Syn-D-CNN Dataset

Text summarization condenses extensive content into concise summaries; however, current approaches often rely on large language models (LLMs), which can lack interpretability and are susceptible to generating hallucinated content. To address these issues, we propose Docusage, an interpretable framework that replicates human summaries through a hierarchical clustering approach combined with extractive summarization, augmented by selective, LLM-based abstraction. Docusage minimizes the risk of hallucinations, ensures contextual relevance, and mitigates the computational costs inherent in leveraging an LLM.<br>Our results show that Docusage aligns closely with journalist-generated summaries, outperforming foundational and specialized models. Additionally, Docusage offers an interpretable framework that is not constrained by context size, ensures transparency regarding the role of extracted sentences within the narrative, and adapts to the style of the training data.文本摘要的核心目标是将海量内容压缩为简洁凝练的摘要，但当前主流的摘要生成方法多依赖大语言模型（Large Language Model，LLM），这类模型往往存在可解释性不足的缺陷，且极易产生幻觉式生成内容。为解决上述问题，本文提出Docusage框架：这是一种可解释的摘要生成方案，通过分层聚类结合抽取式摘要的方式复刻人工撰写的摘要，并辅以基于大语言模型的选择性抽象生成环节。Docusage能够有效降低幻觉生成风险，保障摘要与原文的上下文相关性，同时缓解了直接依赖大语言模型时固有的计算开销问题。 实验结果表明，Docusage生成的摘要与新闻从业者撰写的人工摘要高度贴合，其性能优于基础大模型与垂直领域专用模型。此外，Docusage提供的可解释框架不受上下文长度限制，可清晰展示抽取语句在原文叙事中的作用，同时能够适配训练数据的文本风格。