DebiasPI: Inference-time Debiasing by Prompt Iteration of a Text-to-Image Generative Model.

Citation: Sarah Bonna, Yu-Cheng Huang, Ekaterina Novozhilova, Sejin Paik, Zhengyang Shan, Michelle Yilin Feng, Ge Gao, Yonish Tayal, Rushil Kulkarni, Jialin Yu, Nupur Divekar, Deepti Ghadiyaram, Derry Wijaya, and Margrit Betke. DebiasPI: Inference-time Debiasing by Prompt Iteration of a Text-to-Image Generative Model. Proceedings of the European Conference on Computer Vision 2024 Workshop "Fairness and ethics towards transparent AI: facing the chalLEnge through model Debiasing" (FAILED), Milano, Italy, September 29, 2024. 16 pages.

Abstract:

Ethical intervention prompting has emerged as a tool to counter demographic biases of text-to-image generative AI models. Existing solutions either require to retrain the model or struggle t generatee images that reflect desired distributions on gender and race. We propose an inference-time process called DebiasPI for Debiasing-by-PromptIteration that provides prompt intervention by enabling the user to control the distributions of individuals’ demographic attributes in image generation. DebiasPI keeps track of which attributes have been generated either by probing the internal state of the model or by using external attribute classifiers. Its control loop guides the text-to-image model to select not yet sufficiently represented attributes, With DebiasPI, we were able to create images with equal representations of race and gender that visualize challenging concepts of news headlines. We also experimented with the attributes age, body type, profession, and skin tone, and measured how attributes change when our intervention prompt targets the distribution of an unrelated attribute type. We found, for example, if the text-to-image model is asked to balance racial representation, gender representation improves but the skin tone becomes less diverse. Attempts to cover a wide range of skin colors with various intervention prompts showed that the model struggles to generate the palest skin tones. We conducted various ablation studies, in which we removed DebiasPI’s attribute control, that reveal the model’s propensity to generate young, male characters. It sometimes visualized career success by generating two-panel images with a pre-success dark-skinned person becoming light-skinned with success, or switching gender from pre-success female to post-success male, thus further motivating ethical intervention prompting with DebiasPI.

Codebook used to evaluate the baseline prompt experiment

Related Papers by BU's AIEM Team:

Lei Guo, Chao Su, Sejin Paik, Vibhu Bhatia, Vidya Prasad Akavoor, Ge Gao, Margrit Betke and Derry Wijaya. Proposing an open-sourced tool for computational framing analysis of multilingual data. Digital Journalism, 11(2): 276-297. 2023. pdf. Open Framing Code

Ge Gao, Sejin Paik, Carley Reardon, Yanling Zhao, Lei Guo, Prakash Ishwar, Margrit Betke, and Derry Wijaya. Prediction of People's Emotional Response towards Multi-modal News. The 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (AACL-IJCNLP), November 2022. 10 pages.

Carley Reardon, Sejin Paik, Ge Gao, Meet Parekh, Yanling Zhao, Lei Guo, Margrit Betke, and Derry Wijaya. NEmo: An Affective Dataset of Gun Violence News. Proceedings of the 13th Conference on Language Resources and Evaluation ( LREC 2022) Palais du Pharo, Marseille, France, June 20-25, 2022, pages 2507-2516, pdf.

Sha Lai, Yanru Jiang, Lei Guo, Margrit Betke, Prakash Ishwar, and Derry Tanti Wijaya. An Unsupervised Approach to Discover Media Frames. Proceedings of The LREC 2022 Workshop on Natural Language Processing for Political Science, Marseille, France, June 20-25, 2022, ISBN: 979-10-95546-88-7. Pages 22-31 pdf.

Isidora Tourni, Lei Guo, Taufiq Husad Daryanto, Fabian Zhafransyah, Edward Edberg Halim, Mona Jalal, Boqi Chen, Sha Lai, Hengchang Hu, Margrit Betke, Prakash Ishwar, and Derry Tanti Wijaya. Detecting Frames in News Headlines and Lead Images in U.S. Gun Violence Coverage. In Findings of the Assocation for Computational Linguistics: 2021 Conference on Empirical Methods in Natural Language Processing. November 2021, pages 4037-4050, Punta Cana, Dominican Republic. DOI 10.18653/v1/2021.findings-emnlp.339

Lei Guo, Kate Mays, Yiyan Zhang, Derry Wijaya, and Margrit Betke. What makes gun violence a (less) prominent issue? A computational analysis of compelling arguments and selective agenda setting. Mass Communication & Society:24(5):651-675, April 2021. https://doi.org/10.1080/15205436.2021.1898644

Vibhu Bhatia, Vidya Prasad Akavoor, Sejin Paik, Lei Guo, Mona Jalal, Alyssa Smith, David Assefa Tofu, Edward Edberg Halim, Yimeng Sun, Margrit Betke, Prakash Ishwar and Derry Tanti Wijaya. OpenFraming: Open-sourced Tool for Computational Framing Analysis of Multilingual Data. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 242-250. Online and Punta Cana, Dominican Republic. November 2021. DOI 10.18653/v1/2021.emnlp-demo.28

Afra Feyza Akyürek, Lei Guo, Randa Elanwar, Margrit Betke, Prakash Ishwar, and Derry T. Wijaya. Multi-label and Multilingual News Framing Analysis Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), July 2020, pages 8614-8624.

L. Guo, K. Mays, S. Lai, M. Jalal, P. Ishwar, and M. Betke. Accurate, Fast, But Not Always Cheap: Evaluating "Crowdcoding" as an Alternative Approach to Analyze Social Media Data. Journalism & Mass Communication Quarterly, 97(3): 811-834. September 2020. https://doi.org/10.1177/1077699019891437

Siyi Liu, Lei Guo, Kate Mays, Margrit Betke, and Derry Wijaya. Detecting Frames in News Headlines and Its Application to Analyzing News Framing Trends Surrounding U.S. Gun Violence. Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), Hong Kong, November 2019, 504-514. Gun Violence Frame Corpus (GVFC) dataset