Research map / Label-efficient and robust learning / Recognition and long-tail
Understanding the Origins of Bias in Word Embeddings
Marc-Etienne Brunet, Colleen Alkalay-Houlihan, Ashton Anderson, Richard Zemel
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use influence-function estimates to select documents for removal | GloVe bias attribution · Wiki and NYT, WEAT1 and WEAT2 | correlation r^2 with retraining results: 0.986; 0.993; 0.995; 0.997 | Improved |
| Select the 10 most bias-increasing documents using the proposed method instead of PPMI | GloVe, Wiki WEAT1 · Wiki WEAT1, remove 10 documents | WEAT effect size change: -4% -> -40% | Improved |
| Remove the 10k most bias-contributing documents identified for GloVe | word2vec · NYT WEAT1 | WEAT effect size: 1.35 -> 0.11 | Hurt |
| Remove randomly selected document sets | GloVe embeddings · Wiki and NYT, WEAT1 and WEAT2 | significant difference from baseline bias: For 36 random perturbation sets, only 2 differed significantly (p < 0.05). | No clear effect |
| Remove documents targeted by the proposed method | GloVe embeddings · Wiki and NYT, WEAT1 and WEAT2 | significant difference from baseline bias: For 40 targeted perturbation sets, only 2 did not significantly differ from their respective baselines. | Improved |
| Remove documents influential for GloVe | GloVe TOP-1 analogy performance · Perturbed embeddings | TOP-1 analogy performance | No clear effect |
| Remove GloVe-identified bias-influencing documents | word2vec TOP-1 analogy performance · NYT | TOP-1 analogy performance | No clear effect |
| Remove documents containing rare WEAT words | WEAT sensitivity to corpus perturbations · NYT WEAT1 | WEAT effect size: r^2 = 0.828; 0.07% | Mixed |
Closest papers by meaning
- Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings (NeurIPS 2016 · similarity 0.92)
- On Measuring and Mitigating Biased Inferences of Word Embeddings (AAAI 2020 · similarity 0.92)
- Are Models Biased on Text without Gender-related Language? (ICLR 2024 · similarity 0.86)
- Debiasing Algorithm through Model Adaptation (ICLR 2024 · similarity 0.85)
- By Tying Embeddings You Are Assuming the Distributional Hypothesis (ICML 2024 · similarity 0.85)
- Biases Mitigation and Expressiveness Preservation in Language Models: A Comprehensive Pipeline (Student Abstract) (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
