Combatting abusive behavior in online communities using cross-community learning

Abstract

This dissertation aims to develop a deep understanding of abusive online behavior via statistical machine learning techniques to build tools that help counter it. I have developed computational approaches to model abusive online behavior, aiming to address two of the major gaps in this line of research—the scarcity of labeled ground truth required to train effective ML models, and the contextual nature of online moderation by accounting for community-specific norms. First, I introduced a new class of machine learning tools that are based on cross-community linguistic similarity. Next, I discovered the existence of widely overlapping norms, across distinct online communities, suggesting that new automated tools for moderation could find traction in borrowing data from communities which share similar values. The abuse models that I build will enable a brand new class of interactive machine learning systems that can sidestep the need for site-specific classifiers. My thesis brings these pieces together in the form of open source software to detect abusive behavior online through cross-community learning, and thereby socio-algorithmically govern speech on large-scale Internet platforms like Reddit.