Grouping business news stories based on salience of named entities

Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 1096–1106, Valencia, Spain, April 3-7, 2017. c⃝2017 Association for Computational Linguistics Grouping business news stories based on salience of named entities Llorenc¸ Escoter, Lidia Pivovarova, Mian Du, Anisia Katiskaya and Roman Yangarber Department of Computer Science University of Helsinki, Finland first.last@cs.helsinki.fi Abstract In news aggregation systems focused on broad news domains, certain stories may appear in multiple articles. Depending on the relative importance of the story, the number of versions can reach dozens or hundreds within a day. The text in these versions may be nearly identi- cal or quite different. Linking multiple versions of a story into a single group brings several important benefits to the end-user—reducing the cognitive load on the reader, as well as signaling the rela- tive importance of the story. We present a grouping algorithm, and explore several vector-based representations of input doc- uments: from a baseline using keywords, to a method using salience—a measure of importance of named entities in the text. We demonstrate that features beyond keywords yield substantial improvements, verified on a manually-annotated corpus of business news stories. 1 Introduction We address the problem of grouping multiple ver- sions of the same story in a system that contin- uously processes articles incoming from a large number of news streams. Our problem setting is PULS—an on-line information extraction (IE) system, which analyses news in the business do- main.1 PULS collects articles from over 1000 on-line sources that provide RSS feeds. Among approximately 4000–6000 articles arriving daily, some of the stories appear multiple times. The role of the aggregation component is to cluster the articles into a set of stories—a story is a set of articles that report the same piece of news. 1http://puls.cs.helsinki.fi/ The purpose of grouping is that when a user is- sues a query, the system should show one item per story, rather than one item per article, so the same fact is not shown over and over again. Informa- tion should be presented in a way that minimizes redundancy of the returned results; this implies narrowly clustering news articles that describe the same events or facts. Another goal is to identify trending stories. In the business domain, if a story is globally impor- tant, it will appear in many feeds. When repeated news items are identified, the number and variety of sources covering the story is an indicator of that story’s relative importance. Thus, from the end-user’s perspective, we have at least two motivations for grouping different news that describe the same story: reducing re- dundancy and ranking stories by importance. The main contributions of this paper are: • We try to identify the most effective doc- ument representation for news clustering. We demonstrate that automatically extracted named entities (NE) are better features than words. Moreover, considering the salience of NEs—a measure that combines frequency and prominence of the NE—gives further im- provement in clustering performance. We in- troduce a …

Content truncated to 500 words