Updated Headline Generation: Creating Updated Summaries for Evolving News Stories

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 6438 - 6461 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Updated Headline Generation: Creating Updated Summaries for Evolving News Stories Sheena Panthaplackel1∗ Adrian Benton2† Mark Dredze2,3 1 Computer Science, The University of Texas at Austin, TX, USA 2 Bloomberg, New York, NY USA 3 Computer Science, Johns Hopkins University, Baltimore, MD USA spantha@cs.utexas.edu, adbenton@google.com, mdredze@cs.jhu.edu Abstract We propose the task of updated headline gen- eration, in which a system generates a head- line for an updated article, considering both the previous article and headline. The system must identify the novel information in the ar- ticle update, and modify the existing headline accordingly. We create data for this task us- ing the NewsEdits corpus (Spangher and May, 2021) by automatically identifying contiguous article versions that are likely to require a sub- stantive headline update. We find that models conditioned on the prior headline and body re- visions produce headlines judged by humans to be as factual as gold headlines while making fewer unnecessary edits compared to a standard headline generation model. Our experiments establish benchmarks for this new contextual summarization task. 1 Introduction Automatic text summarization condenses the most important and salient information from a large quantity of text. The task takes many different forms depending on the type of information be- ing summarized, the modality of the information, the type of summary desired and the needs of the end user. Examples include news headline gener- ation (Banko et al., 2000; Zajic et al., 2002; Dorr et al., 2003; Takase et al., 2016; Matsumaru et al., 2020), summarization of social media (Liu et al., 2012; Ding and Jiang, 2015; Kim et al., 2019), and medical documents (Schulze and Neves, 2016; Liang et al., 2019; Adams et al., 2021). In many settings, users encounter information progressively instead of all at once. For instance, ∗Work done during an internship at Bloomberg. †Now at Google Research. news stories are revised as events unfold (Tannier and Moriceau, 2013), social media streams evolve as people post content (Tarnpradab et al., 2021), and biomedical texts are revised as clinical trial results emerge (uptodate, 2021). In such dynamic settings, existing summaries should be updated as new information becomes available. To address this, we could in principle leverage static summa- rization systems for generating a summary of the underlying content at any given point in time. How- ever, a more natural approach would be to produce a new summary based on what the reader already knows and what content changed. Consider the case of a news article being updated as events unfold (Figure 1). The article first reports that a man is charged with stealing an ice cream van, and the article is later updated when the man admits to the crime. By the time the article is updated, the reader already knows what was stolen, who was charged, and where it happened. At this point, the reader is most interested …

Content truncated to 500 words