Investigating Annotator Bias in Abusive Language Datasets

Proceedings of Recent Advances in Natural Language Processing, pages 1515–1525 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_170 1515 Investigating Annotator Bias in Abusive Language Datasets Maximilian Wich, Christian Widmer, Gerhard Hagerer, Georg Groh Technical University of Munich, Germany {maximilian.wich,christian.widmer,gerhard.hagerer}@tum.de, grohg@in.tum.de Abstract Nowadays, social media platforms use classi- fication models to cope with hate speech and abusive language. The problem of these mod- els is their vulnerability to bias. A prevalent form of bias in hate speech and abusive lan- guage datasets is annotator bias caused by the annotators subjective perception and the com- plexity of the annotation task. In our paper, we develop a set of methods to measure an- notator bias in abusive language datasets and to identify different perspectives on abusive language. We apply these methods to four different abusive language datasets. Our pro- posed approach supports annotation processes of such datasets and future research address- ing different perspectives on the perception of abusive language. 1 Introduction A challenge that social media platforms are fac- ing in recent years is the large amount of hate speech and other forms of abusive language (Dug- gan, 2017). Manual monitoring, however, is no longer possible due to the vast volume of user- generated content. Therefore, machine learning models are trained and used by social media plat- forms, such as Facebook, to automatically detect such content (Kantor, 2020). According to Rose (2021), these models are a key component of Face- book’s fight against hate speech. A problem with such machine learning models is that they are vulnerable to bias (Vidgen and Der- czynski, 2021; Dixon et al., 2018). Biased models can strongly impair the fairness of a system, which can lead to discrimination (Dixon et al., 2018). Bias in abusive language detection is already a topic that researchers have started to investigate (Vidgen and Derczynski, 2021; Wich et al., 2021). The type of bias we will focus on in this study is annotator bias. This form of bias is a result of an- notators who perceive abusive language differently from each other and have different levels of expe- rience and knowledge (Ross et al., 2016; Waseem, 2016; Geva et al., 2019; Wich et al., 2020). We aim to investigate two aspects of annotator bias. (1) Assuming that there is only one perspec- tive (one truth) on whether a text is abusive or not, we develop an approach to measure and visualize annotator bias. This approach optimizes the an- notation process (e.g., outlier detection, adapting appropriate annotation guidelines). (2) Acknowl- edging multiple valid views on a text (e.g., a group has a more liberal attitude towards abusive texts, while another is stricter), we aim to identify annota- tor groups to model different, yet valid perspectives. The questions resulting from these research objec- tives are the following: • RQ1: How can we measure and visualize an- notator bias in abusive language datasets? • RQ2: How can we identify and visualize dif- ferent annotator perspectives on abusive lan- guage? Our contributions are the following: 1. To characterize annotators, …

Content truncated to 500 words