The Need For Inclusive NLP: Addressing Sociodemographic Bias and Enhancing Sociotechnical Systems through Interdisciplinary Frameworks

Abstract

Natural Language Processing systems are increasingly integrated into diverse sociotechnical contexts, playing a pivotal role in essential societal functions. These systems are employed across various domains, such as education, healthcare, and policy-making, offering technical solutions to social issues. However, a critical concern is the opaque nature of these models, often presented and utilized as 'black boxes'. Proprietary restrictions or knowledge gaps obscure the underlying interactions, making them challenging to interpret. The widespread presence of these systems necessitates a thorough examination of their inherent biases and societal implications. Despite a growing body of research on bias in NLP, significant challenges remain. Predominantly, studies focus on race and gender biases, adopt a model-centric approach to analysis, and employ technocentric methods for bias mitigation, often neglecting the broader societal ramifications. Addressing these challenges requires a more comprehensive understanding of how NLP systems, as integral components of broader sociotechnical systems, impact society. This gap underscores the need for a deeper understanding of the societal implications of NLP systems. This research aims to address the complexities posed by sociodemographic biases in NLP systems and the resultant harms within these frameworks. The goal is to develop more inclusive methodologies for identifying, quantifying, and mitigating these biases, bridging the interdisciplinary divide between technical and social sciences, and proposing socially aware frameworks that address the impact of biases in these systems.