Weak Supervision and Machine Learning for Online Harassment Detection

Abstract

Automated detection tools can enable both the study of online harassment and technology that mitigates its harm. Machine learning methods allow these tools to adapt and improve using data. Yet current well-established machine learning approaches require amounts of data that are often unmanageable by practitioners aiming to train harassment detectors. Emerging methods that learn models from weak supervision represent one important avenue to address this challenge. In contrast to the full supervision used in most traditional machine learning methods, weak supervision does not require annotators to label individual examples of the target concept. Instead, annotators provide approximate descriptions of the target concept, sucsh as rule-of-thumb indicators. In this chapter, we describe the weak supervision paradigm and some general principles that drive emerging methods. And we detail a weakly supervised method for detection of online harassment that uses key-phrase indicators as the form of weak supervision. This method considers multiple aspects of the online harassment phenomenon, using interplay between these aspects to bolster the weak supervision into a useful model. We describe experimental results demonstrating this approach on detecting harassment in social media data. Finally, we discuss the ongoing challenges for using machine learning methods to build harassment detectors.