摘要: This article is an exposition of a recent study on automatic generation of a structured overview (SOV) over a very large corpus of documents, where an SOV is organized as sections and subsections according to the latent hierarchy ... 展开 This article is an exposition of a recent study on automatic generation of a structured overview (SOV) over a very large corpus of documents, where an SOV is organized as sections and subsections according to the latent hierarchy of topics contained in the documents. We present a new framework called AutoOverview that includes and extends our previous scheme called NDORGS (best paper runner-up in ACM DocEng'2019) [47]. Different from the standard NLP task of generating a coherent summary typically over a handful of documents, AutoOverview needs to balance between two competitive objectives of accuracy and efficiency over thousands of documents. It incorporates hierarchical topic clustering, single-document summarization, multiple-document summarization, title generation, and other text mining techniques into a single platform. To assess the quality of an SOV generated over many documents, while it is possible to rely on human annota-tors to judge its readability, the sheer size of the inputs would make it formidable for human judges to determine if an SOV has covered all major points contained in the original texts. To overcome this obstacle, we present a text mining mechanism to evaluate topic coverage of the SOV against the topics contained in the original documents. We use multi-attribute decision making to help determine a suitable suite of algorithms to implement AutoOverview and the values of parameters for achieving a satisfactory SOV with respect to both accuracy and efficiency. We use NDORGS as an implementation example to address these issues and present evaluation results over a corpus of over 2,000 classified news articles and a corpus of over 5,000 unclassified news articles in a span of 10 years obtained from a search of the same keyword. 收起