【摘要】 多文档抽取式摘要任务(MDES)旨在从多个相关文档中提取一个简明且包含显著信息的摘要。通常，在同主题的多个文档中冗余信息不可避免，例如，因不同表达方式造成的重复描述等。现有大多数方法在抽取摘要时，仅关注显著性内容的检测或冗余信息的过滤二者之一，导致摘要信息不全面、不准确。因此，在建模抽取式多文档摘要任务时如何权衡两者间的协作是个挑战。考虑到多文档领域缺乏大规模训练数据，该文提出了一个新的多阶段的多文档无监督文本摘要抽取模型，该模型在摘要级别上进行提取，并通过以下三个步骤依次解决冗余性去除问题和显著性检测问题：引入外部知识的噪声过滤机制、冗余感知的排序策略，以及显著性感知的重排序策略。实验结果表明，该文框架可在多文档数据集Multi-News上取得无监督方法的最优结果，并在两个单文档数据集上获得有竞争力的结果。更多还原

【Abstract】 Multi-document extractive summarization(MDES) task aims to extract a concise, informative, and comprehensive summary from several related documents within the same topic. It is a challenge for MDES to focus on both saliency detection and diversity(de-redundancy). In this paper, we propose a novel multi-stage content selection framework for unsupervised MDES. It deals with both the saliency detection and redundancy removal by three stages at summary-level: noise filtering stage with external knowledge, redundancy-aware ranking stage with key information coverage and saliency-aware re-ranking stage with hybrid centrality in a heterogeneous graph. Empirical results show that the proposed method achieves new record performance on Multi-News, as well as competitive results on two single document summarization datasets.更多还原