Bias Unveiled: Enhancing Fairness in German Word Embeddings with Large Language Models

Gender bias in word embedding algorithms has garnered significant attention due to its integration into machine learning systems and its potential to reinforce stereotypes. Despite ongoing efforts, the root causes of biases in training word embeddings, specifically for the German language, remain unclear. This research presents a novel approach to tackling this problem, paving the way for new avenues of investigation. Our methodology involves a comprehensive analysis of word embeddings, focusing on how training data manipulations impact resulting biases. By examining how biases originate within specific training documents, we identify subsets that can be removed to effectively mitigate these effects. Additionally, we explore both conventional methods and new approaches using large language models (LLMs) to ensure the generated text adheres to concepts of fairness. Using few-shot prompting, we generate gender bias-free text, employing GPT-4 as a benchmark to evaluate the fairness of this process for the German language. Our method explains the intricate origins of biases within word embeddings, validated through rigorous application to German Wikipedia corpora. Our findings robustly demonstrate the efficacy of our method, showing that removing certain document subsets significantly diminishes bias in word embeddings. This is further detailed in our analysis, "Unlocking the Limits: Document Removal with an Upper Bound," in the experimental results section. Ultimately, this research presents a practical framework to uncover and mitigate biases in word embedding algorithms during training. Our goal is to advance machine learning systems that prioritize fairness and impartiality by revealing and addressing latent sources of bias.