SupMMD: A Sentence Importance Model for Extractive Summarization using Maximum Mean Discrepancy

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4108–4122 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 4108 SupMMD: A Sentence Importance Model for Extractive Summarization using Maximum Mean Discrepancy Umanga Bista⋆ Alexander Patrick Mathews⋆ Aditya Krishna Menon‡ Lexing Xie⋆ ⋆Australian National University, Canberra, ACT, Australia ‡Google Research, New York, NY, United States ⋆{umanga.bista,alex.mathews,lexing.xie}@anu.edu.au, ‡adityakmenon@google.com Abstract Most work on multi-document summarization has focused on generic summarization of infor- mation present in each individual document set. However, the under-explored setting of update summarization, where the goal is to identify the new information present in each set, is of equal practical interest (e.g., presenting readers with updates on an evolving news topic). In this work, we present SupMMD, a novel technique for generic and update summarization based on the maximum mean discrepancy from ker- nel two-sample testing. SupMMD combines both supervised learning for salience and un- supervised learning for coverage and diversity. Further, we adapt multiple kernel learning to make use of similarity across multiple infor- mation sources (e.g., text features and knowl- edge based concepts). We show the efficacy of SupMMD in both generic and update sum- marization tasks by meeting or exceeding the current state-of-the-art on the DUC-2004 and TAC-2009 datasets. 1 Introduction Multi-document summarization is the problem of producing condensed digests of salient information from multiple sources, such as articles. Concretely, suppose we are given two sets of articles (denoted set A and set B) on a related topic (e.g., climate change, the COVID-19 pandemic), separated by publication timestamp or geographic region. We may then identify three possible instantiations of multi-document summarization (see Figure 1): (i) generic summarization, where the goal is to summarize a set (A or B) individually. (ii) comparative summarization, where the goal is to summarize a set (B) against another set (A) while highlighting the differences. (iii) update summarization, where the goal is both generic summarization of set A and compara- tive summarization of set B versus A. A (a) (b) (c) B B A A Figure 1: Different summarization tasks: (a) Generic (b) Comparative (c) Update. Two sets of articles (set A and B) are denoted by red and blue circles, respectively. Summary prototypes are bold circles, and information coverage of each tasks is filled with respective colors. Most existing work on this topic has focused on the generic summarization task. However, update summarization is of equal practical interest. Intu- itively, the comparative aspect of this setting aims to inform a user of new information on a topic they are already familiar with. Multi-document extractive summarization meth- ods can be unsupervised or supervised. Unsuper- vised methods typically define salience (or cov- erage) using a global model of sentence-sentence similarity. Methods based on retrieval (Goldstein et al., 1999), centroids (Radev et al., 2004), graph centrality (Erkan and Radev, 2004), or utility maxi- mization (Lin and Bilmes, 2010, 2011; Gillick and Favre, 2009) have been well explored. However, sentence salience also depends on surface features (e.g., position, length, presence of cue words); …

Content truncated to 500 words