Public justice

Mining user opinions relating to content moderation

Thales Bertaglia, Giovanni de Gregorio, Catalina Goanta

In recent years, the banning of streamers from platforms like Twitch because of alleged violations of community guidelines or terms of service (e.g. on hate speech or harassment) is increasingly becoming an example of how platforms exercise discretion unilaterally.

As an illustration, in the e-sports world, popular Brazilian streamers Jukes or Yoda were banned for using racial slurs and other derogatory words during their streams. In the political arena, Renan Bolsonaro was banned from Twitch for denying the covid-19 pandemic. In the United States, Youtuber Onision was banned from Patreon for doxing a person on Twitter, after the latter publicly accused him of sexual harassment. In spite of the viral hashtag #deplatformpredators, Onision remains active on Youtube and Twitter. Platform interventions or lack thereof leave audiences unhappy, and viewers, fans or followers take it to social media to express their opinions about these incidents and what justice means in this context.

Taking the public perception of morals as a starting point, we aim to use machine learning methods toanalyse public reactions to Youtubers/streamers being banned from the platforms (or sanctioned in other ways) due to controversial content. Such a context allows us to mine for opinions about three different topics:

  1. The cause of the ban itself: people are expressing what they perceive to be offensive or not and why.
  2. The streamer: people are either defending or attacking the streamers’ actions based on personal reasons, like their audience, personality, type of activity etc.
  3. The platform: an overwhelming amount of audience comments e.g. with respect to Twitch are about content moderation policies.

By applying computational methods such as sentiment analysis, stance detection or ordinal quantification on a data set of comments/posts taken from online streaming/social media platforms (e.g. Twitter, Youtube, Twitch), we can determine the overall sentiment expressed by the author of a post, the stance of authors with respect to a proposition, as well as the prevalence of positive or negative posts about a given incident/person.

The empirical findings will then be compared with a legal assessment of the validity of these removals, which will combine both constitutional and contractual implications, in order to understand potential tensions between fundamental rights and freedom of contract.

Research output