AI Papers Reader

Personalized digests of latest AI research

View on GitHub

AI Learns to Break Down Facts for Better Verification

South Bend, IN - In the quest to automate fact-checking, researchers at the University of Notre Dame have developed a new AI framework that dynamically breaks down complex claims into smaller, more manageable pieces, leading to more accurate verification. The study, recently published on arXiv, addresses a critical gap in how current fact-checking systems operate.

The prevailing “Decompose-Then-Verify” paradigm first breaks down a long, complex claim into several subclaims, and then verifies each subclaim independently. However, existing methods often treat these two steps in isolation, leading to suboptimal results. The key insight of the new research is that the way a claim is decomposed significantly impacts the verifier’s ability to assess its truthfulness.

“Think of it like teaching a student,” explains Yining Lu, the lead author of the study. “If you present them with a very dense, complicated concept, they might struggle to understand it. But if you break it down into smaller, more digestible pieces, they’re more likely to grasp the underlying facts.”

The researchers introduced a novel metric called “atomicity” to quantify the information density of a subclaim. A subclaim with high atomicity contains a lot of information packed into a single statement, while a subclaim with low atomicity is more granular and focused.

For example, consider the claim: “In 2015, Owen suffered a serious head injury while surfing in Hawaii, which forced him to take a break from the sport. However, he made a remarkable comeback in 2017 and has continued to compete at the highest level since then.”

A high-atomicity subclaim might be: “Owen made a comeback in 2017.” A low-atomicity subclaim might be: “Owen suffered a serious head injury in 2015.” The study found that different verifiers, even when using the same verification policy (e.g., retrieving evidence from the internet), perform optimally at different levels of atomicity. Some prefer the high-level overview, while others excel at scrutinizing the details.

To address this, the researchers developed a reinforcement learning framework called “dynamic decomposition.” This framework learns a decomposition policy that dynamically adjusts the atomicity of subclaims based on feedback from the verifier. In essence, the AI learns to tailor the decomposition process to the specific strengths and weaknesses of the verification system.

The dynamic decomposition framework works by framing the decomposition process as a series of decisions. At each step, the AI decides whether to further decompose a subclaim or pass it to the verifier. The verifier then provides a “confidence score,” reflecting how certain it is about the subclaim’s veracity. This score serves as a reward signal, guiding the AI to decompose claims in a way that maximizes the verifier’s confidence.

Experimental results showed that dynamic decomposition significantly outperforms existing decomposition policies, improving verification confidence by 0.07 and accuracy by 0.12 on a 0-1 scale. This improvement was consistent across different verifiers, datasets, and atomicity levels of the original claims.

“Our work shows that it’s not just about breaking down claims, but about breaking them down in the right way,” Lu emphasizes. “By learning to align the decomposition process with the verifier’s preferences, we can significantly improve the accuracy and reliability of automated fact-checking systems.”

The researchers hope that this work will pave the way for more sophisticated and adaptable fact-checking systems that can better navigate the complexities of long-form text. The code for dynamic decomposition is available at github.com/yining610/dynamic-decomposition.