四川大学学报（自然科学版）2025，Vol.62Issue(2)：347-358,12.DOI:10.19907/j.0490-6756.240241
生成式摘要的事实一致性与文本质量的平衡性研究
The study on balancing factual consistency and text quality in abstractive summarization
摘要
Abstract
Enhancing factual consistency has become a research hotspot in the field of abstractive summariza-tion.Current mainstream methods can be categorized into two approaches:post-editing of generated summa-ries and optimization of model mechanisms.While these methods effectively improve factual consistency,they often sacrifice text quality and readability.To address this issue,the authors propose an abstractive sum-marization model named SumRCL(Summarization with Reinforcement and Contrastive Learning)that com-bines reinforcement learning with ranking-based contrastive learning.On the one hand,the authors leverage ranking-based contrastive learning on candidate summaries to enhance the correlation between the probability assigned to a summary by the model and its factual consistency.On the other hand,the authors employ rein-forcement learning based on text quality metrics to preserve high-quality text.Specifically,the authors utilize Monte Carlo search to address the issue of intermediate summary evaluation.Experiments on the CNN/DM and XSUM datasets demonstrate that our proposed SumRCL model indeed contributes to generating summa-ries with both high factual consistency and text quality.The authors analyze the effects of the number of candi-date summaries and the choice of ranking metrics in contrastive learning on the final performance.Finally,through manual evaluation,the authors demonstrate that SumRCL exhibits superior factual consistency com-pared to popular large language models.关键词
生成式摘要/事实一致性/强化学习/对比学习/大语言模型Key words
Abstractive summarization/Factual consistency/Reinforcement learning/Contrastive learning/Large language model分类
计算机与自动化引用本文复制引用
杨昱睿,何禹瞳,琚生根..生成式摘要的事实一致性与文本质量的平衡性研究[J].四川大学学报（自然科学版）,2025,62(2):347-358,12.基金项目
国家自然科学基金重点项目(62137001) （62137001）
四川省重点研发项目(2023YFG0265) （2023YFG0265）
