Society and Bias: Uncovering Automated Prejudices in Sociotechnical Natural Language Processing Systems
Guardat en:
| Publicat a: | ProQuest Dissertations and Theses (2025) |
|---|---|
| Autor principal: | |
| Publicat: |
ProQuest Dissertations & Theses
|
| Matèries: | |
| Accés en línia: | Citation/Abstract Full Text - PDF |
| Etiquetes: |
Sense etiquetes, Sigues el primer a etiquetar aquest registre!
|
| Resum: | The applications of AI (Artificial Intelligence) have rapidly expanded in the past decade, finding their way into diverse sectors such as finance and healthcare, and evolving into complex sociotechnical systems. These systems play a crucial role in shaping both the social and technical elements with which they interact. However, as AI deployment grows across these domains, these systems frequently exhibit human-like biases that undermine their effectiveness, particularly in intricate sociotechnical applications. While ethical inquiries in NLP (Natural Language Processing) research have emphasized specific sociodemographic aspects, such as race and gender, they often approach these issues through a predominantly technical lens. To address this limitation, further research must focus on developing comprehensive methods that encompass all sociodemographic groups, thereby promoting a more holistic understanding of NLP’s societal implications.In this thesis, therefore, I examine the biases present in human-language technologies through an interdisciplinary lens using a mixed-method approach. The investigation focuses on three distinct facets, each addressing a specific aspect of the broader problem. The first facet, Facet MODEL, investigates sociodemographic biases related to disability and nationality across various NLP frameworks, examining them primarily through a technical perspective. This analysis centers on three key NLP technologies: sentiment and opinion mining models, word embedding models, and large language models. To quantify and understand the impact of these biases and their potential harmful effects, I employ automatic indicators such as sentiment analysis scores and word vector distances. Additionally, this facet explores potential strategies for mitigating these identified biases.The second facet, Facet GAP, examines the gap between AI researchers and society regarding ‘bias’, ‘emotions’, ‘sentiment’, and ‘hallucination’. These constructs, deeply rooted in societal contexts, are widely employed in natural language processing to develop models intended for social use. However, the AI field often redefines these terms for its own purposes, disregarding the original structures and definitions from which they emerged. This gap has created a lack of shared understanding between AI and societal perspectives, leading to potential biases and harms that disproportionately affect certain groups. Through an interdisciplinary approach grounded in social informatics, philosophy, and AI, I investigate these conceptual differences, emphasizing the risks posed by this misalignment. The analysis reveal how these discrepancies contribute to biases in NLP applications deployed in social contexts. The outcome of this facet leads with the creation of an ethics sheet which can then be used to create socially sensitive technologies, that adhere to both social and technological principles of a shared construct.Building on the need to address gaps in sociotechnical systems, the third facet, Facet SOCIAL, examines the interplay between human and AI actors through the lens of actor-network theory (ANT). I propose a framework designed to foster collaborative development networks involving developers, practitioners, and other key stakeholders to create inclusive and holistic solutions that avoid discriminating against specific populations. This facet adopts a social perspective to investigate the societal impact of bias and its manifestations. To achieve this, I conduct two human subject studies aimed at identifying biases through the perspectives of social actors and establishing improved methods for evaluating such biases. The findings provide insights into how AI models influence society, uncovering the harms caused by biases and offering strategies for defining, understanding, and mitigating these issues. This facet demonstrates how biases in a language model, when left unaddressed, can evolve into systemic ‘harms’ within sociotechnical systems, perpetuating inequities and reinforcing negative dynamics.This research makes contributions to both AI and NLP fields through its interdisciplinary investigation of social biases in human language technologies. Facet MODEL advances the field by demonstrating systematic approaches to measuring biases across a broader spectrum of societal groups, moving beyond the traditional focus on gender and race to encompass previously understudied demographic categories. Building on these technical insights, Facet GAP introduces a comprehensive ethics sheet that emphasizes the critical importance of developing language models with careful attention to their sociotechnical implications, thereby promoting more context-aware development practices. Facet SOCIAL complements these contributions by providing researchers, practitioners, and developers with concrete methodologies to evaluate the holistic impact of AI technologies within social contexts. Together, these three facets create an integrated resource for identifying, understanding, and mitigating social biases across diverse AI architectures and applications, ultimately advancing the development of more equitable human language technologies that better serve all communities of society. |
|---|---|
| ISBN: | 9798290650623 |
| Font: | ProQuest Dissertations & Theses Global |