A multi-platform dataset for detecting cyberbullying in social media

Recent work on cyberbullying detection relies on using machine learning models with text and metadata in small datasets, mostly drawn from single social media platforms. Such models have succeeded in predicting cyberbullying when dealing with posts containing the text and the metadata structure as found on the platform. Instead, we develop a multi-platform dataset that consists purely of the text from posts gathered from seven social media platforms. We present a multi-stage and multi-technique annotation system that initially uses crowdsourcing for post and hashtag annotation and subsequently utilizes machine-learning methods to identify additional posts for annotation. This process has the benefit of selecting posts for annotation that have a significantly greater than chance likelihood of constituting clear cases of cyberbullying without limiting the range of samples to those containing predetermined features (as is the case when hashtags alone are used to select posts for annotation). We show that, despite the diversity of examples present in the dataset, good performance is possible for models trained on datasets produced in this manner. This becomes a clear advantage compared to traditional methods of post selection and labeling because it increases the number of positive examples that can be produced using the same resources and it enhances the diversity of communication media to which the models can be applied.