Expert-annotated dataset to study cyberbullying in Polish language

We present the first dataset for the Polish language containing annotations of harmful and toxic language. The dataset was created to study harmful Internet phenomena such as cyberbullying and hate speech, which have dramatically gained in numbers in recent years both on Polish Internet as well as worldwide. The dataset was automatically collected and annotated in two ways. Firstly, by two trained layperson volunteers under the supervision of a cyberbullying and hate-speech expert. To improve the quality of annotations, the second turn of annotations was performed by a group of trained expert annotators specializing in the annotation of cyberbullying and hate-speech data, which was also additionally supervised by an additional experienced expert annotator (or super-annotator). We initially utilize the dataset in the classification of cyberbullying in Polish. In particular, the dataset is utilized in two tasks: 1) binary classification of harmful and non-harmful messages, and 2) multi-class classification between two types of harmful information (cyberbullying and hate speech), and others. Apart from the dataset itself, we also share the classification model which achieved the highest classification results for the dataset to be freely applied by third parties in cyberbullying prevention architectures.