Supplementary material

Clone this repository right under your home directory. You need a machine with CUDA GPU(s) and Singularity installed.

We explain how to reproduce the results.

Environment

We use four virtual Python environments (ontonotes_converter, bling_fire, allennlp, and in-ex-cor). Python 3.8.12 or later is required for bling_fire, allennlp, and in-ex-cor.

ontonotes_converter

This environment is used to run ontonotes_converter.sh

This is an environment with python >= 2.7.18, < 3.

bling_fire

This environment is used to run data_cleaning/wiki_data_cleaning/preprocess_wiki_dump.py.

1. Create a virtual environment named bling_fire and activate it.
2. pip install -r bling_fire.txt

allennlp

This environment is used to run coref/evaluate_allennlp_*.sh.

1. Create a virtual environment named allennlp and activate it.
2. Move to allennlp.
3. pip install -U pip setuptools wheel
4. pip install --editable .[dev,all]
5. Move to allennlp-models.
6. ALLENNLP_VERSION_OVERRIDE='allennlp' pip install -e .
7. pip install -r dev-requirements.txt

in-ex-cor

This environment is used to run any other script.

1. Create a virtual environment named in-ex-cor and activate it.
2. pip install -r in-ex-cor.txt
3. Install PyTorch 2.0.1 or later.
4. Install TorchText >= 0.15.2, < 1.0.0.
5. Install spaCy 3.7.1 and run python -m spacy download en_core_web_lg.

Data preparation

Wikipedia

cd data_cleaning/wiki_data_cleaning
bash download_wiki_dump.sh en  # This downloads the latest English Wikipedia dump.
python preprocess_wiki_dump.py enwiki-latest-pages-articles.txt
python tokenise_corpus.py enwiki-latest-pages-articles_preprocessed.txt enwiki-latest-pages-articles_tokenized.txt
python lowercase.py
python final_preprocess.py enwiki-latest-pages-articles_tokenized_lc.txt

Twitter

Go to data_cleaning/twitter_data_cleaning_en. Download Twitter Stream 2017-04. Create stream directory, and put the content in it. Unfreeze files in the Twitter stream. Run python language.py.

Go to the directory hatespeech. Read README.txt, and follow its instruction. Run split.py

Go back to the parent directory, and run preprocess.py and generate_strict_dataset.py.

OntoNotes 5.0

Download OntoNotes 5.0, and put it as ontonotes-release-5.0. Run ontonotes_converter.sh.

Wordlists extraction

Go to data_cleaning/twitter_data_cleaning_en/hatespeech, and run extract.py. The mixed-membership demographic-language model developed by Blodgett et al. (2016) is in twitteraae. The final wordlists are in /in-ex-correlation/packages/wordlists/wordlists/__init__.py

Wordlists expansion

packages/wordlists/wordlists/expand_wordlist.py is the code to expand the wordlists. The expanded wordlists after removing odd terms are in /in-ex-correlation/packages/wordlists/wordlists/__init__.py. They are weat_race_exp, weat_gender_exp, weat_gender_twitter_exp, winobias_exp, hatespeech_gender_exp, and hatespeech_race_exp.

Dataset balancing

Run dataset_balancing.py.

Word embedding training

When training word2vec, go to directory w2v and run mkdir vectors && python train_w2v_embeddings.py. When training fastText, go to directory fasttext and run mkdir vectors && python train_ft_embeddings.py.

Attract-repel

Go to attract-repel directory, and run mkdir vectors && python attract-repel_new_pytorch.py. You may want to modify variable GPUS in the code.

Bias evaluation

Empty result directory in WEAT, RNSB, coref, and hatespeech directories.

WEAT

cd WEAT
python wefe-weat.py

RNSB

cd RNSB
python rnsb.py

WinoBias

cd coref
singularity pull docker://allennlp/models:latest
mkdir model
bash coref_train.sh
bash evaluate_allennlp.sh

You may want to modify variable GPUS in coref_train.sh.

Hate speech detection

cd hatespeech
bash detector.sh

Analysis

Empty results directory.

python format_results.py
python analyse_data.py

Scatter plots are generated in results/charts. Spearman's correlations are recorded in results/{task}_{word_emb}_{bias_modification_wordlist}_spearman.txt where

• task = coref or hatespeech,
• word_emb = w2v or ft, and
• bias_modification_wordlist = winobias, weat_gender (the merged wordlists of WEAT 6, 7, and 8), hatespeech_gender (the wordlists extracted from the HSD dataset for gender bias), weat_gender_twitter (similar to weat_gender, but the word "computation" was removed. This is used in the Twitter experiment.), hatespeech_race (the wordlists extracted from the HSD dataset for racial bias), or weat_race (the merged wordlists of WEAT 3, 4, and 5).

*_spearman.txt is in the format {intrinsic_bias_metric} v. {extrinsic_bias_metric}: {correlation}. extrinsic_bias_metric that ends with _strict is measured with the HSD dataset where tweets not containing a corresponding target word were removed.

License and Copyright Notice of Third Party Software

Refer to NOTICE.

Acknowledgement

This repository is based on Goldfarb-Tarrant et al. (2021).

Goldfarb-Tarrant, S., Marchant, R., Muñoz Sánchez, R., Pandya, M., & Lopez, A. (2021). Intrinsic Bias Metrics Do Not Correlate with Application Bias. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 1926–1940. https://doi.org/10.18653/v1/2021.acl-long.150

Acknowledgement of AI use

We use Code Llama as a coding assistant. We also used Google Bard to ask questions related to coding.

Acknowledgement of Third Party Software/Dataset

Internet Archive

The pandas development team. pandas-dev/pandas: Pandas [Computer software]. https://doi.org/10.5281/zenodo.3509134

Blodgett, S. L., Green, L., & O’Connor, B. (2016). Demographic Dialectal Variation in Social Media: A Case Study of African-American English. In J. Su, K. Duh, & X. Carreras (Eds.), Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (pp. 1119–1130). Association for Computational Linguistics. https://doi.org/10.18653/v1/D16-1120

Founta, A., Djouvas, C., Chatzakou, D., Leontiadis, I., Blackburn, J., Stringhini, G., Vakali, A., Sirivianos, M., & Kourtellis, N. (2018). Large Scale Crowdsourcing and Characterization of Twitter Abusive Behavior. Proceedings of the International AAAI Conference on Web and Social Media, 12(1). https://doi.org/10.1609/icwsm.v12i1.14991

Honnibal, M., Montani, I., Van Landeghem, S., & Boyd, A. (2020). spaCy: Industrial-strength Natural Language Processing in Python [Python]. https://doi.org/10.5281/zenodo.1212303 (Original work published 2014)

Hunter, J. D. (2007). Matplotlib: A 2D Graphics Environment. Computing in Science & Engineering, 9(3), 90–95. https://doi.org/10.1109/MCSE.2007.55

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., … Chintala, S. (2019). PyTorch: An Imperative Style, High-Performance Deep Learning Library. Advances in Neural Information Processing Systems, 32. https://proceedings.neurips.cc/paper_files/paper/2019/hash/bdbca288fee7f92f2bfa9f7012727740-Abstract.html

Řehůřek, R., & Sojka, P. (2010). Software Framework for Topic Modelling with Large Corpora. Proceedings of the LREC 2010 Workshop on New Challenges for NLP Frameworks, 45–50.