Mitigating Language-Dependent Ethnic Bias in BERT

This repository contains the code and data for the paper "Mitigating Language-Dependent Ethnic Bias in BERT" (EMNLP 2021)

Setup

Python 3.7

For externel libraries,

pip install -r requirements.txt

Evaluate

Evaluation Templates

Template-based evaluation is employed in this paper. The attributes and templates used in the experiments are available in templates/.

Templates don't give any significant clue that can guess one's ethnicity.
Attributes are mostly occupations and include social positions (e.g., immigrant, refugee)

Evaluation Configuration

Before the evaluation, check configuration.py and set the model and evaluation conditions. In addition, we can change the target ethnicities in here.

Evaluation Script

If you have any custom model trained for this task, you should the path to the trained model as argument (endswith .pt). Unless, the code will load pretrained model from the transformers hub.

export LANG=de

python score.py --language $LANG --custom_model_path /path/to/model/(optional)

casszhao / ethnic_bias