Presenting a labelled dataset for real-time detection of abusive user posts

Abstract

Social media sites facilitate users in posting their own personal comments online. Most support free format user posting, with close to real-time publishing speeds. However, online posts generated by a public user audience carry the risk of containing inappropriate, potentially abusive content. To detect such content, the straightforward approach is to filter against blacklists of profane terms. However, this lexicon filtering approach is prone to problems around word variations and lack of context. Although recent methods inspired by machine learning have boosted detection accuracies, the lack of gold standard labelled datasets limits the development of this approach. In this work, we present a dataset of user comments, using crowdsourcing for labelling. Since abusive content can be ambiguous and subjective to the individual reader, we propose an aggregated mechanism for assessing different opinions from different labellers. In addition, instead of the typical binary categories of abusive or not, we introduce a third class of 'undecided' to capture the real life scenario of instances that are neither blatantly abusive nor clearly harmless. We have performed preliminary experiments on this dataset using best practice techniques in text classification. Finally, we have evaluated the detection performance of various feature groups, namely syntactic, semantic and context-based features. Results show these features can increase our classifier performance by 18% in detection of abusive content.