Mining Both Commonality and Specificity From Multiple Documents for Multi-Document Summarization

The multi-document summarization task requires the designed summarizer to generate a short text that covers the important information of original multiple documents and satisfies content diversity. To fulfill the dual requirements of coverage and diversity in multi-document summarization, this study introduces a novel method. Initially, a class tree is constructed through hierarchical clustering of documents. Subsequently, a sentence selection method based on class tree is proposed for generating a summary. Specifically, a top-down traversal is performed on the class tree, during which sentences are selected from each node based on their similarity to the centroid of the documents within the node and their dissimilarity to the centroid of documents not belonging to the node. Sentences selected from the root node reflect the commonality of all document, and sentences selected from the sub nodes reflect the distinct specificity of the respective subclasses. Experimental results on standard text summarization datasets DUC'2002, DUC'2003, and DUC'2004 demonstrate that the proposed method significantly outperforms the variant method that considers only commonality of all documents, achieving average improvements of up to 1.54 and 1.42 in ROUGE-1 and ROUGE-L scores, respectively. Additionally, the method demonstrates significant superiority over another variant method that considers only the specificity of subclasses, achieving average improvements of up to 2.16 and 2.01 in ROUGE-1 and ROUGE-L scores, respectively. Furthermore, extensive experiments on DUC'2004 and Multi-News datasets show that the proposed method outperforms lots of competitive supervised and unsupervised multi-document summarization methods and yields considerable results.