Probing Gender Bias in BERT and RoBERT
The Experiments are done on profession dataset. The jason file for this data set is in experiment_data folder
You can run the experiments by runing the run.py file. To run the experiment for the models in the following table type pyhton3 run.py -model model-name for example for roberta-base model: run.py -model roberta-base
model-name | model-name |
---|---|
bert-base-cased | bert-base-uncased |
bert-large-cased | bert-large-uncased |
bert-base-multilingual-cased | bert-base-multinlingual-uncased |
distilbert-base-cased | distilbert-base-uncased |
distilbert-base-multilingual-cased | |
roberta-base | roberta-large |
roberta-large-mnli | distilroberta-base |
After runing the run.py the results of prompt tuning are saved in "results" folder. For measuring the bias you need to run:
evaluation.py folder_name model-name
example:
python3 evaluation.py 20211206__bias_probing_roberta-base roberta-base
For visualsation have the model_name_final_results.csv in your directory and run the note book attached to this project