Identifying and Mitigating Social Bias Knowledge in Language Models

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Chen, Ruizhe, Li, Yichen, Yang, Jianfei, Zhou, Joey Tianyi, Wu, Jian, Liu, Zuozhu
Format: Preprint
Veröffentlicht: 2024
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866909513278291968
author Chen, Ruizhe
Li, Yichen
Yang, Jianfei
Zhou, Joey Tianyi
Wu, Jian
Liu, Zuozhu
author_facet Chen, Ruizhe
Li, Yichen
Yang, Jianfei
Zhou, Joey Tianyi
Wu, Jian
Liu, Zuozhu
contents Generating fair and accurate predictions plays a pivotal role in deploying large language models (LLMs) in the real world. However, existing debiasing methods inevitably generate unfair or incorrect predictions as they are designed and evaluated to achieve parity across different social groups but leave aside individual commonsense facts, resulting in modified knowledge that elicits unreasonable or undesired predictions. In this paper, we first establish a new bias mitigation benchmark, BiaScope, which systematically assesses performance by leveraging newly constructed datasets and metrics on knowledge retention and generalization. Then, we propose a novel debiasing approach, Fairness Stamp (FAST), which enables fine-grained calibration of individual social biases. FAST identifies the decisive layer responsible for storing social biases and then calibrates its outputs by integrating a small modular network, considering both bias mitigation and knowledge-preserving demands. Comprehensive experiments demonstrate that FAST surpasses state-of-the-art baselines with superior debiasing performance while not compromising the overall model capability for knowledge retention and downstream predictions. This highlights the potential of fine-grained debiasing strategies to achieve fairness in LLMs.
format Preprint
id arxiv_https___arxiv_org_abs_2408_11843
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Identifying and Mitigating Social Bias Knowledge in Language Models
Chen, Ruizhe
Li, Yichen
Yang, Jianfei
Zhou, Joey Tianyi
Wu, Jian
Liu, Zuozhu
Computation and Language
Artificial Intelligence
Generating fair and accurate predictions plays a pivotal role in deploying large language models (LLMs) in the real world. However, existing debiasing methods inevitably generate unfair or incorrect predictions as they are designed and evaluated to achieve parity across different social groups but leave aside individual commonsense facts, resulting in modified knowledge that elicits unreasonable or undesired predictions. In this paper, we first establish a new bias mitigation benchmark, BiaScope, which systematically assesses performance by leveraging newly constructed datasets and metrics on knowledge retention and generalization. Then, we propose a novel debiasing approach, Fairness Stamp (FAST), which enables fine-grained calibration of individual social biases. FAST identifies the decisive layer responsible for storing social biases and then calibrates its outputs by integrating a small modular network, considering both bias mitigation and knowledge-preserving demands. Comprehensive experiments demonstrate that FAST surpasses state-of-the-art baselines with superior debiasing performance while not compromising the overall model capability for knowledge retention and downstream predictions. This highlights the potential of fine-grained debiasing strategies to achieve fairness in LLMs.
title Identifying and Mitigating Social Bias Knowledge in Language Models
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2408.11843