Benevolent Sexism Detection in Text: A Data-Centric Approach

Abstract

The Ambivalent Sexism theory divides sexism into two-dimensional ideologies: benevolent sexism and hostile sexism. Hostile sexism has been associated with short-term harmful impacts, and benevolent sexism has been proven to have more severe long-term effects on women’s well-being, their representation, and gender equality in societies. Recently, research has been directed toward the detection and mitigation of hostile sexism, and minimal efforts have been done with the aim of detecting and mitigating benevolent sexism. Adversely, since benevolent sexism is associated with a seemingly positive expression, detecting and mitigating its online spread is a challenge that needs the attention of social scientists, gender scholars, and data scientists. In this paper, we aim toward creating a benevolent sexism detection system. To the best of our knowledge, the research area lacks a representative benevolent sexism dataset. Thus, to be able to train supervised machine learning models, we collected and labeled a dataset of benevolent, hostile, and non-sexist statements collected from quotes’ websites, online articles, and the Google Advanced Search tool. Further, we trained several machine learning models and incrementally tuned and optimized the best classifier for the detection of benevolent sexism. Then, we validated our model’s performance on similar and broader context datasets and detailed its strengths, weaknesses, and areas of improvement. Our final results confirm our model’s ability to detect benevolent sexism in a generalized context. To emphasize, the dataset collected was proven to perform well in the representation of the benevolent sexism expression. In conclusion, this research is a steppingstone to creating a self-learning, data-centric benevolent sexism detection system.