Verlag der Österreichischen Akademie der Wissenschaften

Abstract:
In this paper, we describe two systems for predicting message-level offensive language in German tweets: one discriminates between offensive and not offensive messages, and the second performs a fine-grained classification by recognizing also classes of offense. Both systems are based on the same approach, which builds upon Recurrent Neural Networks used with the following features: word embeddings, emoji embeddings and social-network specific features. The model is able to combine word-level information and tweet-level information in order to perform the classification tasks.