A Macroscopic Analysis of News Content in Twitter

Abstract

Previous literature has considered the relevance of Twitter to journalism, for example as a tool for reporters to collect information and for organizations to disseminate news to the public. We consider the reciprocal perspective, carrying out a survey of news media-related content within Twitter. Using a random sample of 1.8 billion tweets over four months in 2014, we look at the distribution of activity across news media and the relative dominance of certain news organizations in terms of relative share of content, the Twitter behavior of news media, the hashtags used in news content versus Twitter as a whole, and the proportion of Twitter activity that is news media-related. We find a small but consistent proportion of Twitter is news media-related (0.8 percent by volume); that news media-related tweets focus on a different set of hashtags than Twitter as a whole, with some hashtags such as those of countries of conflict (Arab Spring countries, Ukraine) reaching over 15 percent of tweets being news media-related; and we find that news organizations’ accounts, across all major organizations, largely use Twitter as a professionalized, one-way communication medium to promote their own reporting. Using Latent Dirichlet Allocation topic modeling, we also examine how the proportion of news content varies across topics within 100,000 #Egypt tweets, finding that the relative proportion of news media-related tweets varies vastly across different subtopics. Over-time analysis reveals that news media were among the earliest adopters of certain #Egypt subtopics, providing a necessary (although not sufficient) condition for influence.

KEYWORDS:

• computational
• news media
• social media
• topic modeling
• Twitter

ACKNOWLEDGEMENT

Thanks to Cornelia Brantner for discussion and guidance around journalism theory.

Notes

1. See also “Success Stories: Media, News & Publishing” (https://biz.twitter.com/success-stories/industry/media-news-publishing, accessed December 10, 2015), where Twitter pitches “success stories” of “curated content” to potential clients. However, note that success is shown by many different metrics. While this may be attributed to companies having different goals in their Twitter use, and Twitter itself gives a typology of strategies at https://business.twitter.com/, when we look at comparable products with presumably comparable goals (such as the “products” of Mitt Romney and Barack Obama’s respective 2012 presidential campaigns), we see that even then different metrics are presented as evidence of success.

2. Accessed August 16, 2015.

3. The list of news media organization Twitter handles used in this article can be accessed at http://www.pfeffer.at/data/news-on-twitter/.

4. From http://www.alexa.com/topsites/category/Top/News (accessed August 16, 2015).

5. This means we miss tweets linking to news media websites through these shortened URLs, so we likely undercount the number of tweets with news URLs. For reference, there are 37,462,621 bit.ly links, 8,713,644 ow.ly links, 7,988,958 goo.gl links, and 7,714,077 tinyurl links among the 1.8 billion collected tweets. Note that since February 2013, Twitter forcibly displays all URLs in 23 characters using its own “t.co” shortening service (see “Twitter Now Reducing Some Tweets to 117 Characters,” http://mashable.com/2013/02/20/twitter-tco-length/, and “Posting Links in a Tweet,” https://support.twitter.com/articles/78124, accessed December 10, 2015), removing the incentive that users would have previously had to use URL shortening services to save space (Antoniades et al. 2011; Grier et al. 2010; Maggi et al. 2013; Wang et al. 2013). The one paper analyzing URL shortening services after the changeover (Gupta, Aggarwal, and Kumaraguru 2014) does not consider the uses of such services other than for spam, and do not have an estimate of what proportion of tweets with shortened URLs are spam.

6. The eponym Dirichlet is of the nineteenth-century German mathematician; a probability distribution based on his work was named after him, and this probability distribution is the basis for LDA.

7. This part of the process is very similar to a technique common in social science, Principle Component Analysis, in how the analyst interprets what loadings represent.

8. Note that for estimating just this proportion, it was not necessary to have a sample as large as the decahose; we could have also estimated this with the Sample API, or even 1/10,000th of the Sample API. The advantage of having the full data, or a far larger sample, is in being able to get accurate estimates of observations in the tail of the rankings, as here observations are very sparse and hence it is much harder to get accurate estimates from smaller samples. Furthermore, large samples allow for drawing subsamples (such as the one we take for #Egypt) large enough to perform meaningful inference.

9. See http://www.theguardian.com/fashion/2014/may/29/roberto-cavalli-perfume-offends-sufi-students (accessed December 10, 2015).

10. See https://about.twitter.com/company (accessed December 10, 2015).

Additional information

Funding

This work was supported in part by the Office of Naval Research under MINVERVA [grant number N000141310835]. Momin is supported in part by a grant from the ARCS Foundation.