Date of Award

12-9-2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Information Science

First Advisor

Mariofanna Milanova

Abstract

Artificial intelligence (AI) is rapidly transforming industries and markets, from healthcare to entertainment, revolutionizing decision-making processes. However, as AI grow more influential, they also risk amplifying existing biases, potentially leading to harmful consequences. Recent advancements in large language models (LLMs), such as GPT-4 and Llama, have heightened concerns about bias in natural language processing (NLP) tasks, driving the need for robust methods to detect and mitigate bias. Current approaches, such as the Word Embedding Association Test (WEAT) and its sentence-level extension the Sentence Encoder Association Test (SEAT) often fall short in capturing the nuances of biases in the input embeddings of AI language models, relying on predefined word groups and single effect sizes. This dissertation introduces SD-WEAT, a novel method that leverages the standard deviation (SD) of effect sizes across multiple permutations of WEAT benchmarks to provide a more comprehensive and reliable measure of bias in AI language models. SD-WEAT addresses key limitations of existing measures by allowing for non-binary, multi-class bias analysis and incorporating a negative control phase that distinguishes genuine bias from random variation or noise. Applying SD-WEAT to healthcare applications revealed significant biases in the biomedicine-focused BioBERT compared to general-purpose models, highlighting disparities related to sex and ethnicity among medical conditions. Additionally, experiments with the debiasing technique Auto-Debias demonstrated potential for mitigating healthcare-related biases, but underscored the challenges that still remain in manipulating biases in language models. Overall, SD-WEAT advances the field by offering a more robust and versatile tool for assessing bias in the input embeddings of language models, contributing to the development of socially responsible AI and emphasizing the ongoing need for innovation in this critical area.

Download

Share

COinS