Detecting and Mitigating Bias in LLMs through Knowledge Graph-Augmented Training

Rajeev Kumar, Harishankar Kumar, Kumari Shalini

Keywords: Bias Detection, Bias Mitigation, Large Language Models, Knowledge Graph-Augmented Training, Real-World Datasets

TL;DR: Integrating Knowledge Graph-Augmented Training with LLMs reduces biases from training data, enhances fairness using datasets like Gender Shades and FairFace, and improves overall model performance for ethical AI applications.

Abstract:

Large language models have revolutionized natural language processing with their surprising capability to understand and generate human-like text. However, most of these models often inherit and further amplify biases present in their training data, raising ethical and fairness concerns in their applications. The detection and mitigation of such biases are important to make LLMs act responsibly and equitably across diverse domains. This work investigates the integration of Knowledge Graph-Augmented Training as a novel approach to mitigate bias in LLMs. By leveraging structured and domain-specific knowledge from real-world knowledge graphs, we improve the model's understanding and contextual awareness to reduce biased outputs. The public datasets employed in the assessment of biases in gender classification include Gender Shades, for stereotype analysis in professional description Bias in Bios, and FairFace for racial and gender biases. This process involves augmenting the GPT-4 training process with the relevant knowledge graphs tuned towards these datasets and rigorous bias detection using metrics like demographic parity and equal opportunity. We also conduct several mitigation strategies that exploit this richer knowledge base to rectify such biased associations and to eventually achieve fairness in the model's predictions. The results have been very promising, showing a huge drop in biased outputs with an increase in many of the metrics for bias. The Knowledge Graph-Augmented Training approach not only mitigates existing biases but also enhances overall model performance and reliability by providing richer information about context. This research underlines the value of structured knowledge representations together with advanced language models as the basis for more ethical and unbiased AI. Equipped with real-world data sets and knowledge graphs, our framework makes for a scalable and effective system that detects and mitigates bias, paving the way towards responsible deployment in sensitive and high-stakes applications.

Submission Number: 10