Gender and Representation Bias in GPT-3 Generated Stories

Li Lucy, David Bamman


Abstract
Using topic modeling and lexicon-based word similarity, we find that stories generated by GPT-3 exhibit many known gender stereotypes. Generated stories depict different topics and descriptions depending on GPT-3’s perceived gender of the character in a prompt, with feminine characters more likely to be associated with family and appearance, and described as less powerful than masculine characters, even when associated with high power verbs in a prompt. Our study raises questions on how one can avoid unintended social biases when using large language models for storytelling.
Anthology ID:
2021.nuse-1.5
Volume:
Proceedings of the Third Workshop on Narrative Understanding
Month:
June
Year:
2021
Address:
Virtual
Editors:
Nader Akoury, Faeze Brahman, Snigdha Chaturvedi, Elizabeth Clark, Mohit Iyyer, Lara J. Martin
Venues:
NUSE | WNU
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
48–55
Language:
URL:
https://aclanthology.org/2021.nuse-1.5
DOI:
10.18653/v1/2021.nuse-1.5
Bibkey:
Cite (ACL):
Li Lucy and David Bamman. 2021. Gender and Representation Bias in GPT-3 Generated Stories. In Proceedings of the Third Workshop on Narrative Understanding, pages 48–55, Virtual. Association for Computational Linguistics.
Cite (Informal):
Gender and Representation Bias in GPT-3 Generated Stories (Lucy & Bamman, NUSE-WNU 2021)
Copy Citation:
PDF:
https://preview.aclanthology.org/nschneid-patch-3/2021.nuse-1.5.pdf
Code
 lucy3/gpt3_gender