Semi-supervised Multi-task Learning for Multi-label Fine-grained Sexism Classification

Proceedings of the 28th International Conference on Computational Linguistics, pages 5810–5820 Barcelona, Spain (Online), December 8-13, 2020 5810 Semi-supervised Multi-task Learning for Multi-label Fine-grained Sexism Classification Harika Abburi∗ IIIT-Hyderabad, India harika.a@research.iiit.ac.in Pulkit Parikh∗ IIIT-Hyderabad, India pulkit.parikh@research.iiit.ac.in Niyati Chhaya Adobe Research, India nchhaya@adobe.com Vasudeva Varma IIIT-Hyderabad, India vv@iiit.ac.in Abstract Sexism, a form of oppression based on one’s sex, manifests itself in numerous ways and causes enormous suffering. In view of the growing number of experiences of sexism reported online, categorizing these recollections automatically can assist the fight against sexism, as it can facili- tate effective analyses by gender studies researchers and government officials involved in policy making. In this paper, we investigate the fine-grained, multi-label classification of accounts (re- ports) of sexism. To the best of our knowledge, we work with considerably more categories of sexism than any published work through our 23-class problem formulation. Moreover, we propose a multi-task approach for fine-grained multi-label sexism classification that leverages several supporting tasks without incurring any manual labeling cost. Unlabeled accounts of sex- ism are utilized through unsupervised learning to help construct our multi-task setup. We also devise objective functions that exploit label correlations in the training data explicitly. Multi- ple proposed methods outperform the state-of-the-art for multi-label sexism classification on a recently released dataset across five standard metrics. 1 Introduction Sexism, defined as prejudice, stereotyping, or discrimination based on a person’s sex, occurs in various overt and subtle forms, permeating personal as well as professional spaces. While men and boys are also harmed by sexism, women and girls suffer the brunt of sexist mindsets and resultant wrongdoings. With increasingly many people sharing recollections of sexism experienced or witnessed by them, the automatic classification of these accounts into well-conceived categories of sexism can help fight this oppression, as it can better equip authorities formulating policies and researchers of gender studies to analyze sexism. The detection of sexism differs from and can complement the classification of sexism. In a forum where instances of sexism are mixed with other posts unrelated to sexism, sexism detection can be used to identify the posts on which to perform sexism classification. Moreover, we observe the distinction be- tween sexist statements (e.g., posts whereby one perpetrates sexism) and the accounts of sexism suffered or witnessed (e.g., personal recollections shared as part of the #metoo movement). We also note the prior work on detecting or classifying personal stories of sexual harassment and/or assault (Chowdhury et al., 2019; Karlekar and Bansal, 2018). In this paper, we focus on classifying an account (report) of sexism involving any set of categories of sexism. Most of the existing research on sexism classification (Anzovino et al., 2018; Jafarpour et al., 2018; Jha and Mamidi, 2017) considers at most five categories of sexism. Further, the majority of prior approaches associate only one category of sexism with an instance of sexism. Having mutually exclusive categories of sexism is unreasonable and limiting, as substantiated by Table 1. To the best of our knowledge, Parikh et al. (2019) is …

Content truncated to 500 words