Accepted to NAACL2022, Workshop on Gender Bias in Natural Language Processing

👉 arxiv paper 👈

We freeze weights of a pre-trained BERT and we fine-prune it on a gender debiasing loss. Optimized are only the pruning scores -- they act a gate to the BERT's weights. We utilzie block movement pruning.

Setup

conda env create -f envs/pruning-bias.yaml
conda activate debias
pip uninstall nn_pruning
pip install git https://github.com/[anonymized]/nn_pruning.git@automodel

Block pruning

python run.py --multirun \
    experiment=debias_block_pruning_frozen \
    model.embedding_layer=last,all \
    model.debias_mode=sentence,token \
    prune_block_size=32,64

Pruning enitre heads

python run.py --multirun \
    experiment=debias_head_pruning_frozen_values_only \
    model.embedding_layer=last,all \
    model.debias_mode=sentence,token

Debiasing-only:

python run.py --multirun \
    model.embedding_layer=first,last,all,intermediate \
    model.debias_mode=sentence,token

• The first run will download, process, and cache datasets.
• By default, debiasing will run on a single GPU. For more options, see configs.
  • This project uses hydra for config managements and pytorch lightning for training loops.
  • All experiments are defined in configs/experiment/
• We use run_glue.py to evaluate GLUE. To evaluate pruned models, we manually load the pruning scores state dicts.

• Block pruning:

@article{Lagunas2021BlockPF,
  title={Block Pruning For Faster Transformers},
  author={Franccois Lagunas and Ella Charlaix and Victor Sanh and Alexander M. Rush},
  journal={ArXiv},
  year={2021},
  volume={abs/2109.04838}
}

• The original debiaing idea:

@inproceedings{kaneko-bollegala-2021-context,
    title={Debiasing Pre-trained Contextualised Embeddings},
    author={Masahiro Kaneko and Danushka Bollegala},
    booktitle = {Proc. of the 16th European Chapter of the Association for Computational Linguistics (EACL)},
    year={2021}
}

• Hydra lightning template by ashleve.