Learning to Detect Online Harassment on Twitter with the Transformer

This paper describes our submission to the SIMAH challenge (SocIaL Media And Harassment). The proposed competition addresses the challenge of harassment detection on Twitter posts as well as the identification of a harassment category. Automatically detecting content containing harassment could be the basis for removing it. Accordingly, the task is considered to be an essential step to distinguishing different types of harassment provides the means to control such a mechanism in a fine-grained way. In this work, we classify a set of Twitter posts into non-harassment or harassment tweets where the last ones are classified as indirect harassment, sexual harassment, or physical harassment. We explore how to use self-attention models for harassment classification in order to combine different baselines' outputs. For a given post, we use the transformer architecture to encode each baseline output exploiting relationships between baselines and posts. Then, the transformer learns how to combine the outputs of these methods with a BERT representation of the post, reaching a macro-averaged F-score of 0.481 on the SIMAH test set.