New hybrid fact-checker combines knowledge graphs, LLMs, and web search for 93% accuracy

A new paper by Shaghayegh Kolli and colleagues (accepted at the 9th wiNLP workshop at EMNLP) presents a hybrid fact-checking approach that combines knowledge graphs, large language models, and real-time search agents. The system operates in three autonomous steps: first, a rapid one-hop lookup in DBpedia; second, an LLM-based classification guided by a task-specific labeling prompt that produces outputs with internal rule-based logic; and third, a web search agent invoked only when the knowledge graph coverage is insufficient.

On the FEVER benchmark, the pipeline achieves an F1 score of 0.93 on the Supported/Refuted split without any task-specific fine-tuning. For claims labeled 'Not Enough Information' (NEI), the authors conducted a targeted reannotation study showing their approach frequently uncovers valid evidence, verified by both expert annotators and LLM reviewers. The modular, open-source pipeline is designed for generalization across datasets and includes fallback strategies for robust fact verification.