Automated Detection of Toxicity in Twitch Live Streaming Communities

• In Event: AI in Criminal Justice: Empirical and Normative Perspectives

Abstract

Social media encompass different microenvironments where criminogenic features might vary, constituting each one of them a cyber place. Routine activities on cyberspace have shifted to new places, between them live streaming communities. These are environments led by a content creator, whose economic success depends on building a strong sense of belonging of users to a community. If some studies have studied some crimes on these spaces (hate crimes), and some forms of victimization (content creators victimization) the criminogenic characteristics of these places are relatively unknown.
The present study attempts to measure criminogenic variables of streamers' channels on Twitch. The hypothesis is that some characteristics of the cyber place will influence prevalence on victimization. Particularly, we hypothesize that visibility of the stream, the gender of the streamer, and the type of activity of the stream will affect victimization rates.
To that end, we use Google's Perspective API algorithm to detect severe toxicity, insults, identity attacks, threats, and sexually explicit content in a corpus of Twitch chats. One hundred and eighty streams were collected from 60 streamers, collecting data on channel sizes, streamer gender and the content of the stream (just chatting, video game related content and sports-oriented content) to examine correlation between the independent variables and victimization.
Toxicity prevalence analysis is pending, and preliminary results are expected to be presented on the American Society of Criminology 79th Annual Meeting.

Author

• Mario Santisteban Galarza, UPV / EHU Euskal Herriko Unibertzitatea