Read Top News First: A Document Reordering Approach for Multi-Document News Summarization

Findings of the Association for Computational Linguistics: ACL 2022, pages 613 - 621 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Read Top News First: A Document Reordering Approach for Multi-Document News Summarization Chao Zhao1∗ Tenghao Huang1∗ Somnath Basu Roy Chowdhury1 Muthu Kumar Chandrasekaran Kathleen McKeown2 Snigdha Chaturvedi1 {zhaochao, tenghao, somnath, snigdha}@cs.unc.edu cmkumar087@gmail.com, kathy@cs.columbia.edu 1 UNC Chapel Hill 2 Columbia University Abstract A common method for extractive multi- document news summarization is to re- formulate it as a single-document summariza- tion problem by concatenating all documents as a single meta-document. However, this method neglects the relative importance of documents. We propose a simple approach to reorder the documents according to their relative impor- tance before concatenating and summarizing them. The reordering makes the salient content easier to learn by the summarization model. Experiments show that our approach outper- forms previous state-of-the-art methods with more complex architectures. 1 Introduction Multi-document news extractive summarization (MDS) aims to extract the salient information from multiple related news documents into a concise summary. Some approaches use task-specific archi- tectures for this problem. For example, Wang et al. (2020) organize multiple documents as a hetero- geneous graph before summarizing them. Zhong et al. (2020) formulate the extractive summariza- tion task as a semantic matching problem. Recent works also explored reformulating this problem as a single-document summarization (SDS) prob- lem by concatenating all documents into a single meta-document and then using an SDS model to summarize it (Cao et al., 2017; Liu et al., 2018; Lebanoff et al., 2018; Fabbri et al., 2019). Due to the conventions of news writing (Hong and Nenkova, 2014; Hicks et al., 2016), salient in- formation often appears at the beginning of a news article. As a result, many summarization systems, including recent neural models (Kedzie et al., 2018; Zhong et al., 2019), pay more attention to the be- ginning of the document. Therefore, in MDS, it is important to consider the order in which the docu- ∗ Equal Contribution ments are concatenated to form the meta-document before applying the summarization model. Specifically, we argue that the various docu- ments in the input are not equally important. Some documents contain more salient or detailed infor- mation and are more important. Therefore, com- pared with concatenating documents in an arbitrary order, it would be beneficial to reorder the docu- ments such that the important ones are in the front of the meta-document and it becomes easier for the summarization model to learn the salient content. Motivated by these factors, we propose a simple yet effective approach to reorder the input docu- ments according to their relative importance before applying a summarization model. We evaluate the effectiveness of our approach on Multi-News (Fab- bri et al., 2019) and DUC-2004. 1 Results show that our simple reordering approach significantly outperforms the state-of-the-art methods with more complex model architectures. We also observe that this approach brings more performance gain with the increase in the number of input documents. 2 Method We refer to D as a …

Content truncated to 500 words