Abstract
Numerous works have analyzed biases in vision and pre-trained language models individually - however, less attention has been paid to how these biases interact in multimodal settings. This work extends text-based bias analysis methods to investigate multimodal language models, and analyzes intra- and inter-modality associations and biases learned by these models. Specifically, we demonstrate that VL-BERT (Su et al., 2020) exhibits gender biases, often preferring to reinforce a stereotype over faithfully describing the visual scene. We demonstrate these findings on a controlled case-study and extend them for a larger set of stereotypically gendered entities.- Anthology ID:
- 2022.gebnlp-1.10
- Volume:
- Proceedings of the 4th Workshop on Gender Bias in Natural Language Processing (GeBNLP)
- Month:
- July
- Year:
- 2022
- Address:
- Seattle, Washington
- Editors:
- Christian Hardmeier, Christine Basta, Marta R. Costa-jussà, Gabriel Stanovsky, Hila Gonen
- Venue:
- GeBNLP
- SIG:
- Publisher:
- Association for Computational Linguistics
- Note:
- Pages:
- 77–85
- Language:
- URL:
- https://aclanthology.org/2022.gebnlp-1.10
- DOI:
- 10.18653/v1/2022.gebnlp-1.10
- Cite (ACL):
- Tejas Srinivasan and Yonatan Bisk. 2022. Worst of Both Worlds: Biases Compound in Pre-trained Vision-and-Language Models. In Proceedings of the 4th Workshop on Gender Bias in Natural Language Processing (GeBNLP), pages 77–85, Seattle, Washington. Association for Computational Linguistics.
- Cite (Informal):
- Worst of Both Worlds: Biases Compound in Pre-trained Vision-and-Language Models (Srinivasan & Bisk, GeBNLP 2022)
- PDF:
- https://preview.aclanthology.org/ingest-acl-2023-videos/2022.gebnlp-1.10.pdf