/biased-rulers

A survey of fairness in contextualized language models

Primary LanguageJupyter Notebook

Biased rulers

Tests

A Survey on Quantifying Biases in Pretrained Language Models

Metrics

  • metrics.seat: Implementations of SEAT, and variations of SEAT by Lauscher et al. (2021) and Tan et al. (2019).
  • metrics.disco: Implementation of DisCo.
  • metrics.lpbs: Implementation of the log probability bias score.

Notebooks

-The notebooks folder contains subfolders LM_correlations, sublist correlations, winobias

  • LM_correlations: contains correlation results on how different template types and embedding types correlate based on scores from different language models.
  • sublist_correlations: contains correlation results on how different template types and embedding types correlate based on scores from different sublist of jobs.
  • correlation_analysis*.ipynb in each folder contains the results from the correlation experiments
  • The codes used for Winobias experiments are directly adopted from Manela et al. (2021) https://github.com/12kleingordon34/NLP_masters_project