ECML Workshop 2023 submission: How Different is Stereotypical Bias across Languages?

This repository contains the codebase to measure stereotypical bias in pretrained monolingual and multilingual models for different languages, as well as the code to replicate our results. It allows to create one's own dataset in the language of interest, and then obtain the bias evaluation for that language on your preferred pre-trained model. The repository assists in conducting an analysis for various pretrained models of different architectures, i.e. encoder-based (e.g. BERT), decoder-based (e.g. GPT-2), and encoder-decoder based (e.g. T5).

Install

git clone <this repository> (original path deleted for now (due to anonymization)
cd stereotypes-multi
python -m pip install -r requirements.txt

Experiments

There are 3 main steps:

1. Creating the data set

One can create a custom data set in the language that one is interested in. We provided the English, German, French, Spanish and Turkish versions in the respective folders (create_dataset/data/intersentence and create_dataset/data/intrasentence). Hence, for these languages Step 2 can be directly applied. The pickled data files, provided in version 1.0 and 2.0, are the versions just after the translation without any pre-processing. Therefore, these should only be use if a different pre-processing than ours is intended. The main English dataset is directly copied from the StereoSet repo and is located in create_dataset/data/dev.json. It can be used for translations in custom languages.

To create a data set in a new language and for more details please follow the instructions in the create_dataset folder

2. Acquiring probabilities for candidate words and sentences

After creating the dataset in the pandas format by following Step 1, the probabilities for each candidate can be obtained. The results of this step (predictions) will be saved in the results/predictions folder. We provide all the prediction files for the languages we examined, therefore, you can skip to Step 3 if you want to work on the same models / languages.

• Model Types: The tests are carried out with BERT, GPT-2 and T5 models. If you need direct support for more models, please get in touch by opening an issue. A list of the specific models for each langauge can be found in the Appendix of our paper.

• Complexity: Utilizing GPU, inferring in batch mode, using multiprocessing is possible for each of the models by configuring the hyperparameters accordingly. Using eight CPU cores, one test will take (on average) around 20 minutes per model. On a Tesla V100-SXM2-16GB, using maximum batch size possible, it takes only a few minutes (depending on the model).

Fine-tuning Models for Next Sentence Prediction:

Since GPT and T5 models were not pre-trained with Next Sentence Prediction objective (for intersentence tests), we inferred the probabilities using their generative structure. In addition, we also fine-tuned these models for NSP downstream task to make them comparable with BERT models. To Fine-tune models for Next Sentence Prediction (NSP) please follow the instructions in the code/nsp_training folder. For this, you don't need to run the Step 1.

To obtain the probability predictions for each candidate sentence in the dataset and for more details please follow the instructions in the code folder.

3. Evaluting the Results and Computing Bias

After completing the Step 2 and obtaining the predictions file, one can measure the stereotype score, language modelling score, ICAT score and many more by running the code/evaluation.py.

To obtain the final scores and for more details about it please follow the instructions in the code folder.

Acknowledgements

This repository makes use of code from the following repositories:

• StereoSet: Measuring Stereotypical Bias in Pre-trained Language Models