MDR Cluster-Debias: A Nonlinear WordEmbedding Debiasing Pipeline
This addresses the issue of incomplete debiasing in word embeddings for NLP applications, though it is incremental as it builds on prior methods and reveals limitations in current bias tests.
The authors tackled the problem of residual gender bias in word embeddings that persists after debiasing, identifying reasons for this clustering and developing the MDR Cluster-Debias pipeline. Their method significantly outperforms existing approaches on upstream bias tests, with improvements of up to 30% in some metrics, but shows limited gains in downstream tasks.
Existing methods for debiasing word embeddings often do so only superficially, in that words that are stereotypically associated with, e.g., a particular gender in the original embedding space can still be clustered together in the debiased space. However, there has yet to be a study that explores why this residual clustering exists, and how it might be addressed. The present work fills this gap. We identify two potential reasons for which residual bias exists and develop a new pipeline, MDR Cluster-Debias, to mitigate this bias. We explore the strengths and weaknesses of our method, finding that it significantly outperforms other existing debiasing approaches on a variety of upstream bias tests but achieves limited improvement on decreasing gender bias in a downstream task. This indicates that word embeddings encode gender bias in still other ways, not necessarily captured by upstream tests.