Measuring Gender Bias in Natural Language Processing: Incorporating Gender-Neutral Linguistic Forms for Non-Binary Gender Identities in Abusive Speech Detection

Proceedings of Recent Advances in Natural Language Processing, pages 1121–1131 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_119 1121 Measuring Gender Bias in Natural Language Processing: Incorporating Gender-Neutral Linguistic Forms for Non-Binary Gender Identities in Abusive Speech Detection Nasim Sobhani SFI Centre for Research Training in Machine Learning Technological University Dublin nasim.x.sobhani@mytudublin.ie Kinshuk Sengupta Microsoft Dublin, Ireland kinshuk.sengupta@microsoft.com Sarah Jane Delany SFI Centre for Research Training in Machine Learning Technological University Dublin sarahjane.delany@tudublin.ie Abstract Predictions from Machine Learning models can reflect bias in the data on which they are trained. Gender bias has been shown to be prevalent in Natural Language Processing models. The re- search into identifying and mitigating gender bias in these models predominantly considers gender as binary, male and female, neglecting the fluidity and continuity of gender as a vari- able. In this paper, we present an approach to eval- uate gender bias in a prediction task, which recognises the non-binary nature of gender. We gender-neutralise a random subset of existing real-world hate speech data. We extend the ex- isting template approach for measuring gender bias to include test examples that are gender- neutral. Measuring the bias across a selection of hate speech datasets we show that the bias for the gender-neutral data is closer to that seen for test instances that identify as male than those that identify as female. 1 Introduction Natural Language Processing (NLP) models and systems are developed by using text content cre- ated by humans and they may incorporate biases that exist in the data. These biases can then be reflected in the results produced by these models and systems when they are used in downstream applications (Dixon et al., 2018; Park et al., 2018). Additionally, word embeddings, which are repre- sentations of words and sentences generated from large amounts of natural language text, may also exhibit and even magnify certain features of the data, such as gender stereotypes (Bolukbasi et al., 2016; Caliskan et al., 2017; Zhao et al., 2017). An issue with existing research is that it con- siders gender as binary neglecting the fluidity and continuity of gender as a variable (Stanczak and Augenstein, 2021). Many of the data resources in NLP currently are inadequate for identifying gender bias as they often have a significant under- representation of female or non-binary instances. (Stanczak and Augenstein, 2021). There is a need to incorporate gender-neutral linguistic forms in datasets and algorithms to recognise the non-binary nature of gender. This impacts on algorithms too, such as language models which learn meaningless unstable representations for non-binary associated pronouns and terms (Dev et al., 2021). In this paper, we present an approach to measure gender bias in a downstream task to identify abu- sive or hate speech that considers male, female, and gender-neutral gender identities. Due to the lack of datasets that include gender-neutral linguistic forms, we adjust existing real-world datasets by gender-neutralising a random subset of instances. A challenge with measuring gender bias in nat- ural language training data is the lack of gender identification in …

Content truncated to 500 words