Towards Coherent Multi-Document Summarization

Proceedings of NAACL-HLT 2013, pages 1163–1173, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Towards Coherent Multi-Document Summarization Janara Christensen, Mausam, Stephen Soderland, Oren Etzioni Computer Science & Engineering University of Washington Seattle, WA 98195, USA {janara,mausam,soderlan,etzioni}@cs.washington.edu Abstract This paper presents G-FLOW, a novel system for coherent extractive multi-document sum- marization (MDS).1 Where previous work on MDS considered sentence selection and or- dering separately, G-FLOW introduces a joint model for selection and ordering that balances coherence and salience. G-FLOW’s core rep- resentation is a graph that approximates the discourse relations across sentences based on indicators including discourse cues, deverbal nouns, co-reference, and more. This graph en- ables G-FLOW to estimate the coherence of a candidate summary. We evaluate G-FLOW on Mechanical Turk, and find that it generates dramatically bet- ter summaries than an extractive summarizer based on a pipeline of state-of-the-art sentence selection and reordering components, under- scoring the value of our joint model. 1 Introduction The goal of multi-document summarization (MDS) is to produce high quality summaries of collections of related documents. Most previous work in ex- tractive MDS has studied the problems of sentence selection (e.g., (Radev, 2004; Haghighi and Vander- wende, 2009)) and sentence ordering (e.g., (Lapata, 2003; Barzilay and Lapata, 2008)) separately, but we believe that a joint model is necessary to produce coherent summaries. The intuition is simple: if the sentences in a summary are first selected—without regard to coherence—then a satisfactory ordering of the selected sentences may not exist. 1System and data at http://knowitall.cs.washington.edu/gflow/ doc1: Bomb- ing in Jerusalem doc1: Anger from Israelis doc1: Suspen- sion of peace accord due to bombing doc2: Hamas claims respon- sibility doc5: Pales- tinians con- demn attack doc4: Mubarak urges peace accord doc5: Pales- tinians urge peace accord doc3: Clinton urges peace accord Figure 1: An example of a discourse graph covering a bombing and its aftermath, indicating the source docu- ment for each node. A coherent summary should begin with the bombing and then describe the reactions. Sen- tences are abbreviated for compactness. An extractive summary is a subset of the sen- tences in the input documents, ordered in some way.2 Of course, most possible summaries are in- coherent. Now, consider a directed graph where the nodes are sentences in the collection, and each edge represents a pairwise ordering constraint necessary for a coherent summary (see Figure 1 for a sample graph). By definition, any coherent summary must obey the constraints in this graph. Previous work has constructed similar graphs au- tomatically for single document summarization and manually for MDS (see Section 2). Our system, G-FLOW extends this research in two important ways. First, it tackles automatic graph construction for MDS, which requires novel methods for identi- fying inter-document edges (Section 3). It uses this 2We focus exclusively on extractive summaries, so we drop the word “extractive” henceforth. 1163 State-of-the-art MDS system G-FLOW • The attack took place Tuesday near Cailaco in East Timor, a former Portuguese colony, according to a statement issued by the …

Content truncated to 500 words