Decoupling Generation and Selection for Budget-Constrained Faithful Summarization
A modular generation-and-selection framework for sentence-budget-constrained summarization is proposed. 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 indicates higher perceived consistency, relevance, clarity, and conciseness, with a small reduction in coherence.
A new modular framework decouples summary generation from sentence selection to improve factual consistency under a sentence budget. A generator creates multiple candidate summaries, which are broken into sentences; a selector then picks the best combination using objectives like MMR, ILP, or a DPP-inspired log-determinant. Tested on CNN/DailyMail, Multi-News, FaithBench, and TofuEval, the method boosts factuality and source-grounding, especially for multi-document tasks, though reference-overlap scores drop. Human judges rate the summaries as more consistent, relevant, clear, and concise, with slightly lower coherence.
The framework treats summarization as a combinatorial optimization problem over sentence candidates, enabling explicit trade-offs between relevance, factuality, and redundancy without retraining the generator. The DPP-inspired log-determinant objective promotes diversity while maintaining budget constraints, and the modular design allows plug-and-play with any pretrained abstractive model.
This approach addresses a critical pain point in enterprise summarization tools: factual reliability. By decoupling generation from selection, organizations can deploy existing models and improve output trustworthiness without costly retraining, potentially accelerating adoption in legal, financial, and medical domains where accuracy is paramount.
The method offers a cost-effective path to more faithful AI-generated summaries, reducing the risk of misinformation in high-stakes applications. It can be retrofitted to existing summarization systems, lowering the barrier to improved factual grounding and potentially reducing liability for AI vendors.
Next signals include integration of the selector with large language model (LLM)-based generators, extension to multilingual and domain-specific summarization, and real-world deployment in fact-checking pipelines. The trade-off between coherence and factuality may drive further research into hybrid selection criteria.