Title: An Interdisciplinary Approach to Understanding and Mitigating Harms Against Marginalized Groups in Intelligent Systems

Date: Tuesday, December 3, 2024

Time:  11:30am-1:30pm(Eastern Time)

Location (in-person): Coda 9th Floor, Room C0908 Home Park

Location (virtual): https://gatech.zoom.us/j/98489741017?pwd=AcLSKTibcaMckqZFWc5AmUfMzB7nn2.1

Camille Harris

School of Interactive Computing

College of Computing

Georgia Tech

Committee

Dr. Diyi Yang (Advisor) - School of Computer Science, Stanford University

Dr. Neha Kumar (Advisor) - School of Interactive Computing, Georgia Tech

Dr. Munmun De Choudhury - School of Interactive Computing, Georgia Tech

Dr. Kartik Goyal - School of Interactive Computing, Georgia Tech

Dr. Andre Brock - School of Literature Media and Communication , Georgia Tech

Dr. Christina Harrington - Human Computer Interaction Institute, Carnegie Mellon University 

Abstract

As AI systems become increasingly adopted in new domains and a greater part of everyday life, examining, understanding, and ultimately preventing these systems from replicate existing societal harms of sexism, racism, and other oppressive systems is an increasing concern. Within a United States context, these issues pose a large potential to harm to groups historically oppressed marginalized groups such as Black/African people and Indigenous people, and other People of Color (BIPOC). Within Natural Language Processing (NLP), dialects of English spoken by minority populations such as African American English and Chicano English that have distinct grammar, vocabulary and other linguistic differences from White Mainstream English (WME) also called Standard American English (SAE). While most works that explore such biases concerns in NLP or other machine learning (ML) bias and fairness evaluations focus on the model level, most marginalized users who experience the impacts of such systems experience it at the level of the downstream application, where multiple other components including the context of application, the party creating and implementing the model, the user interface, etc. all impact the end user in addition to model performance issues alone. For this reason, my work focuses on downstream applications of intelligent systems, with an emphasis on NLP systems, and the harms of these applications towards marginalized groups in the United States. Further, my work engages traditional machine learning fairness and NLP methods to better quantify harmful outcomes against marginalized groups with these technologies, along with qualitative methodologies which provide an analysis that centers the lived experiences of marginalized groups.