Structure‐Guided Extractive Compression for Long‐Document Summarization via Budget Allocation and Redundancy Filtering

ABSTRACT

Long‐document summarization with large language models is limited by context‐window size and inference cost. A training‐free compressor can reduce this cost by selecting source sentences before generation, but a single global ranking leaves the distribution of a fixed token budget across document units implicit. We separate allocation from sentence selection in a two‐stage method that requires no task‐specific training. It first assigns initial quotas to pre‐defined units from their lengths and fixed structural weights. It then ranks sentences by similarity to the document centroid, applies a within‐unit redundancy threshold and globally refills unused capacity. The resulting implementation is greedy and does not solve a linear or integer program. We evaluate one fixed Qwen2.5‐7B‐Instruct generator on GovReport, ArXiv, QMSum and an 84‐case Multi‐LexSum set whose median input length is 221,986 tokens. Against matched global‐ranking controls, the method improves ROUGE‐1 by 0.83 points on ArXiv and 4.04 points on QMSum, whereas GovReport results are mixed. Against the supervised MemSum extractor under a matched 1024‐token evidence budget, it is weaker on ArXiv and mixed on GovReport. On Multi‐LexSum, it exceeds a full‐document hierarchical Qwen baseline by 3.66/1.39/1.29 ROUGE‐1/2/L points overall and reduces end‐to‐end runtime from approximately 7.9 h to 50.6 min, although the 200K+ word subset does not show a conclusive quality advantage. The results support quota allocation as a practical, interpretable control while also showing that it is not consistently better than supervised or simple positional baselines.