The consistently high prevalence of hate speech on the Internet continues to pose significant social and individual challenges. Given the centrality of social networks in public discourse, automating the identification of criminally relevant content is a pressing challenge. This study addresses the challenge of developing an automated system that is capable of classifying online comments in a criminal justice context and categorising them into relevant sections of the criminal code. Not only technical, but also ethical and legal requirements must be considered. To this end, 351 comments were annotated by public prosecutors from the Central Office for Combating Internet and Computer Crime (ZIT) according to previously formed paragraph classes. These groupings consist of several German criminal law statutes that most hate comments violate. In the subsequent phase of the research, a further 839 records were assigned to the classes by student annotators who had been trained previously.

In this work, we present a new publicly available offensive language dataset of 10.278 German social media comments collected in the first half of 2021 that were annotated by in total six annotators. With twelve different annotation categories, it is far more comprehensive than other datasets, and goes beyond just hate speech detection. The labels aim in particular also at toxicity, criminal relevance and discrimination types of comments. Furthermore, about half of the comments are from coherent parts of conversations, which opens the possibility to consider the comments’ contexts and do conversation analyses in order to research the contagion of offensive language in conversations.

In this work, we present our approaches on the toxic comment classification task (subtask 1) of the GermEval 2021 Shared Task. For this binary task, we propose three models: a German BERT transformer model; a multilayer perceptron, which was first trained in parallel on textual input and 14 additional linguistic features and then concatenated in an additional layer; and a multilayer perceptron with both feature types as input. We enhanced our pre-trained transformer model by re-training it with over 1 million tweets and fine-tuned it on two additional German datasets of similar tasks. The embeddings of the final fine-tuned German BERT were taken as the textual input features for our neural networks. Our best models on the validation data were both neural networks, however our enhanced German BERT gained with a F1-score = 0.5895 a higher prediction on the test data.