Measuring Gender Bias in German Language Generation

Research output: Contributions to collected editions/worksArticle in conference proceedingsResearchpeer-review

Authors

Most existing methods to measure social bias in natural language generation are specified for English language models. In this work, we developed a German regard classifier based on a newly crowd-sourced dataset. Our model meets the test set accuracy of the original English version. With the classifier, we measured binary gender bias in two large language models. The results indicate a positive bias toward female subjects for a German version of GPT-2 and similar tendencies for GPT-3. Yet, upon qualitative analysis, we found that positive regard partly corresponds to sexist stereotypes. Our findings suggest that the regard classifier should not be used as a single measure but, instead, combined with more qualitative analyses.

Original languageEnglish
Title of host publicationINFORMATIK 2022 - Informatik in den Naturwissenschaften
EditorsDaniel Demmler, Daniel Krupka, Hannes Federrath
Number of pages18
Place of PublicationBonn
PublisherGesellschaft für Informatik e.V.
Publication date2022
Pages1257-1274
ISBN (electronic)978-3-88579-720-3
DOIs
Publication statusPublished - 2022
Externally publishedYes
Event52. Jahrestagung der Gesellschaft für Informatik - INFORMATIK 2022: Informatik in den Naturwissenschaften - UHH Gebäude in der Edmund-Siemers-Allee 1, Hamburg, Germany
Duration: 26.09.202230.09.2022
Conference number: 52
https://informatik2022.gi.de/

Bibliographical note

This work presents and extends the results of Angelie Kraft's Master thesis at Universität Hamburg and inovex GmbH. Regarding any additional research and experimentation, we acknowledge the financial support from the Federal Ministry for Economic Affairs and Energy of Germany in the project CoyPu (project number 01MK21007[G]).

Publisher Copyright:
© 2022 Gesellschaft fur Informatik (GI). All rights reserved.

DOI