Hi-SCCG: Hierarchical Structure-aware Chunking and Context Generation for document-level information extraction

Abstract

Large Language Model (LLM)-based generative Information Extraction (IE) over long documents is constrained by finite context windows, and the necessary text-chunking step disrupts semantic continuity, causing subject drift, unresolved coreference, and loss of the global constraints required for fine-grained knowledge graph construction. Existing context-augmentation methods either provide only limited local context or incur substantial computational cost while introducing irrelevant information, largely because they do not fully exploit document hierarchy. We propose Hi-SCCG (Hierarchical Structure-aware Chunking and Context Generation), a framework that models document structure by reconstructing the logical skeleton as a semantic tree through LLM-based heading correction and by generating cached multi-granularity summaries with a bottom-up recursive strategy. During inference, Hi-SCCG injects a lightweight path-aware context formed by concatenating ancestor summaries from the leaf chunk to the document root, thereby helping anchor local text to its document-level and section-level subjects. We further introduce LDE-500, a focused benchmark for document-level IE over Chinese hierarchical documents, and use Subject Restoration Rate (SRR) as a diagnostic subset-recall measure for triples that require subject restoration. On LDE-500, Hi-SCCG achieves an F1-score of 0.82 and a diagnostic SRR of 82.1%, with a 9.4 percentage-point absolute SRR improvement over the strongest evaluated baseline. In a representative token-accounting case involving a 45K-token document, after including one-time heading correction, recursive summary-generation calls, extraction input tokens, and extraction output tokens, Hi-SCCG reduces total KG-construction token consumption by 75.4–77.7% relative to SLIDE. These results suggest that structure-aware path injection is a promising strategy for fine-grained IE in structure-rich long documents when SRR is interpreted together with standard triple-level metrics and subset-level precision. A supplementary fully human-annotated Legal-OOD-60 evaluation shows a similar trend in this compact legal-domain sample, where Hi-SCCG achieves an F1-score of 0.77 and an SRR of 75.0%, although broader cross-domain validation remains necessary.