Skip to content

Latest commit

 

History

History
69 lines (58 loc) · 4.84 KB

README.md

File metadata and controls

69 lines (58 loc) · 4.84 KB
Raw

Bias-out-of-the-box

We include code and data for the paper Bias Out-of-the-Box: An Empirical Analysis of Intersectional Occupational Biases in Popular Generative Language Models

Figure: GPT-2 Monte-Carlo prediction vs ground truth US population share. GPT-2’s predictions with regards to intersec-tional characteristics are highly stereotypical – yet they are closely aligned to the US population data. We show the predicted valuesfor gender intersected with ethnicity along with the [Mean-Squared Errors] and annotate example jobs for the gender-only predictions. For details, please look at the paper.

Code

We include the following python scripts:

  • Generate_Freq_Matrices.py: Generates frequency matrices for occupations associated for each gender-intersection and name generated by the prefix templates
  • Generate_Figures.py: Generates the following figures: distribution plots of occupations, Gini indices, Lorenz curves, gender parity bar and scatter plots
  • Logistic_Regression.py: Performs logistic regression and generates heatmaps of regression coefficients
  • Comparison_to_RealWorld.py: Compares occupational distribution generated by GPT-2 with real-world US Labor Bureau data

Data

We include the following CSV files and folder structure:

.
├── scripts
├── regression_results
├── data
│   ├── XLNET
|     |-- NER_output
|           |-- names_occupations_template.csv
|           |-- identity_occupations_template.csv
│   ├── GPT-2
|     |-- NER_output
|           |-- names_occupations_template.csv
|           |-- identity_occupations_template.csv
|     |-- US_data
|           |-- gpt_vs_us_data.csv
|           |-- us_rowwise_data.csv
│   ├── shared_data
|     |-- job_replacements.csv
|     |-- names_to_continent.csv
            
└── ...

Where:

  • names_occupations_template.csv: a list of occupations associated by (GPT-2/XLNet), extracted by Stanford’s NER tool, for geographic-specific namesf
  • identity_occupations_template.csv: a list of occupations associated by (GPT-2/XLNet), extracted by Stanford’s NER tool, or ‘man’/’woman’ intersected with the categories for ethnicity, sexuality, political affiliation, and religion
  • job_replacements.csv: renames duplicate occupations (i.e. nurse and nurse practitioner)
  • us_rowwise_data.csv: percentage workers in each of the most granular level of occupation as listed in the US Labor Bureau data, for each gender and ethnicity pair (Ex: Within Asian women, what percentage of them work as CEOs, lawyers, etc.?)
  • gpt_vs_us_data.csv: percentage of gender and ethnicity in each job (Ex: Within CEOs, what percentage are Asian women, Black men, etc.?)
  • names_to_continent.csv: top names chosen for each geographic region, extracted from Wikipedia

Regression Results

In the main paper, we report a summary of logistic regression results where the dependent variable is the probability a simulated individual (i.e. a gender-intersectional prompt to GPT-2 like "the asian woman works as a...") is associated with a given job token and the independent variables are that individual's gender and intersection. In this repository, we provide:

  • summary_regression_results.csv: an aggregated summary of job regressions for each intersectional category in one file. We report three metrics: (1) Pct_Signif, which is the percentage of all job regressions where the set of identity main effects and interactions are significant, (2) Gen_R2_Change, which measures the mean change in Psuedo-R2 on the addition of the main effect from the woman dummy, and (3) Int_R2_Change, which measures the mean change in Pseudo-R2 from the additional of the gender-intersection interaction terms.
  • {identity}_logistic_regressions.csv for identity = {ethnicity, sexuality, political affiliation, religion}. There is a seperate file for each intersectional category which reports the p-values of main effects and interaction terms for all job regressions. The magnitude of significant coefficients (p<0.05) is reported and non-significant coefficients are set to 0.

Dependencies

For the NER pipeline we ran Stanford CoreNLP on localhost. To setup Stanford CoreNLP we followed instructions listed here: https://stanfordnlp.github.io/CoreNLP/corenlp-server.html

Citation

If you find the code in this repo useful, please consider citing:

@inproceedings{Kirk2021bias,
      title={Bias Out-of-the-Box: An Empirical Analysis of Intersectional Occupational Biases in Popular Generative Language Models}, 
      author={Hannah Kirk and Yennie Jun and Haider Iqbal and Elias Benussi and Filippo Volpin and Frederic A. Dreyer and Aleksandar Shtedritski and Yuki M. Asano},
      year={2021},
      booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
}