Unsupervised Discovery of Implicit Gender Bias

By Anjalie Field, Yulia Tsvetkov. 2020

55 citations recorded by the source corpus.

Despite their prevalence in society, social biases are difficult to define and identify, primarily because human judgements in this domain can be unreliable. Therefore, we take an unsupervised approach to identifying gender bias at a comment or sentence level, and present a model that can surface text likely to contain bias. The main challenge in this approach is forcing the model to focus on signs of implicit bias, rather than other artifacts in the data. Thus, the core of our methodology relies on reducing the influence of confounds through propensity score matching and adversarial learning. Our analysis shows how biased comments directed towards female politicians contain mixed criticisms and references to their spouses, while comments directed towards other female public figures focus on appearance and sexualization. Ultimately, our work offers a way to capture subtle biases in various domains without relying on subjective human judgements.

Subfields: Artificial Intelligence, Gender Studies, Political Science and International Relations, Natural Language Processing, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Gender Studies

In Political Science and International Relations

In Natural Language Processing

In Safety, Robustness & Interpretability

Citations in the corpus

Language Generation Models Can Cause Harm: So What Can We Do About It? An Actionable Survey (2022, methodology): “Hence, unsupervised or distantly supervised approaches have been adopted to detect them (Korzeniowski et al., 2019; Field and Tsvetkov, 2020; Sabri et al., 2021).”
Controlled Text Generation as Continuous Optimization with Multiple Constraints (2021, methodology): “These approaches can be used as a useful tool for mitigating many problematic biases already encoded in large language models [15, 2, 37], for anonymizing personal attributes [50], and even as aids for humans to avoid implicit biases in their writing [40, 14].”
Stepmothers are mean and academics are pretentious: What do pretrained language models learn about you? (2021, unknown): “…studying speciﬁc phenomena such as the infamous ‘Angry Black Woman’ stereotype and the ‘double bind’ (Heilman et al., 2004) theory (Kir-itchenko and Mohammad, 2018; May et al., 2019; Tan and Celis, 2019), or relating model predictions to gender stereotype lexicons (Field and Tsvetkov, 2020).”
Confronting Abusive Language Online: A Survey from the Ethical and Human Rights Perspective (2020, background): “Furthermore, implicit abuse presents additional challenges to human annotators as sometimes specific background knowledge and experience are required in order to understand the hidden meaning behind implicit statements (Sap et al., 2020; Breitfeller et al., 2019; Field & Tsvetkov, 2020).”
Gradient-based Constrained Sampling from Language Models (2022, background): “They also find useful applications in anonymizing protected attributes in written text (Reddy and Knight, 2016), as writing assistants to avoid users’ implicit biases (Ma et al., 2020; Field and Tsvetkov, 2020”
Gendered Mental Health Stigma in Masked Language Models (2022, background): “We draw this list of 66 nouns from Field and Tsvetkov (2020).”