Iterative Nullspace Projection (INLP)

This repository contains the code for the experiments and algorithm from the paper "Null it out: guarding protected attributes by iterative nullspsace projection" (accepted as a long paper in ACL 2020).

To cite:

@inproceedings{DBLP:conf/acl/RavfogelEGTG20,
  author    = {Shauli Ravfogel and
               Yanai Elazar and
               Hila Gonen and
               Michael Twiton and
               Yoav Goldberg},
  editor    = {Dan Jurafsky and
               Joyce Chai and
               Natalie Schluter and
               Joel R. Tetreault},
  title     = {Null It Out: Guarding Protected Attributes by Iterative Nullspace
               Projection},
  booktitle = {Proceedings of the 58th Annual Meeting of the Association for Computational
               Linguistics, {ACL} 2020, Online, July 5-10, 2020},
  pages     = {7237--7256},
  publisher = {Association for Computational Linguistics},
  year      = {2020},
  url       = {https://www.aclweb.org/anthology/2020.acl-main.647/},
  timestamp = {Wed, 24 Jun 2020 17:15:07 +0200},
  biburl    = {https://dblp.org/rec/conf/acl/RavfogelEGTG20.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Algorithm

The implementaion of the Iterative Nullspace Projection (INLP) method is available under src/inlp-oop. Given a dataset of examples X and annotation for any kind of information Z we want to remove from X (e.g. gender, sentiment, tense) the algorithm learns a proejction matrix P which aims to exhaustively remove the ability to linearly predict Z from X ("linear guardness").

src/inlp-oop contains an implementations of the algorithm that allows learning a projection matrix for several common objectives, such as neutralizing the information captured by linear classifiers (a classification objective), or by linear siamese networks (a metric learning objective). The notebook usage_example.ipynb demonstrates the use of the algorithm for those purposes. A more bare-bone implementaton of the same algorithm for the common use case of classification (removing the ability to classify Z based on X - the focus of the paper) is avaialble under src/debias.py.

Experiments

Start a new virtual environment:

conda create -n null_space python=3.7 anaconda
conda activate null_space

Setup

download the data used for this project:

./download_data.sh

Word Embedding experiments (Section 6.1 in the paper)

python src/data/to_word2vec_format.py data/embeddings/glove.42B
.300d.txt

python src/data/filter_vecs.py \
--input-path data/embeddings/glove.42B.300d.txt \
--output-dir data/embeddings/ \
--top-k 150000  \
--keep-inherently-gendered  \
--keep-names 

And run the notebook word_vectors_debiasing.ipynb (under "notebooks")

Controlled Demographic experiments (Section 6.2 in the paper)

export PYTHONPATH=/path_to/nullspace_projection

./run_deepmoji_debiasing.sh

Bias Bios experiments (Section 6.3 in the paper)

Assumes the bias-in-bios dataset from De-Arteaga, Maria, et al. 2019 saved at data/biasbios/BIOS.pkl.

python src/data/create_dataset_biasbios.py \
        --input-path data/biasbios/BIOS.pkl \
        --output-dir data/biasbios/ \
        --vocab-size 250000

./run_bias_bios.sh

And run the notebooks biasbios_fasttext.ipynb and biasbios_bert.ipynb.

Trained Models

We release the following models and projection layers: "debiased" GloVe embeddings (glove.42B.300d.projected.txt) and "gender-neutralizing projection" (P.glove.dim=300.iters=35.npy) from Section 6.1 in the paper, and BERT-base "gender-neutralizing projection" (P.bert_base.iters=300.npy), over the biographies dataset, from Section 6.3 in the paper. glove.42B.300d.projected.txt was created by applying the transformation P.glove.dim=300.iters=35.npy over the original 300-dim GloVe embeddings. Note that the BERT projection P.bert_base.iters=300.npy is designed to remove the ability to predict gender from the CLS token, over a specific profession-prediction dataset. It is to be applied on layer 12 of BERT-base, and requires the finetuning of the subsequent linear layer.

Usage guidelines: We urge practitioners not to treat those as "gender-neutral embeddings": naturally, as a research paper, the debiasing process was guided by one relatively simple definition of gender association, and was evaluated only on certain benchmarks. As such, it is likely that various gender-related biases are still present in the vectors. Rather, we hope that this model would encourage the community to explore which kinds of biases were mitigated by our intervention -- and which were not, shedding light on thw ways by which bias is manifested.