeline is designed to be domain-agnostic, meaning it can generalize across different topics and levels of label granularity. It has demonstrated superior performance compared to state-of-the-art approaches on various well-known baselines. For instance, it outperformed RAWFC (a fact-checking dataset) by 25.28% and HOVER (an encyclopedia-based dataset) by significant margins across different complexity levels (1.54% on 2-hop, 4.94% on 3-hop, and 1.78% on 4-hop claims).
Furthermore, RA V proved to be more robust when evaluated on the PFO dataset compared to its unfiltered counterpart, showing a much smaller performance drop of 16.3% in macro-f1, especially with larger LLM backbones like LLaMA-3.1-70B, which saw only a 7.36% drop. The study also emphasized the importance of the reasoning steps within the RA V pipeline, as removing them led to an average performance degradation of 3.11%.
This research marks a significant step towards more transparent and reliable automated fact-checking systems, addressing critical issues of data realism and model interpretability. For more details, you can read the full research paper here.
