Finding a representative subset from large-scale documents

Large-scale information, especially in the form of documents, is potentially useful for decision-making but intensifies the information overload problem. To cope with this problem, this paper proposes a method named RepExtract to extract a representative subset from large-scale documents. The extracted representative subset possesses three desirable features: high coverage of the content of the original document set, low redundancy within the extracted subset, and consistent distribution with the original set. Extensive experiments were conducted on benchmark datasets, demonstrating the superiority of RepExtract over the benchmark methods in terms of the three features above. A user study was also conducted by collecting human evaluations of different methods, and the results indicate that users can gain an understanding of large-scale documents precisely and efficiently through a representative subset extracted by the proposed method.