Theory-Grounded Measurement of U.S. Social Stereotypes in English Language Models

Cao, Yang Trista; Sotnikova, Anna; Daumé, Hal; Rudinger, Rachel; Zou, Linda

doi:10.18653/v1/2022.naacl-main.92

Cited by 10 publications

(9 citation statements)

References 25 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Many of the significant differences in Figure 2 did confirm our hypotheses: high-Agency words generated more men than women, as did low-Communion words, and in the case of Midjourney, low-Agency words were associated with younger age. As well, both Midjourney and Stable Diffusion showed a significant tendency to associate progres-sive Beliefs with lighter skin, primarily driven by the traits modern-traditional and science-oriented-religious, as also reported by (Cao et al 2022). In contrast to our hypotheses, lighter skin was associated with low-communion adjectives for both DALL-E and Stable Diffusion.…”

Section: Summary and Discussioncontrasting

confidence: 61%

“…The experimental design does not directly compare communion values for men and women. However, other related literature confirms many of these predictions and also reports that women are seen as more Communal (warm, friendly) than men (Fiske et al 2002;Nicolas, Bai, and Fiske 2022), and white people are seen as more modern and science-oriented (high-Beliefs) than Black people (Cao et al 2022).…”

Section: Social Stereotypesmentioning

confidence: 79%

“…List of stereotype dimensions and corresponding traits in the ABC model, adapted fromCao et al (2022).All three systems evaluated in this study, DALL-E 2, 2 Midjourney, 3 and Stable Duffusion, 4 generate original images from textual prompts and/or uploaded images. They are based on state-of-the-art image generation technology, like diffusion models(Sohl-Dickstein et al 2015;Nichol et al …”

mentioning

confidence: 99%

See 2 more Smart Citations

A Friendly Face: Do Text-to-Image Systems Rely on Stereotypes when the Input is Under-Specified?

Fraser¹,

Kiritchenko²,

Nejadgholi³

2023

Preprint

View full text Add to dashboard Cite

As text-to-image systems continue to grow in popularity with the general public, questions have arisen about bias and diversity in the generated images. Here, we investigate properties of images generated in response to prompts which are visually under-specified, but contain salient social attributes (e.g., 'a portrait of a threatening person' versus 'a portrait of a friendly person'). Grounding our work in social cognition theory, we find that in many cases, images contain similar demographic biases to those reported in the stereotype literature. However, trends are inconsistent across different models and further investigation is warranted.

show abstract

Section: Summary and Discussioncontrasting

confidence: 61%

Section: Social Stereotypesmentioning

confidence: 79%

mentioning

confidence: 99%

See 1 more Smart Citation

A Friendly Face: Do Text-to-Image Systems Rely on Stereotypes when the Input is Under-Specified?

Fraser¹,

Kiritchenko²,

Nejadgholi³

2023

Preprint

View full text Add to dashboard Cite

show abstract

“…Bias is also a key issue in current research on generative AI. There is research indicating that the content generated by generative AI may contain bias [42][43][44]. This bias is considered inappropriate and may affect the user's cognition.…”

Section: The Application Of This Work In Eliminating Bias In Generati...mentioning

confidence: 99%

Automatically Distinguishing People’s Explicit and Implicit Attitude Bias by Bridging Psychological Measurements with Sentiment Analysis on Large Corpora

Wang,

Zhang,

Xue

et al. 2024

Applied Sciences

View full text Add to dashboard Cite

Social psychological studies show that people’s explicit attitude bias in public expression can differ from their implicit attitude bias in mind. However, the current automatic attitude analysis does not distinguish between explicit and implicit attitude bias. Simulating the psychological measurements of explicit and implicit attitude bias, i.e., self-report assessment (SRA) and implicit association test (IAT), we propose an automatic language-based analysis to distinguish explicit and implicit attitude bias in a large population. By connecting the criteria of SRA and IAT with the statements containing patterns of special words, we derive explicit and implicit attitude bias with the sentiment scores of the statements, which are obtained by pre-trained machine-learning methods. Extensive experiments on four English and Chinese corpora and four pairs of concepts show that the attitude biases obtained by our method on a large population are consistent with those of traditional psychological experiments in the costly small-scale experiments. The maximum gap between the sentiment scores of explicit and implicit biases reaches 0.9329. Furthermore, we achieve new findings on the difference between the evolution of explicit and implicit attitude bias. The maximum variance gap of sentiment scores in the dynamic changes between explicit and implicit biases reaches 0.249.

show abstract

“…Demo video at https://vimeo.com/800847322/9848219f33. patterns in the training data that are not desirable such as offensiveness (Gehman et al, 2020) or unequal treatment (Cao et al, 2022), but also from failing to replicate other more desirable patterns which are also present in the data but are hard to capture by the neural network model, such as truthful information (Lin et al, 2022). For these reasons, there is a growing interest in controlling the generations to align with human values (Ouyang et al, 2022;Askell et al, 2021).…”

Section: Introductionmentioning

confidence: 99%

disco: a toolkit for Distributional Control of Generative Models

Kruszewski¹,

Rozen²,

Dymetman³

2023

Preprint

View full text Add to dashboard Cite

Pre-trained language models and other generative models have revolutionized NLP and beyond. However, these models tend to reproduce undesirable biases present in their training data. Also, they may overlook patterns that are important but challenging to capture. To address these limitations, researchers have introduced distributional control techniques. These techniques, not limited to language, allow controlling the prevalence (i.e. expectations) of any features of interest in the model's outputs. Despite their potential, the widespread adoption of these techniques has been hindered by the difficulty in adapting the complex, disconnected code. Here, we present disco, an open-source Python library that brings these techniques to the broader public. 1

show abstract

Theory-Grounded Measurement of U.S. Social Stereotypes in English Language Models

Cited by 10 publications

References 25 publications

A Friendly Face: Do Text-to-Image Systems Rely on Stereotypes when the Input is Under-Specified?

A Friendly Face: Do Text-to-Image Systems Rely on Stereotypes when the Input is Under-Specified?

Automatically Distinguishing People’s Explicit and Implicit Attitude Bias by Bridging Psychological Measurements with Sentiment Analysis on Large Corpora

disco: a toolkit for Distributional Control of Generative Models

Contact Info

Product

Resources

About