Scoring Sentence Singletons and Pairs for Abstractive Summarization

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2175–2189 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 2175 Scoring Sentence Singletons and Pairs for Abstractive Summarization Logan Lebanoff† Kaiqiang Song† Franck Dernoncourt§ Doo Soon Kim§ Seokhwan Kim§ Walter Chang§ Fei Liu† †Computer Science Department, University of Central Florida, Orlando, FL 32816 {loganlebanoff, kqsong}@knight.ucf.edu feiliu@cs.ucf.edu §Adobe Research, San Jose, CA 95110 {dernonco,dkim,seokim,wachang}@adobe.com Abstract When writing a summary, humans tend to choose content from one or two sentences and merge them into a single summary sentence. However, the mechanisms behind the selec- tion of one or multiple source sentences remain poorly understood. Sentence fusion assumes multi-sentence input; yet sentence selection methods only work with single sentences and not combinations of them. There is thus a cru- cial gap between sentence selection and fusion to support summarizing by both compressing single sentences and fusing pairs. This pa- per attempts to bridge the gap by ranking sen- tence singletons and pairs together in a uni- fied space. Our proposed framework attempts to model human methodology by selecting ei- ther a single sentence or a pair of sentences, then compressing or fusing the sentence(s) to produce a summary sentence. We conduct ex- tensive experiments on both single- and multi- document summarization datasets and report findings on sentence selection and abstraction. 1 Introduction Abstractive summarization aims at presenting the main points of an article in a succinct and coherent manner. To achieve this goal, a proficient editor can rewrite a source sentence into a more succinct form by dropping inessential sentence elements such as prepositional phrases and adjectives. She can also choose to fuse multiple source sentences into one by reorganizing the points in a coherent manner. In fact, it appears to be common practice to summarize by either compressing single sen- tences or fusing multiple sentences. We investi- gate this hypothesis by analyzing human-written abstracts contained in three large datasets: DUC- 04 (Over and Yen, 2004), CNN/Daily Mail (Her- mann et al., 2015), and XSum (Narayan et al., 2018). For every summary sentence, we find its ground-truth set containing one or more source CNN/DM DUC−04 XSum 0% 20% 40% 60% 80% InstanceType Compression (1) Fusion (2) Fusion (3+) Figure 1: Portions of summary sentences generated by compression (content is drawn from 1 source sentence) and fusion (content is drawn from 2 or more source sen- tences). Humans often grab content from 1 or 2 docu- ment sentences when writing a summary sentence. sentences that exhibit a high degree of similarity with the summary sentence (details in §4). As shown in Figure 1, across the three datasets, 60- 85% of summary sentences are generated by fus- ing one or two source sentences. Selecting summary-worthy sentences has been studied in the literature, but there lacks a mecha- nism to weigh sentence singletons and pairs in a unified space. Extractive methods focus on select- ing sentence singletons using greedy (Carbonell and Goldstein, 1998), optimization-based (Gillick and Favre, 2009; …

Content truncated to 500 words