Joint Mitigation of Interactional Bias

Anonymous

Abstract: Machine learning algorithms have been found discriminative against groups of different social identities, e.g., gender and race. With the detrimental effects of these algorithmic biases, researchers proposed promising approaches for bias mitigation, typically designed for individual bias types. Due to the complex nature of social bias, we argue it is important to study how different biases interact with each other, i.e., how mitigating one bias type (e.g., gender) influences the bias results regarding other social identities (e.g., race and religion). We further question whether jointly debiasing multiple types of bias is desired in different contexts, e.g., when correlations between biases are different. To address these research questions, we examine bias mitigation in two NLP tasks -- toxicity detection and word embeddings -- on three social identities, i.e., race, gender, and religion. Empirical findings based on benchmark datasets suggest that different biases can be correlated and therefore, warranting attention for future research on joint bias mitigation.

Paper Type: long

Consent To Share Data: yes