EZ-check: Explainable zero-shot knowledge extraction and fact checking using knowledge graphs

The unchecked proliferation of online information underscores the need for transparent and efficient fact-checking systems. While some methods leverage structured representations like knowledge graphs, significant gaps persist in efficiently and accurately extracting relevant information. To address this, we introduce EZ-Check, a novel framework that combines explainability with a fast, reliable, and effective querying mechanism for knowledge graphs, leveraging resources such as ConceptNet. At its core, EZ-Check integrates structured knowledge graph relationships with advanced language models to assess textual authenticity and extract relevant information precisely. EZ-Check achieves more than a 10% relative F1 improvement on FEVER and UKP Snopes compared to representative claim-only and evidence-based baselines, including BERT-as-KB, GPT-3.5 Turbo, KGAT, and TARSA, while reducing inference time by approximately 50%. On fact-checking datasets such as FEVER and UKP Snopes, it achieved state-of-the-art results, surpassing baselines without relying on external evidence sentences. Moreover, EZ-Check delivers clear, interpretable justifications, setting a new standard for both performance and explainability in fact verification based on our testing against baselines.