German Abusive Language Dataset with Focus on COVID-19

German Abusive Language Dataset with Focus on COVID-19 Maximilian Wich∗, Svenja Räther∗, and Georg Groh Technical University of Munich, Department of Informatics, Germany maximilian.wich@tum.de, svenja.raether@tum.de, grohg@in.tum.de Abstract The COVID-19 pandemic has had a significant impact on human lives globally. As a result, it is unsurprising that it has influenced hate speech and other sorts of abusive language on social media. Machine learning models have been designed to automatically detect such posts and messages, which necessitate a signif- icant amount of labeled data. Despite the rele- vance of the COVID-19 topic in the context of abusive language, no annotated datasets with this focus are available. To solve these short- falls, we target to create such a dataset. Our contributions are as follows: (1) a methodol- ogy for collecting abusive language data from Twitter with a substantial amount of abusive and hateful content, and (2) a German abusive language dataset with 4,960 annotated tweets centered on COVID-19. Both the methodol- ogy and the dataset are intended to aid re- searchers in improving abusive language detec- tion. 1 Introduction Hate speech is a serious challenge that social me- dia platforms are currently confronting (Duggan, 2017). However, it is not limited to the online world. According to a study, there is a link be- tween online hate and physical crime (Williams et al., 2020). As a result, it is critical to combat hate speech and other forms of abusive language on social media platforms to improve the conversation atmosphere and prevent spillover. Owed to the large amounts of content created by billions of users, it is inefficient to detect this phenomenon manually. Therefore, its automatic detection is an essential part of the fight against this. Machine learning is a promising technology that aids in the training of classification models for detecting hate speech. ∗These authors contributed equally to this work. The success of a classification model depends largely on its training data. It requires data to learn patterns that can be used for solving the task. Large amounts of labeled data are required in the con- text of hate speech because hate speech is mul- tifaceted and diversified (e.g., misogyny, racism, anti-Semitism) (Rieger et al., 2021). As a result, researchers have published many abusive language datasets in recent years (Vidgen and Derczynski, 2020; Wich et al., 2021b; Schmidt and Wiegand, 2017). The majority of the datasets are in English, and only a small portion is in German. Another shortcoming of the existing datasets is that, with some exceptions, they do not cover COVID-19- related hate (Vidgen et al., 2020; Alshalan et al., 2020; Ziems et al., 2020). COVID-19, however, has become a popular topic in the hate and extremist communities (Guhl and Gerster, 2020; Velásquez et al., 2020), making it a suitable topic in the hate speech and abusive language detection community as well. Our research goal is to develop a Ger- man abusive language dataset with an emphasis on COVID-19 to solve both shortcomings. Contribution: • With a topical focus, we present a …

Content truncated to 500 words