Research map / Label-efficient and robust learning / Recognition and long-tail
On Measuring and Mitigating Biased Inferences of Word Embeddings
Sunipa Dev, Tao Li, Jeff M. Phillips, Vivek Srikumar
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 |
|---|---|---|---|
| Project out the he-she direction | GloVe · Gender-occupation | NN: 0.387 -> 0.480; +24.7% | Improved |
| Project out the demonym subspace | GloVe · Demonym-polarity | NN: 0.713 -> 0.808; +13.3% | Improved |
| Project out the adherent subspace | GloVe · Religion-polarity | NN: 0.710 -> 0.794; +11.8% | Improved |
| Project out a random vector instead of the gender direction (ablation) | GloVe · Gender-occupation | NN: 0.387 -> 0.362; -6.0% | No clear effect |
| Project out the gender direction from all ELMo layers (ablation) | ELMo · Gender-occupation | NN: 0.417 -> 0.423; +1.6% | No clear effect |
| Project out a random direction from all ELMo layers (ablation) | ELMo · Gender-occupation | T:0.7: 0.063 -> 0.115; +82.5% | Mixed |
| Project out the he-she direction from ELMo layer 1 | ELMo · Gender-occupation | NN: 0.417 -> 0.488; +17.3% | Improved |
| Project out the demonym direction from ELMo layer 1 (ablation) | ELMo · Demonym-polarity | NN: -10.7% | Hurt |
| Project out the adherent direction from ELMo layer 1 (ablation) | ELMo · Religion-polarity | NN: -13.2% | Hurt |
| Apply the BERT gender projection only at test time (ablation) | BERT · Gender-occupation | NN: 0.421 -> 0.396; -5.9% | Hurt |
| Apply the BERT gender projection during training and at test time | BERT · Gender-occupation | NN: 0.421 -> 0.516; +22.6% | Improved |
| Apply the BERT gender projection during training and at test time | BERT · SNLI test | accuracy: 90.22 -> 90.23 | No clear effect |
Closest papers by meaning
- Understanding the Origins of Bias in Word Embeddings (ICML 2019 · similarity 0.92)
- Biases Mitigation and Expressiveness Preservation in Language Models: A Comprehensive Pipeline (Student Abstract) (AAAI 2024 · similarity 0.86)
- 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)
- SocialStigmaQA: A Benchmark to Uncover Stigma Amplification in Generative Language Models (AAAI 2024 · similarity 0.85)
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.
