Social bias in language models can potentially exacerbate social inequalities. Despite it having garnered wide attention, most research focuses on English data. In a low-resource scenario, the models often perform worse due to insufficient training data. This study aims to leverage high-resource language corpora to evaluate bias and experiment with debiasing methods in low-resource languages. We evaluated the performance of recent multilingual models in five languages: English, Chinese, Russian, Indonesian and Thai, and analyzed four bias dimensions: gender, religion, nationality, and race-color. By constructing multilingual bias evaluation datasets, this study allows fair comparisons between models across languages. We have further investigated three debiasing methods-CDA, Dropout, SenDeb-and demonstrated that debiasing methods from high-resource languages can be effectively transferred to low-resource ones, providing actionable insights for fairness research in multilingual NLP.

Pre-trained large language models (LLMs) reflect the inherent social biases of their training corpus. Many methods have been proposed to mitigate this issue, but they often fail to debias or they sacrifice model accuracy. We use conceptors--a soft projection method--to identify and remove the bias subspace in LLMs such as BERT and GPT. We propose two methods of applying conceptors (1) bias subspace projection by post-processing by the conceptor NOT operation; and (2) a new architecture, conceptor-intervened BERT (CI-BERT), which explicitly incorporates the conceptor projection into all layers during training. We find that conceptor post-processing achieves state-of-the-art (SoTA) debiasing results while maintaining LLMs' performance on the GLUE benchmark. Further, it is robust in various scenarios and can mitigate intersectional bias efficiently by its AND operation on the existing bias subspaces. Although CI-BERT's training takes all layers' bias into account and can beat its post-processing counterpart in bias mitigation, CI-BERT reduces the language model accuracy. We also show the importance of carefully constructing the bias subspace. The best results are obtained by removing outliers from the list of biased words, combining them (via the OR operation), and computing their embeddings using the sentences from a cleaner corpus.

Bolukbasi et al. (2016) presents one of the first gender bias mitigation techniques for word embeddings. Their method takes pre-trained word embeddings as input and attempts to isolate a linear subspace that captures most of the gender bias in the embeddings. As judged by an analogical evaluation task, their method virtually eliminates gender bias in the embeddings. However, an implicit and untested assumption of their method is that the bias sub-space is actually linear. In this work, we generalize their method to a kernelized, non-linear version. We take inspiration from kernel principal component analysis and derive a non-linear bias isolation technique. We discuss and overcome some of the practical drawbacks of our method for non-linear gender bias mitigation in word embeddings and analyze empirically whether the bias subspace is actually linear. Our analysis shows that gender bias is in fact well captured by a linear subspace, justifying the assumption of Bolukbasi et al. (2016).

Debiasing word embeddings has been largely limited to individual and independent social categories. However, real-world corpora typically present multiple social categories that possibly correlate or intersect with each other. For instance, "hair weaves" is stereotypically associated with African American females, but neither African American nor females alone. Therefore, this work studies biases associated with multiple social categories: joint biases induced by the union of different categories and intersectional biases that do not overlap with the biases of the constituent categories. We first empirically observe that individual biases intersect non-trivially (i.e., over a one-dimensional subspace). Drawing from the intersectional theory in social science and the linguistic theory, we then construct an intersectional subspace to debias for multiple social categories using the nonlinear geometry of individual biases. Empirical evaluations corroborate the efficacy of our approach. Data and implementation code can be downloaded at https://github.com/GitHubLuCheng/Implementation-of-JoSEC-COLING-22.

Online texts -- across genres, registers, domains, and styles -- are riddled with human stereotypes, expressed in overt or subtle ways. Word embeddings, trained on these texts, perpetuate and amplify these stereotypes, and propagate biases to machine learning models that use word embeddings as features. In this work, we propose a method to debias word embeddings in multiclass settings such as race and religion, extending the work of (Bolukbasi et al., 2016) from the binary setting, such as binary gender. Next, we propose a novel methodology for the evaluation of multiclass debiasing. We demonstrate that our multiclass debiasing is robust and maintains the efficacy in standard NLP tasks.