Detection of Abusive Language: the Problem of Biased Datasets

Proceedings of NAACL-HLT 2019, pages 602–608 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 602 Detection of Abusive Language: the Problem of Biased Datasets Michael Wiegand∗0, Josef Ruppenhofer†‡, Thomas Kleinbauer∗ ∗Spoken Language Systems, Saarland University, Saarbr¨ucken, Germany †Leibniz ScienceCampus, Heidelberg/Mannheim, Germany ‡Institute for German Language, Mannheim, Germany michael.wiegand@lsv.uni-saarland.de ruppenhofer@ids-mannheim.de thomas.kleinbauer@lsv.uni-saarland.de Abstract We discuss the impact of data bias on abu- sive language detection. We show that clas- sification scores on popular datasets reported in previous work are much lower under realis- tic settings in which this bias is reduced. Such biases are most notably observed on datasets that are created by focused sampling instead of random sampling. Datasets with a higher proportion of implicit abuse are more affected than datasets with a lower proportion. 1 Introduction Abusive or offensive language is commonly de- fined as hurtful, derogatory or obscene utterances made by one person to another person.1 Examples are (1)-(3). In the literature, closely related terms include hate speech (Waseem and Hovy, 2016) or cyber bullying (Zhong et al., 2016). While there may be nuanced differences in meaning, they are all compatible with the general definition above. (1) stop editing this, you dumbass. (2) Just want to slap the stupid out of these bimbos!!! (3) Go lick a pig you arab muslim piece of scum. Due to the rise of user-generated web content, in particular on social media networks, the amount of abusive language is also steadily growing. NLP methods are required to focus human review ef- forts towards the most relevant microposts. In this paper, we examine the issue of data bias. For the creation of manually annotated datasets, randomly sampling microposts from large social media platforms typically results in a too small proportion of abusive comments (Wulczyn et al., 2017; Founta et al., 2018). Therefore, more fo- cused sampling strategies have to be applied which 0Present affiliation: berg/Mannheim, GermanyLeibniz ScienceCampus, Heidel- 1http://thelawdictionary.org/ cause biases in the resulting datasets. We show what implications this has on classifiers trained on these datasets: Previous evaluations reported high classification performance on datasets with diffi- cult cases of abusive language, e.g. implicit abuse (§2). Contrarily, we find that the high classifica- tion scores are likely to be the result of modeling the bias in those datasets. Although we will explicitly name shortcomings of existing individual datasets, our paper is not in- tended as a reproach of those who created them. On the contrary, we acknowledge the great ef- forts the researchers have taken to provide these resources. Without them, much existing research would not have been possible. However, we also noticed a lack of awareness of the special prop- erties of those datasets among researchers using them. As we will illustrate with specific examples, this may result in unforeseen results of particular classification approaches. 2 Explicit and Implicit Abuse One major distinction that has been proposed in the literature is the division into explicitly and im- plicitly abusive language (Waseem et al., 2017). The former are microposts that …

Content truncated to 500 words