Disability Bias Detection in Electra-based Masked Language Model

Disability bias is one of the most challenging sociodemographic bias to mitigate. To promote a more just and inclusive society, developers and researchers are strongly advised to create language models by prioritizing ethical considerations, where the benefits and opportunities of AI are accessible to all users and groups. The creation of new design guidelines and datasets are therefore essential to help AI systems realize their enormous potential for the benefit of people with disabilities. This paper presents a study on disability bias detection in Electra-Large-based Masked Language. Participatory research was conducted by involving three disability organizations, in collecting disability information for deaf and hard of hearing people. The results of our initial analysis significantly reveal the sensitivity of the model studied to the different identity references used to designate deaf and hard of hearing people.