The final day at AoIR 2019 begins for me with a panel on social media bots, and the first speakers are Felix Münch and Ben Thies who present a paper that I have also contributed to; the slides are below. Social bots have become quite prominent in media coverage of social media in recent times, with particular focus on platforms like Twitter, but it is difficult to assess just how prevalent they are on such platforms, partly also because it is difficult to get a sense of the make-up of larger social media populations.
This project examines the prevalence, influence, and roles of bots in the German-speaking Twittersphere. It also aims to make available the tools to do similar work for other Twitter populations. We have already developed an approach to establish a solid and representative sample of the German Twittersphere (covered elsewhere), and in this paper apply the Botometer tool to these accounts as well as manually checking whether some such accounts are truly bots in any sense of that definition.
The project used the rank degree method to identify a solid subset of the German-language Twittersphere (that is, accounts that have their interface language set to German), over a period of several months, and captured some 1 million German accounts that had 1.6 million edges between them; from these, it selected the 200,000 most central accounts and identified the network clusters they belonged to. Keyword analysis of the themes that are prevalent in each cluster also make it possible to identify what themes these clusters are focussing on. Tests show that for any random German-language Twitter account, some 40% of the accounts followed by that account show up in our sample.
We then used the Botometer tool to assess the ‘botness’ of the accounts in our sample, and – according to the assessment by that tool – there are relatively few bots in the sample; they are also especially prominent in just a handful of the network clusters (especially clusters focussing on spam and politics, with a handful also in a far right politics cluster).
Yet Botometer may not be entirely reliable, and its assessment criteria for accounts are not always entirely clear. We therefore selected the 50 bots with that highest page rank that had received a Botometer rating of 0.75 or higher, and manually examined those accounts for their botness; (only) 65% of those accounts were rated as actual bots by the manual coding. Such bots included both benign (interesting images, famous quotes, news updates) bots, automated followback accounts, automated repost bots that shared links to blogs and similar Websites, and malignant bots (promoting pirated software etc.).
This means that Botometer scores should not necessarily be trusted – but they are useful as a starting point for developing bot typologies. Also, bot studies should not necessarily limit themselves to political issues, since many of the bots found are non-political.
