Bias and Mitigation in Large Language Models: Addressing Inequalities and Promoting Ethical AI Development
ID:128
View protection:Participant Only
Updated time:2025-12-23 13:12:30 Views:93
Online
Abstract
Large Language Models (LLMs) are used extensively in natural language processing but are biased and hence yield unfair output. In this paper, the bias present in four prominent models—BERT, XLNet, RoBERTa, and ALBERT—is examined using the Crows-Pairs dataset, which is employed for identifying biased language patterns. The paper discusses the working of the models and the type of their biases. Bias mitigation methods address various factors of bias through techniques such as Counterfactual Data Augmentation (CDA), Adversarial Debiasing, and the AI Fairness 360 Toolkit (AIF360), which aim to ensure fairness in AI systems. It aims to create more balanced and dependable AI systems, which will lead to the development of ethical and unbiased language models. The project also aims to showcase how training data, model architecture, and interventions outside of the model can be employed to prevent bias. By outlining how to avoid and identify bias, the article sets the stage for the potential future development of responsible AI. Through research, it has been determined that there is a need to continually probe and improve in light of changing needs, to develop equitable AI for all applications.
Keywords
Large Language Models (LLMs), Bias, Mitigation, BERT, RoBERTa, XLNet, ALBERT, Stereotypical, anti-stereotypical.
Post comments