InriaFBK Drawing Attention to Offensive Language at Germeval2019

In this paper we describe the system developed by InriaFBK team and submitted to the Germeval2019 task on offensive language detection and classification. With the same architecture we participate to all subtasks: binary classification of offensive and not offensive tweets, 4-class message categorisation based on offense type (Profanity, Insult, Abuse and Other), and classification of explicit and implicit offensive language. The two runs submitted for each subtask are obtained with and without attention mechanism. After evaluating our system performance on Germeval2018 test set, we observe that attention is remarkably beneficial in the more challenging tasks of implicit offense detection and offense categorisation.