MultiOpEd: A Corpus of Multi-Perspective News Editorials

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4345–4361 June 6–11, 2021. ©2021 Association for Computational Linguistics 4345 MULTIOPED: A Corpus of Multi-Perspective News Editorials Siyi Liu Sihao Chen Xander Uyttendaele Dan Roth University of Pennsylvania {siyiliu, sihaoc, xanderu, danroth}@seas.upenn.edu Abstract We propose MULTIOPED1, an open-domain news editorial corpus that supports various tasks pertaining to the argumentation structure in news editorials, focusing on automatic per- spective discovery. News editorial is a genre of persuasive text, where the argumentation struc- ture is usually implicit. However, the argu- ments presented in an editorial typically cen- ter around a concise, focused thesis, which we refer to as their perspective. MULTIOPED aims at supporting the study of multiple tasks relevant to automatic perspective discovery, where a system is expected to produce a single- sentence thesis statement summarizing the ar- guments presented. We argue that identifying and abstracting such natural language perspec- tives from editorials is a crucial step toward studying the implicit argumentation structure in news editorials. We first discuss the challenges and define a few conceptual tasks towards our goal. To demon- strate the utility of MULTIOPED and the in- duced tasks, we study the problem of perspec- tive summarization in a multi-task learning set- ting, as a case study. We show that, with the in- duced tasks as auxiliary tasks, we can improve the quality of the perspective summary gener- ated. We hope that MULTIOPED will be a use- ful resource for future studies on argumenta- tion in the news editorial domain. 1 Introduction News editorial is a form of persuasive text that con- veys consensus opinion on a controversial topic from the editors of a newspaper. Much like an ar- gumentative essay, a news editorial centers around a thesis, which represents the authors’ perspec- tive on the topic. Usually, a news editorial ar- gues in favor of the authors’ stance on the topic, and is substantiated by extensive factual evidence. 1The authors would like to thank Daniel Ravner, the CEO of www.theperspective.com, for granting access to data from the site for academic research. Figure 1: Structure of MULTIOPED. For each query on a controversial topic, two (rather long) news editori- als respond to the query from different point-of-views. Each editorial comes with a single paragraph abstract plus a one-sentence perspective, that abstractively sum- marizes the editorial’s key argument in the context of the query. The two resulting perspectives serve as re- sponses with opposite stance to the query. As news editorials function as professionally pro- duced written discourse for conveying media at- titude and guidance, they have traditionally been studied by the community as a rich resource for many argumantation-related tasks. (Wilson and Wiebe, 2003; Yu and Hatzivassiloglou, 2003; Bal and Saint-Dizier, 2009). This work targets the problem of developing computational methods to identify and compara- tively analyze the authors’ perspectives and sup- porting arguments behind news editorials. One challenge to studying the argumentation structure in news editorials is …

Content truncated to 500 words