DrawL: Understanding the Effects of Non-Mainstream Dialects in Prompted Image Generation

TL;DR

This paper investigates whether implicit dialect cues in prompts influence the skin-tone and gender portrayal of people in text-to-image generation. Using a contrastive prompting framework, the authors assemble 1821 MAE baseline prompts and 1821 counterfactual prompts that encode AAE syntactic features, generate four images per prompt with Stable Diffusion, and annotate skin tones with the Monk Skin Tone Scale. They find a moderate overall association (ES=$0.272$) between using AAE features and darker skin-tone outputs, with certain features (e.g., Finna $ES=$0.729$, Habitual Be $ES=$0.410$, Completive Done $ES=$0.437$) producing stronger effects. The study discusses representational and quality-of-service harms, the naturalness of such bias given large web-sourced training data, and calls for sociolinguistic analyses in model evaluation plus potential mitigation or personalization considerations.

Abstract

Text-to-image models are now easy to use and ubiquitous. However, prior work has found that they are prone to recapitulating harmful Western stereotypes. For example, requesting that a model generate an "African person and their house," may produce a person standing next to a straw hut. In this example, the word "African" is an explicit descriptor of the person that the prompt is seeking to depict. Here, we examine whether implicit markers, such as dialect, can also affect the portrayal of people in text-to-image outputs. We pair prompts in Mainstream American English with counterfactuals that express grammatical constructions found in dialects correlated with historically marginalized groups. We find that through minimal, syntax-only changes to prompts, we can systematically shift the skin tone and gender of people in the generated images. We conclude with a discussion of whether dialectic distribution shifts like this are harmful or are expected, possibly even desirable, model behavior.

Figures (9)

• Figure 1: The Monk Skin Tone Scale. The skin tones of humans generated by the model are annotated with a score of 1 (lightest) to 10 (darkest). In this work, we measure how the distribution of skin tones generated by the model changes when prompting in African American English (AAE) as opposed to Standard American English (SAE)
• Figure 2: Examples of User-Submitted Prompts and the resultant contrastive prompt pairs. We construct the dataset to be used for our analysis by choosing in-the-wild user-submitted prompts to an image generation model, and rewording these prompts into a prompt that generates humans and allows us to apply each syntactical feature with a minimal number of changes to the SAE prompt.
• Figure 3: Effect Sizes for the association between dialect, skin tone distribution, and gendered prompts. Bolded cells all have at least a moderate effect on the skin-tones generated by Stable Diffusion. In aggregate, the application of AAE has a moderate effect on the distribution of skin tones -- shifting skin tones darker.
• Figure 4: Distribution of Monk Skin Tones for images generated by our contrastive prompts in SAE and AAE. (\ref{['fig:result_distribution:all']}) Shows the marginal skin tone distributions over the gendered prompt subjects. (\ref{['fig:result_distribution:male']},\ref{['fig:result_distribution:female']},\ref{['fig:result_distribution:neutral']}) show the skin tone distribution conditioned on the prompt specifying male subjects, female subjects, and not specifying gender respectively. The marginal and conditional distributions, all show that prompting Stable Diffusion in AAE generates overall darker subjects in the image compared to prompting with SAE.
• Figure 5: Distribution of Monk Skin Tones for selected features. (\ref{['fig:comparison:finna']},\ref{['fig:comparison:completive']}) The use of 'Finna' as a semi-modal and the use of 'Completive Done' have a relatively strong effects on the distribution of skin tones -- darkening the skin tones of generated humans. (\ref{['fig:comparison:quotative']},\ref{['fig:comparison:aint']}) The use of the 'Quotative All' and the use of 'Ain't' as the negated form of 'be' have little effect on the distribution of skin tones.