Multi-Modal Bias: Introducing a Framework for Stereotypical Bias Assessment beyond Gender and Race in Vision–Language Models

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1725–1735 May 2-6, 2023 ©2023 Association for Computational Linguistics Multi-Modal Bias: Introducing a Framework for Stereotypical Bias Assessment beyond Gender and Race in Vision–Language Models Sepehr Janghorbani Rutgers University New Brunswick, NJ, USA sepehr.janghorbani@rutgers.edu Gerard de Melo HPI / University of Potsdam Potsdam, Germany gdm@demelo.org Abstract Recent breakthroughs in self-supervised train- ing have led to a new class of pretrained vi- sion–language models. While there have been investigations of bias in multimodal models, they have mostly focused on gender and racial bias, giving much less attention to other rel- evant groups, such as minorities with regard to religion, nationality, sexual orientation, or disabilities. This is mainly due to lack of suit- able benchmarks for such groups. We seek to address this gap by providing a visual and textual bias benchmark called MMBias, con- sisting of around 3,800 images and phrases cov- ering 14 population subgroups. We utilize this dataset to assess bias in several prominent self- supervised multimodal models, including CLIP, ALBEF, and ViLT. Our results show that these models demonstrate meaningful bias favoring certain groups. Finally, we introduce a debi- asing method designed specifically for such large pretrained models that can be applied as a post-processing step to mitigate bias, while pre- serving the remaining accuracy of the model. 1 Introduction The recent emergence of large pretrained vision– language models has revolutionized many multi- modal tasks previously considered impractical to solve. Although architectures capable of jointly addressing computer vision and NLP tasks using a single unified model have been around for a while (Lu et al., 2019; Tan and Bansal, 2019; Li et al., 2019), recent advances in self-supervised training methods have amplified the significance and appli- cability of such models. The sheer power of these methods is highly dependent on the scale of the model and the diversity and distributional proper- ties of the dataset on which they are trained. Due to their wide range of diverse applications (Eslami et al., 2021), it is of utmost importance to be aware of the shortcomings of vision–language pretrained (VLP) models as well as their capabilities. One such limitation is that, like any other ma- chine learning system, multimodal models may be prone to exhibiting human-like stereotypical biases such as gender or race-related stereotypes (Nadeem et al., 2020; Garrido-Muñoz et al., 2021). For instance, pretrained language models have been shown to associate male-gendered phrases and sen- tences to a greater extent with certain high-paying professions and even with individual traits such as intelligence, in comparison to female-gendered phrases (Wang et al., 2021a). Similarly, it has been found that Hispanic and African American names may be tied to words representing danger and crime more often than Caucasian names (Manzini et al., 2019). Certain biases have also been identified in computer vision models as well (Wang et al., 2019). Such biases are discriminatory towards affected population groups and can be extremely harmful to society the more these …

Content truncated to 500 words