Decoupling Generation and Selection for Budget-Constrained Faithful Summarization

2026-08-04

AI Summary

This work addresses the challenges of factual inconsistency, redundancy, and weak length control in multi-document abstractive summarization by proposing a modular generate-then-select framework. Under a sentence-level budget constraint, the approach first employs a pretrained model to generate candidate summaries and decomposes them into sentence units; it then applies a combinatorial optimization selector to jointly maximize relevance, factual faithfulness, and diversity while minimizing redundancy. By decoupling generation from selection, the framework flexibly integrates various selection strategies—such as MMR, ILP, and DPP-inspired objectives—without requiring fine-tuning of the generator. Experimental results demonstrate significant improvements in factual consistency and source fidelity across CNN/DailyMail, Multi-News, FaithBench, and TofuEval benchmarks. Human evaluations further confirm consistent superiority over baselines in coherence, relevance, clarity, and conciseness, with only a marginal drop in ROUGE scores.

Abstract

Abstractive summarization models remain vulnerable to factual inconsistency, redundancy, and weak length control. We propose a modular generation-and-selection framework for sentence-budget-constrained summarization. A pretrained generator produces multiple candidate summaries, which are decomposed into sentence-level candidates. A combinatorial selector then constructs the final summary by balancing relevance, factuality, and redundancy under an explicit budget. The framework supports MMR, ILP, and a DPP-inspired log-determinant objective without retraining the generator. Experiments on CNN/DailyMail, Multi-News, FaithBench, and TofuEval show consistent improvements in factuality and source-grounding metrics, especially for multi-document summarization, at the cost of lower reference-overlap scores. Human evaluation further indicates higher perceived consistency, relevance, clarity, and conciseness, with a small reduction in coherence. These results show that decoupling generation from selection provides a model-agnostic mechanism for improving factual grounding. Code is available at https://anonymous.4open.science/r/bcfs-D05E/.

Problem

• abstractive summarization
• factual inconsistency
• budget-constrained summarization
• redundancy
• length control

Innovation

• decoupled generation-and-selection
• budget-constrained summarization
• factuality
• combinatorial selection
• model-agnostic framework