MDR Cluster-Debias: A Nonlinear WordEmbedding Debiasing Pipeline
arXiv:2006.11642
Abstract
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.
References in corpus (6)
- Semantics derived automatically from language corpora contain human-like biases
- Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings
- FRAGE: Frequency-Agnostic Word Representation
- Attenuating Bias in Word Vectors
- Mitigating Gender Bias in Natural Language Processing: Literature Review
- Gender-preserving Debiasing for Pre-trained Word Embeddings