A Logic-Constrained Global-Local Collaborative Method for Long-Text Knowledge Extraction

In long-text knowledge extraction, long-span semantic dependencies and dispersed entity references often cause conventional sliding-window segmentation to break cross-paragraph semantic links, thereby affecting the completeness and consistency of extraction results. To address this problem, this paper proposes GlocalKE (Global-Local Collaborative Knowledge Extraction), a logic-constrained global-local collaborative method for long-text knowledge extraction. The method first uses coreference resolution over the full text to integrate dispersed entity expressions and form a global semantic index that can be transferred across windows. It then introduces fuzzy evidence matching to verify the consistency between generated results and supporting spans in the original text, reducing the risk of hallucination in generative models. Finally, entity alignment and bidirectional soft logic reasoning are combined to check and revise candidate triples under ontology constraints. Experiments on the DuIE-Doc and CS-Cn-50 datasets show that the F1 values of the proposed method reach 65.92% and 42.55%, respectively, and outperform those of the compared baseline methods, improving the accuracy and logical consistency of long-text knowledge extraction.