Abstract: Long-document summarization remains challenging due to complex document structures and excessive input length, which often introduce substantial noise into generative models and hinder effective knowledge acquisition. To mitigate this issue, a method integrating document compression and entity prompting was proposed to filter irrelevant content and reinforce the learning of key information. An end-to-end framework termed DCEPSum (Document Compression and Entity Prompted Summarization) was developed, in which soft-selection document compression was combined with entity prompting to generate coherent summaries. First, within the document compression module, a heterogeneous graph was constructed over sentence, entity, and section nodes, and a multi-head graph attention network was employed to aggregate contextual information and output sentence-importance weights; sentences with high weights were retained to reduce noise and enhance cross-sentence associations. Subsequently, to guide the model to focus on key entity information and reduce noise interference, a key entity selection and entity prompting mechanism was introduced. During generation, within an encoder architecture extended by a state space model, the selected entities were inserted into the input context as prefix prompts, and the final summaries were generated. Experimental results on two benchmark datasets show that the ROUGE (Recall-Oriented Understudy for Gisting Evaluation) -2 scores of DCEPSum are 23.18 and 20.86, and the ROUGE-L scores are 46.63 and 45.83; compared with the strongest baseline LSG (Local, Sparse and Global attention) (16k), ROUGE-2 increases by 0.76 and 0.67, and ROUGE-L increases by 2.31 and 3.14. The combination of document compression and entity prompting improves long-document summarization quality under controllable computational overhead and provides a feasible solution for long-context summarization modeling.

Key words: long document summarization, document compression, entity prompt, Graph Neural Network (GNN), State-Space Model (SSM)

摘要： 长文档摘要任务由于文档结构复杂、篇幅较长而面临显著挑战。现有的大多数生成式框架在训练过程中易受噪声干扰，导致无关语义信息影响有效知识的获取。为缓解这一问题，提出一种融合文档压缩与实体提示的方法，用以过滤无关内容并强化对关键信息的学习。该方法提出基于文档压缩与实体提示的端到端框架——DCEPSum (Document Compression and Entity Prompted Summarization)，通过“软筛选”文档压缩与实体提示相结合，以生成连贯的摘要。首先，文档压缩模块基于句子、实体与章节节点构建异质图，并采用多头图注意力网络聚合上下文信息并输出句子重要性权重，保留高权重句子以降低噪声并增强跨句关联性。随后，为引导模型关注关键实体信息并减少噪声干扰，引入关键实体选择与实体提示机制。在生成阶段，该框架基于状态空间模型扩展的编码器结构中，将选定的实体作为前缀提示插入输入上下文，实现最终摘要生成。在两个基准数据集上的实验结果表明，DCEPSum的ROUGE-2 (Recall-Oriented Understudy for Gisting Evaluation) 分别为23.18与20.86，ROUGE-L分别为46.63与45.83；相较最强基准LSG(Local, Sparse and Global attention) (16k)，ROUGE-2分别提高0.76与0.67，ROUGE-L分别提高2.31与3.14。文档压缩与实体提示的结合能够在可控计算开销下改善长文档摘要质量，为长上下文摘要建模提供可行方案。

关键词: 长文档摘要, 文档压缩, 实体提示, 图神经网络, 状态空间模型