 |
Center for Nonlinear Studies |
Understanding the drivers leading to collective behavior is important for many areas in science from social movements and disaster response to Internet trolling. Abusive speech online, a form of Internet trolling, is a growing threat and can result in devasting consequences such as physical harm, manipulation, and control of public discourse. The ability to detect and track the spread of these behaviors is challenging given the vast amount of data and the variability in the text being used. We utilized reddit, a social news and media aggregation bulletin board network to evaluate the dynamics of hate speech in varying communities. Using multiple machine and deep learning hate speech detection models and graph theoretical approaches, we analyzed several sub-reddit threads. We inferred dynamic characteristics of topics strongly correlated with abuse. We will integrate these into subreddit level network analysis to build a behavior based feature set associated with collective abuse. These features will be integrated with our language models to identify topics that were missed by our language model, and predict troll swarm (brigading) events in near real-time. Chrysm will also discuss previous work in Complex Adaptive Systems modelling, biosurveillance, and IOT security.
Host: Sara Del Valle