From Statistical Fairness to Semantic Fairness: The DBSD Framework for Representation- Space Bias Mitigation in Large Language Models

Abstract

Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of natural language processing tasks. However, numerous studies have shown that these models continue to exhibit measurable demographic, cultural, and social biases embedded within their semantic representations. Existing fairness methodologies primarily evaluate statistical disparities in structured prediction tasks and provide only limited support for understanding or mitigating semantic bias in free-text generation. This paper introduces the Data Bias Semantic Distance (DBSD) framework, a novel representation-space methodology that interprets algorithmic bias as the semantic distance between an observed embedding and an explicitly defined normative representation. Unlike existing debiasing techniques that modify datasets, model parameters, or generated outputs, DBSD operates directly in semantic embedding space through controlled semantic alignment while explicitly preserving informational utility. The framework integrates the CrowS-Pairs benchmark corpus, the BiasBench evaluation platform, SentenceTransformer embeddings, and a complete Python implementation developed specifically for this research, enabling fully reproducible experiments. The empirical evaluation was conducted on the complete CrowS-Pairs benchmark comprising 1,508 demographic sentence pairs spanning nine bias categories. Statistical analysis demonstrates a substantial reduction in semantic bias across all evaluated demographic groups, with an overall mean bias reduction of 45.21%, while preserving 94.85% of the original semantic utility. Paired statistical tests indicate highly significant improvements (t = 263.93, p < 0.001) accompanied by an exceptionally large within-subject effect size (Cohen's [31] d = 6.80), providing strong evidence for the effectiveness and robustness of the proposed framework.

Beyond introducing a practical semantic debiasing methodology, the paper establishes a formal theoretical relationship between semantic representation alignment and conventional statistical fairness measures. The results suggest that improvements achieved
in semantic representation space can systematically propagate to downstream fairness metrics, thereby connecting semantic optimization with statistical fairness evaluation. The proposed DBSD framework contributes a new paradigm for fairness research
in generative AI by unifying benchmark-based evaluation, representation-space optimization, empirical statistical validation, and
reproducible implementation within a single coherent framework. The approach extends existing fairness methodologies beyond
structured datasets and offers a scalable foundation for future semantic fairness research in Large Language Models.