Measuring News Similarity Across Ten U.S. News Sites

TL;DR

This paper presents a method to identify top news stories and measure their similarity across ten U.S. news websites over three months, revealing how news coverage converges around major events like elections.

Contribution

The paper introduces a novel approach combining headline extraction and cosine similarity to quantify news similarity and identify top stories across multiple news sites.

Findings

Similarity increased before Election Day and after, indicating convergence in news coverage.

The method effectively identifies top stories and measures their similarity over time.

News coverage shows notable shifts during major events like elections.

Abstract

News websites make editorial decisions about what stories to include on their website homepages and what stories to emphasize (e.g., large font size for main story). The emphasized stories on a news website are often highly similar to many other news websites (e.g, a terrorist event story). The selective emphasis of a top news story and the similarity of news across different news organizations are well-known phenomena but not well-measured. We provide a method for identifying the top news story for a select set of U.S.-based news websites and then quantify the similarity across them. To achieve this, we first developed a headline and link extractor that parses select websites, and then examined ten United States based news website homepages during a three month period, November 2016 to January 2017. Using archived copies, retrieved from the Internet Archive (IA), we discuss the methods and difficulties for parsing these websites, and how events such as a presidential election can lead news websites to alter their document representation just for these events. We use our parser to extract k = 1, 3, 10 maximum number of stories for each news site. Second, we used the cosine similarity measure to calculate news similarity at 8PM Eastern Time for each day in the three months. The similarity scores show a buildup (0.335) before Election Day, with a declining value (0.328) on Election Day, and an increase (0.354) after Election Day. Our method shows that we can effectively identity top stories and quantify news similarity.