Logic-Guided Extraction Improves LLM Fact Accuracy

Large Language Models (LLMs) are increasingly used for semantic data extraction from unstructured text, generating relational facts. However, they often struggle with tasks requiring complex combinatorial reasoning and global consistency, leading to unreliable outputs. Researchers have developed a novel logic-guided data extraction framework that integrates LLM-based extraction with Answer Set Programming (ASP). In this hybrid approach, the LLM generates candidate facts, while ASP takes on the crucial roles of validation, inference, consistency checking, and control. Unlike traditional pipelines that query LLMs for all predicates independently, this framework interleaves LLM calls with ASP derivation. ASP identifies logically admissible predicates, guiding subsequent extraction queries, inferring implied facts without extra LLM calls, and detecting inconsistencies early. This method reduces LLM calls and significantly improves extraction quality by reducing spurious outputs.