Fairness and Bias Detection in Large Language Models: Assessing and Mitigating Unwanted Biases

EasyChair Preprint 12283

Abstract

This paper examines the critical task of fairness and bias detection within LLMs, focusing on the assessment and mitigation of unwanted biases. The pervasive nature of biases in language data and their impact on downstream tasks is outlined. Existing methodologies for detecting biases in LLMs, encompassing quantitative metrics and qualitative analyses, are surveyed. Challenges associated with bias mitigation techniques, including data preprocessing, model fine-tuning, and post-processing, are scrutinized. Ethical considerations surrounding bias detection and mitigation are also investigated, emphasizing the importance of transparency and accountability in algorithmic decision-making systems. Finally, future research directions are proposed to foster fairer and more inclusive LLMs, emphasizing interdisciplinary collaboration and community engagement.

Keyphrases: language, large, models

@booklet{EasyChair:12283,
  author    = {Kurez Oroy and Adam Nick},
  title     = {Fairness and Bias Detection in Large Language Models: Assessing and Mitigating Unwanted Biases},
  howpublished = {EasyChair Preprint 12283},
  year      = {EasyChair, 2024}}