Debiasing Methods in Natural Language Understanding Make Bias More Accessible

Accepted as a conference paper for EMNLP 2021

Paper | Arxiv

Abstract: Model robustness to bias is often determined by the generalization on carefully designed out-of-distribution datasets. Recent debiasing methods in natural language understanding (NLU) improve performance on such datasets by pressuring models into making unbiased predictions. An underlying assumption behind such methods is that this also leads to the discovery of more robust features in the model’s inner representations. We propose a general probing-based framework that allows for post-hoc interpretation of biases in language models, and use an information-theoretic approach to measure the extractability of certain biases from the model's representations. We experiment with several NLU datasets and known biases, and show that, counter-intuitively, the more a language model is pushed towards a debiased regime, the more bias is actually encoded in its inner representations.

Environment setup

First, clone this repository

git clone https://github.com/technion-cs-nlp/bias-probing

Anaconda/Miniconda

From the root folder of the project run

conda env create -f environment.yml
conda activate probing

Other environments

Dependencies are listed in environment.yml for installing with pip,venv, etc,

Tested Environment

All package versions and environment settings are documented in environment.yml.

Tested on Python 3.7.4 with:

• transformers 4.5.1
• datasets 1.3.0
• pytorch 1.7.0

Compatibility tested with Cuda v10.2. Models were trained on NVIDIA GeForce RTX 2080 Ti.

Code structure

Framework

• bias_probing/ contains models, data files, and implementation of the probing framework.
• scripts/ is the entry point for running the experiments and training models.

Reproducing the Results

TBD

Citation

If you find this repository useful in your work, please cite our paper:

@inproceedings{mendelson-belinkov-2021-debiasing,
    title = "Debiasing Methods in Natural Language Understanding Make Bias More Accessible",
    author = "Mendelson, Michael  and
      Belinkov, Yonatan",
    booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2021",
    address = "Online and Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.emnlp-main.116",
    pages = "1545--1557",
    abstract = "Model robustness to bias is often determined by the generalization on carefully designed out-of-distribution datasets. Recent debiasing methods in natural language understanding (NLU) improve performance on such datasets by pressuring models into making unbiased predictions. An underlying assumption behind such methods is that this also leads to the discovery of more robust features in the model{'}s inner representations. We propose a general probing-based framework that allows for post-hoc interpretation of biases in language models, and use an information-theoretic approach to measure the extractability of certain biases from the model{'}s representations. We experiment with several NLU datasets and known biases, and show that, counter-intuitively, the more a language model is pushed towards a debiased regime, the more bias is actually encoded in its inner representations.",
}

License

This work is open-sourced under the MIT license. For more information check out LICENSE