StrucSum: Graph-Structured Reasoning for Long Document Extractive Summarization with LLMs

TL;DR

Long documents challenge LLM summarization due to limited context and discourse awareness. StrucSum presents a training-free approach by constructing a Text-Attributed Graph (TAG) and applying three prompting strategies—NAP, CAP, and CGM—to inject structural signals into zero-shot extractive summarization. The method leverages Sentence-BERT-based graph construction and defined centrality signals to guide sentence selection, achieving improvements in both summary quality and factual consistency across ArXiv, PubMed, and Multi-News, with analysis showing strategy-specific benefits and limited gains from combining strategies. The work demonstrates that simple graph priors can effectively steer black-box LLMs without fine-tuning, offering a scalable direction for structure-aware long-document summarization and prompting research, with public code to follow up on.

Abstract

Large language models (LLMs) have shown strong performance in zero-shot summarization, but often struggle to model document structure and identify salient information in long texts. In this work, we introduce StrucSum, a training-free prompting framework that enhances LLM reasoning through sentence-level graph structures. StrucSum injects structural signals into prompts via three targeted strategies: Neighbor-Aware Prompting (NAP) for local context, Centrality-Aware Prompting (CAP) for importance estimation, and Centrality-Guided Masking (CGM) for efficient input reduction. Experiments on ArXiv, PubMed, and Multi-News demonstrate that StrucSum consistently improves both summary quality and factual consistency over unsupervised baselines and vanilla prompting. In particular, on ArXiv, it increases FactCC and SummaC by 19.2\% and 8.0\% points, demonstrating stronger alignment between summaries and source content. The ablation study shows that the combination of multiple strategies does not yield clear performance gains; therefore, structure-aware prompting with graph-based information represents a promising and underexplored direction for the advancement of zero-shot extractive summarization with LLMs. Our source code is publicly available.

Figures (9)

• Figure 1: Comparing StrucSum's graph-structured LLM prompting with standard flat-input prompting for long-document extractive summarization.
• Figure 2: Overview of the StrucSum framework: A document is encoded into a Text-Attributed Graph, from which structural signals are extracted via three prompting strategies to guide LLMs in zero-shot extractive summarization.
• Figure 3: Average ROUGE-F1 on PubMed with varying values of $k$ and $\theta$ in the NAP strategy.
• Figure 4: Normalized prompt lengths for vanilla, NAP, CAP, and CGM strategies across ArXiv, PubMed, and Multi-News.
• Figure 5: Spearman correlation coefficients between sentence centrality scores and selection rank under different prompting strategies on GPT-4o, shown separately for ArXiv and PubMed. Asterisks indicate statistical significance ($p < 0.05$).