A multi-LLM slot filling pipeline for real-time bias reduction in textual content

Volume 1, Issue 1 (Jun-Aug 2025)

Authors: Dheeraj Arremsetty

DOI: doi.org/10.64848/WJST.1.1.2025.28-35

Page No: 28-35

Keywords: Bias mitigation, Large language models, Slot filling, Fairness, Real-time processing

Download PDF

Abstract

Large language models (LLMs) frequently perpetuate explicit biases, including gender, racial, cultural, age, and socioeconomic stereotypes, which undermine fairness in critical applications such as chatbots, hiring tools, and educational platforms. These biases, rooted in training data, can lead to discriminatory outputs, eroding trust and equity in automated systems. This study introduces a novel multi-LLM slot filling pipeline designed for real-time bias mitigation, offering a scalable and modular solution to enhance fairness in textual content. The pipeline employs sequence labeling, powered by lightweight models like DistilBERT, to identify bias-sensitive tokens (e.g., gendered or racial terms) and constrained generation, using T5-small, to replace them with neutral alternatives, ensuring semantic coherence. Evaluated on seven manually crafted texts mimicking LLaMA-3-70B-Instruct outputs, the pipeline demonstrates robust bias neutralization across diverse scenarios, achieving a latency of under 100ms, suitable for dynamic, real-time applications. Comprehensive analysis, supported by visualizations, highlights the pipeline’s effectiveness in reducing bias scores while maintaining text quality, validated through fairness classifiers and human evaluations. This work provides an expanded theoretical framework, detailed methodology, and extensive related work, positioning the pipeline as a significant advancement in equitable AI. By addressing gaps in efficiency and scalability, it contributes to ethical AI research and practice, fostering inclusive outcomes in fairness-critical domains and paving the way for future explorations in adaptive and multimodal bias mitigation.