Robust Evaluation Measures for Evaluating Social Biases in Masked Language Models

Code for the paper Robust Evaluation Measures for Evaluating Social Biases in Masked Language Models (accepted by AAAI 2024)

🚴Prepare Model

Download the model from huggingface with the following script:

mkdir models
cd models/
git clone https://huggingface.co/albert-base-v2
git clone https://huggingface.co/bert-base-cased
git clone https://huggingface.co/roberta-large

💻Prepare Datasets

Download CrowS-Pairs (CP) and StereoSet (SS) datasets using the following script:

mkdir data
wget -O data/cp.csv https://raw.githubusercontent.com/nyu-mll/crows-pairs/master/data/crows_pairs_anonymized.csv
wget -O data/ss.json https://raw.githubusercontent.com/moinnadeem/StereoSet/master/data/dev.json

🧘Preprocessing

The original data is already in the data folder, if not, please download it in CrowS-Pairs (CP) and StereoSet (SS)

Then, preprocess the data with the following script:

cd code/
python preprocessing.py --input stereoset --output ../data/paralled_ss.json
python preprocessing.py --input crows_pairs --output ../data/paralled_cp.json

We refer to the method of Kaneko et al. to preprocess the data

💇‍♂️Data Sampling

Use the following script to sample the data, the sampling ratio is 30%, 40%, 50%, 60%, 70% and 80%:

cd code/
python sampling.py --sample_rate [sample_rate]

You can set [sample_rate] to 0.8 for 80% sampling.

🎯Evaluation

Use the following script to get the PLL score of MLMs:

cd code/
python evaluation.py --data [ss, cp] --output ../result/output/ --model [bert-base-cased, roberta-large, albert-large-v2] --sample_rate [sample_rate] --method [aul, cps, sss, gms]

For example, if you execute the following script, you will get result/output/ss_gms_bert-base-cased.json to record the PLL score.

python evaluation.py --data ss --output ../result/output/ --model bert-base-cased --sample_rate 1 --method gms

If you set [sample_rate] to 0.8, the file name will be result/output/0.8_ss_gms_bert-base-cased.json

📄Scoring

Use the following script to score the MLM with the PLL score:

cd code/
python scoring.py --data [ss, cp] --output ../result/output/ --model [bert-base-cased, roberta-large, albert-large-v2] --sample_rate [sample_rate] --method [aul, sss, cps, kls, jss]

For example, if you execute the following script, you will get the result/scoring/ss_kls_bert-base-cased.txt record bias score.

python scoring.py --data ss --output ../result/output/ --model bert-base-cased --sample_rate 1 --method kls

Similarly, if you set [sample_rate] to 0.8, the file name will be result/scoring/0.8_ss_kls_bert-base-cased.json

If this work has helped you in any way, please cite it by the following:

@article{liu2024robust,
    title = {Robust Evaluation Measures for Evaluating Social Biases in Masked Language Models},
    author = {Yang Liu},
    journal = {arXiv preprint arXiv:2401.11601},
    year = {2024},
    doi = {10.48550/arXiv.2401.11601}
}