Measuring Gender Bias in German Language Generation

This repository holds code, data, and a German regard classifier as described in our paper "Measuring Gender Bias in German Language Generation" in the proceedings of INFORMATIK 2022. The project is an updated version of the bias measure developed and applied in my Master's thesis (can be found in this repo).

---

Running bias evaluations

(Python 3.7 recommended)

The pretrained SentenceBERT-based regard classifier is stored in models/sbert_regard_classifier.pth. To evaluate the bias in a list of sentences:

• Classify the generated sentences with the pretrained regard classifier via python run.py run_mode=classifier classifier_mode=predict
• Then run python run.py run_mode=eval_bias

Generally: Switching between modes can be done via python run.py run_mode=MODENAME (classifier with classifier_mode set to train or evaluate the classifier, eval_bias to run a bias analysis). It is definitely recommended checking out the detailed options within the respective config files.

---

Data

Different development data and experiment artifacts are included in the data folder:

For training and evaluation of the classifier:

• GER_regard_data_annotated/combined_all.csv contains the majority-labeled and cleaned regard data (preprocessed_GER/dev_test_indices contains the randomly determined indices used for our dev and test splits)
• The crowd-sourced human-authored dataset with human annotations (used for training) can be found in: annotated_data_raw/crowd_sourced_regard_w_annotations
• The GerPT-2-generated dataset with human annotations (used for classifier evaluation) are in: annotated_data_raw/gerpt2_generated_regard_w_annotations

Jupyter notebooks used for data & annotation exploration can be found in the original thesis repo.

Data for experiments:

• classifier_bias_check was used to confirm that there was no classifier-inherent biases
• gerp2-generated and gpt3-generated contain samples and bias evaluation results with and without triggers Warning: Some samples are explicit or offensive in nature.

Citation

To cite this work:

@inproceedings{mci/Kraft2022,
author = {Kraft,Angelie AND Zorn,Hans-Peter AND Fecht,Pascal AND Simon,Judith AND Biemann,Chris AND Usbeck,Ricardo},
title = {Measuring Gender Bias in German Language Generation},
booktitle = {INFORMATIK 2022},
year = {2022},
editor = {Demmler, Daniel AND Krupka, Daniel AND Federrath, Hannes} ,
pages = { 1257-1274 } ,
doi = { 10.18420/inf2022_108 },
publisher = {Gesellschaft für Informatik, Bonn},
address = {}
}