Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Controlling Gender Bias in Retrieval via a Backpack Architecture

Amirabbas Afzali, Amirreza Velae, Iman Ahmadi, Mohammad Aliannejadi

TL;DR

The paper tackles gender bias in retrieval and ranking systems that leverage large language models. It introduces a bias-controllable ranking framework built on Backpack Language Models that reweights sense vectors during inference to control a targeted social attribute, enabling in situ fairness adjustments. The key contributions are the Sense-Attribute Alignment procedure and a per-sense reweighting mechanism governed by a fairness parameter $\lambda$, which together reduce gender bias while preserving ranking performance on MS MARCO. This approach offers a practical, deployment-friendly method for fair information retrieval without retraining and suggests broad applicability to other biases and larger models.

Abstract

The presence of social biases in large language models (LLMs) has become a significant concern in AI research. These biases, often embedded in training data, can perpetuate harmful stereotypes and distort decision-making processes. When LLMs are integrated into ranking systems, they can propagate these biases, leading to unfair outcomes in critical applications such as search engines and recommendation systems. Backpack Language Models, unlike traditional transformer-based models that treat text sequences as monolithic structures, generate outputs as weighted combinations of non-contextual, learned word aspects, also known as senses. Leveraging this architecture, we propose a framework for debiasing ranking tasks. Our experimental results show that this framework effectively mitigates gender bias in text retrieval and ranking with minimal degradation in performance.

Controlling Gender Bias in Retrieval via a Backpack Architecture

TL;DR

The paper tackles gender bias in retrieval and ranking systems that leverage large language models. It introduces a bias-controllable ranking framework built on Backpack Language Models that reweights sense vectors during inference to control a targeted social attribute, enabling in situ fairness adjustments. The key contributions are the Sense-Attribute Alignment procedure and a per-sense reweighting mechanism governed by a fairness parameter , which together reduce gender bias while preserving ranking performance on MS MARCO. This approach offers a practical, deployment-friendly method for fair information retrieval without retraining and suggests broad applicability to other biases and larger models.

Abstract

The presence of social biases in large language models (LLMs) has become a significant concern in AI research. These biases, often embedded in training data, can perpetuate harmful stereotypes and distort decision-making processes. When LLMs are integrated into ranking systems, they can propagate these biases, leading to unfair outcomes in critical applications such as search engines and recommendation systems. Backpack Language Models, unlike traditional transformer-based models that treat text sequences as monolithic structures, generate outputs as weighted combinations of non-contextual, learned word aspects, also known as senses. Leveraging this architecture, we propose a framework for debiasing ranking tasks. Our experimental results show that this framework effectively mitigates gender bias in text retrieval and ranking with minimal degradation in performance.

Paper Structure

This paper contains 5 sections, 8 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Methodology
  4. Experiments
  5. Conclusion

Figures (3)

  • Figure 1: Overview of our bias-controllable ranking framework. Tokens are decomposed into non-contextual sense vectors, whose sensitivities to a target aspect are measured. A policy $\mathcal{M}_{\mathbf{s}}(\ell)$ reweights these senses, and the aggregated representations are passed through an MLP to yield the final relevance score.
  • Figure 1: Ranking performance on the MS MARCO passage dev set. Decreasing the value of $\lambda$ slightly reduces ranking performance, highlighting the fairness–performance trade-off.
  • Figure 2: Effect of the control weight $\lambda$ on the trade-off between ranking effectiveness (NDCG@10) and gender fairness, quantified as $1 - \mathrm{ARaB}$.