Measuring Gender Bias in German Language Generation
dc.contributor.author | Kraft,Angelie | |
dc.contributor.author | Zorn,Hans-Peter | |
dc.contributor.author | Fecht,Pascal | |
dc.contributor.author | Simon,Judith | |
dc.contributor.author | Biemann,Chris | |
dc.contributor.author | Usbeck,Ricardo | |
dc.contributor.editor | Demmler, Daniel | |
dc.contributor.editor | Krupka, Daniel | |
dc.contributor.editor | Federrath, Hannes | |
dc.date.accessioned | 2022-09-28T17:10:03Z | |
dc.date.available | 2022-09-28T17:10:03Z | |
dc.date.issued | 2022 | |
dc.description.abstract | 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. | en |
dc.identifier.doi | 10.18420/inf2022_108 | |
dc.identifier.isbn | 978-3-88579-720-3 | |
dc.identifier.pissn | 1617-5468 | |
dc.identifier.uri | https://dl.gi.de/handle/20.500.12116/39481 | |
dc.language.iso | en | |
dc.publisher | Gesellschaft für Informatik, Bonn | |
dc.relation.ispartof | INFORMATIK 2022 | |
dc.relation.ispartofseries | Lecture Notes in Informatics (LNI) - Proceedings, Volume P-326 | |
dc.subject | gender bias | |
dc.subject | stereotypes | |
dc.subject | regard | |
dc.subject | natural language generation | |
dc.subject | gpt-2 | |
dc.subject | gpt-3 | |
dc.subject | german | |
dc.title | Measuring Gender Bias in German Language Generation | en |
gi.citation.endPage | 1274 | |
gi.citation.startPage | 1257 | |
gi.conference.date | 26.-30. September 2022 | |
gi.conference.location | Hamburg | |
gi.conference.sessiontitle | Trustworthy AI in Science and Society |
Dateien
Originalbündel
1 - 1 von 1
Vorschaubild nicht verfügbar
- Name:
- trustai_03.pdf
- Größe:
- 356.66 KB
- Format:
- Adobe Portable Document Format