Abusive language is a massive problem in online social platforms. Existing abusive language detection techniques are particularly ill-suited to comments containing heterogeneous abusive language patterns, i.e., both abusive and non-abusive parts. This is due in part to the lack of datasets that explicitly annotate heterogeneity in abusive language. We tackle this challenge by providing an annotated dataset of abusive language in over 11,000 comments from YouTube. We account for heterogeneity in this dataset by separately annotating both the comment as a whole and the individual sentences that comprise each comment. We then propose an algorithm that uses a supervised attention mechanism to detect and categorize abusive content using multi-task learning. We empirically demonstrate the challenges of using traditional techniques on heterogeneous content and the comparative gains in performance of the proposed approach over state-of-the-art methods.

Current computer involvement in adolescent social networks (youth between the ages of 11 and 17) provides new opportunities to study group dynamics, interactions amongst peers, and individual preferences. Nevertheless, most of the research in this area focuses on efficiently retrieving information that is explicit in large social networks (e.g., properties of the graph structure), but not on how to use the dynamics of the virtual social network to discover latent characteristics of the real-world social network. In this paper, we present the analysis of a game designed to take advantage of the familiarity of adolescents with online social networks, and describe how the data generated by the game can be used to identify bullies in 5th grade classrooms. We present a probabilistic model of the game and using the in-game interactions of the players (i.e., content of chat messages) infer their social role within their classroom (either a bully or non-bully). The evaluation of our model is done by using previously collected data from psychological surveys on the same 5th grade population and by comparing the performance of the new model with off-the-shelf classifiers.