Yixin Wan, Kai-Wei Chang

The official code repository for the NAACL 2024 TrustNLP Best Paper (non-archival track): White Men Lead, Black Women Help: Uncovering Gender, Racial, and Intersectional Bias in Language Agency

We provide the code for generating the raw dataset. Alternatively, in the folder ./lac_dataset_construction/lac_dataset, we provide the final version of the cleaned, labeled, and split LAC dataset that you can directly use for training LAC classifiers.

• To run data generation, first go to the folder for generation scripts:

cd lac_dataset_construction

• Then, add in your OpenAI account configurations in ./lac_dataset_construction/generation_util.py and run:

python generate_dataset.py

To train a BERT-based LAC classifier from scratch, first return to the main labe-agency folder and run:

sh ./scripts/run_train_bert.sh

Alternatively, download model checkpoints from this Google Drive link and store all checkpoint subfolders (e.g. bert-base-cased_binary_*_*) in the labe-agency/checkpoints/ folder.

For experiments on ChatGPT, first add in your OpenAI account configurations in generation_util.py. For generation experiments without prompt-based mitigation, run:

sh ./scripts/run_generate.sh

For generation experiments with prompt-based mitigation, run:

sh ./scripts/run_generate_mitigate.sh

For evaluation experiments on language agency gender bias in human-written texts, run:

sh ./scripts/run_infer_calc_agency_aggregated_human_bert.sh

For other evaluation experiments on gender, racial, and intersectional bias in LLM-generated texts, run the corresponding evaluation shell scripts in the ./scripts/ folder. For instance, for the evaluation of Llama3-generated texts without mitigation, run:

sh ./scripts/run_infer_calc_agency_aggregated_llama3_bert.sh