Annotating Online Misogyny

Annotating Online Misogyny Philine Zeinert IT University of Copenhagen Denmark phze@itu.dk Nanna Inie IT University of Copenhagen Denmark nans@itu.dk Leon Derczynski IT University of Copenhagen Denmark leod@itu.dk Abstract Online misogyny, a category of online abusive language, has serious and harmful social con- sequences. Automatic detection of misogynis- tic language online, while imperative, poses complicated challenges to both data gathering, data annotation, and bias mitigation, as this type of data is linguistically complex and di- verse. This paper makes three contributions in this area: Firstly, we describe the detailed design of our iterative annotation process and codebook. Secondly, we present a comprehen- sive taxonomy of labels for annotating misog- yny in natural written language, and finally, we introduce a high-quality dataset of annotated posts sampled from social media posts. 1 Introduction Abusive language is a phenomenon with serious consequences for its victims, and misogyny is no exception. According to a 2017 report from Amnesty International, 23% of women from eight different countries have experienced online abuse or harassment at least once, and 41% of these said that on at least one occasion, these online experi- ences made them feel that their physical safety was threatened (Amnesty International, 2017). Automatic detection of abusive language can help identify and report harmful accounts and acts, and allows counter narratives (Chung et al., 2019; Garland et al., 2020; Ziems et al., 2020). Due to the volume of online text and the mental impact on humans who are employed to moderate online abusive language - moderators of abusive online content have been shown to develop serious PTSD and depressive symptoms (Casey Newton, 2020) - it is urgent to develop systems to automate the de- tection and moderation of online abusive language. Automatic detection, however, presents significant challenges (Vidgen et al., 2019). Abusive language is linguistically diverse (Vid- gen and Derczynski, 2020), both explicitly, in the form of swear words or profanities; implicitly, in the form of sarcasm or humor (Waseem et al., 2017); and subtly, in the form of attitudes and opin- ions. Recognizing distinctions between variants of misogyny is challenging for humans, let alone com- puters. Systems for automatic detection are usually created using labeled training data (Kiritchenko et al., 2020), hence, their performance depends on the quality and representativity of the available datasets and their labels. We currently lack trans- parent methods for how to create diverse datasets. When abusive language is annotated, classes are of- ten created based on each unique dataset (a purely inductive approach), rather than taking advantage of general, established terminology from, for in- stance, social science or psychology (a deductive approach, building on existing research). This makes classification scores difficult to compare and apply across diverse training datasets. This paper investigates the research question: How might we design a comprehensive annotation process which results in high quality data for au- tomatically detecting misogyny? We make three novel contributions: 1. Methodology: We describe our iterative approach to the annotation process in a transparent way which allows for a higher degree …

Content truncated to 500 words