@inproceedings{yang-etal-2025-bias,
title = "Bias A-head? Analyzing Bias in Transformer-Based Language Model Attention Heads",
author = "Yang, Yi and
Duan, Hanyu and
Abbasi, Ahmed and
Lalor, John P. and
Tam, Kar Yan",
editor = "Cao, Trista and
Das, Anubrata and
Kumarage, Tharindu and
Wan, Yixin and
Krishna, Satyapriya and
Mehrabi, Ninareh and
Dhamala, Jwala and
Ramakrishna, Anil and
Galystan, Aram and
Kumar, Anoop and
Gupta, Rahul and
Chang, Kai-Wei",
booktitle = "Proceedings of the 5th Workshop on Trustworthy NLP (TrustNLP 2025)",
month = may,
year = "2025",
address = "Albuquerque, New Mexico",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.trustnlp-main.18/",
doi = "10.18653/v1/2025.trustnlp-main.18",
pages = "276--290",
ISBN = "979-8-89176-233-6",
abstract = "Transformer-based pretrained large language models (PLM) such as BERT and GPT have achieved remarkable success in NLP tasks. However, PLMs are prone to encoding stereotypical biases. Although a burgeoning literature has emerged on stereotypical bias mitigation in PLMs, such as work on debiasing gender and racial stereotyping, how such biases manifest and behave internally within PLMs remains largely unknown. Understanding the internal stereotyping mechanisms may allow better assessment of model fairness and guide the development of effective mitigation strategies. In this work, we focus on attention heads, a major component of the Transformer architecture, and propose a bias analysis framework to explore and identify a small set of biased heads that are found to contribute to a PLM{'}s stereotypical bias. We conduct extensive experiments to validate the existence of these biased heads and to better understand how they behave. We investigate gender and racial bias in the English language in two types of Transformer-based PLMs: the encoder-based BERT model and the decoder-based autoregressive GPT model, LLaMA-2 (7B), and LLaMA-2-Chat (7B). Overall, the results shed light on understanding the bias behavior in pretrained language models."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yang-etal-2025-bias">
<titleInfo>
<title>Bias A-head? Analyzing Bias in Transformer-Based Language Model Attention Heads</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yi</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hanyu</namePart>
<namePart type="family">Duan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ahmed</namePart>
<namePart type="family">Abbasi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">John</namePart>
<namePart type="given">P</namePart>
<namePart type="family">Lalor</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kar</namePart>
<namePart type="given">Yan</namePart>
<namePart type="family">Tam</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 5th Workshop on Trustworthy NLP (TrustNLP 2025)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Trista</namePart>
<namePart type="family">Cao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anubrata</namePart>
<namePart type="family">Das</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tharindu</namePart>
<namePart type="family">Kumarage</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yixin</namePart>
<namePart type="family">Wan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Satyapriya</namePart>
<namePart type="family">Krishna</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ninareh</namePart>
<namePart type="family">Mehrabi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jwala</namePart>
<namePart type="family">Dhamala</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anil</namePart>
<namePart type="family">Ramakrishna</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aram</namePart>
<namePart type="family">Galystan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anoop</namePart>
<namePart type="family">Kumar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rahul</namePart>
<namePart type="family">Gupta</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kai-Wei</namePart>
<namePart type="family">Chang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Albuquerque, New Mexico</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-233-6</identifier>
</relatedItem>
<abstract>Transformer-based pretrained large language models (PLM) such as BERT and GPT have achieved remarkable success in NLP tasks. However, PLMs are prone to encoding stereotypical biases. Although a burgeoning literature has emerged on stereotypical bias mitigation in PLMs, such as work on debiasing gender and racial stereotyping, how such biases manifest and behave internally within PLMs remains largely unknown. Understanding the internal stereotyping mechanisms may allow better assessment of model fairness and guide the development of effective mitigation strategies. In this work, we focus on attention heads, a major component of the Transformer architecture, and propose a bias analysis framework to explore and identify a small set of biased heads that are found to contribute to a PLM’s stereotypical bias. We conduct extensive experiments to validate the existence of these biased heads and to better understand how they behave. We investigate gender and racial bias in the English language in two types of Transformer-based PLMs: the encoder-based BERT model and the decoder-based autoregressive GPT model, LLaMA-2 (7B), and LLaMA-2-Chat (7B). Overall, the results shed light on understanding the bias behavior in pretrained language models.</abstract>
<identifier type="citekey">yang-etal-2025-bias</identifier>
<identifier type="doi">10.18653/v1/2025.trustnlp-main.18</identifier>
<location>
<url>https://aclanthology.org/2025.trustnlp-main.18/</url>
</location>
<part>
<date>2025-05</date>
<extent unit="page">
<start>276</start>
<end>290</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Bias A-head? Analyzing Bias in Transformer-Based Language Model Attention Heads
%A Yang, Yi
%A Duan, Hanyu
%A Abbasi, Ahmed
%A Lalor, John P.
%A Tam, Kar Yan
%Y Cao, Trista
%Y Das, Anubrata
%Y Kumarage, Tharindu
%Y Wan, Yixin
%Y Krishna, Satyapriya
%Y Mehrabi, Ninareh
%Y Dhamala, Jwala
%Y Ramakrishna, Anil
%Y Galystan, Aram
%Y Kumar, Anoop
%Y Gupta, Rahul
%Y Chang, Kai-Wei
%S Proceedings of the 5th Workshop on Trustworthy NLP (TrustNLP 2025)
%D 2025
%8 May
%I Association for Computational Linguistics
%C Albuquerque, New Mexico
%@ 979-8-89176-233-6
%F yang-etal-2025-bias
%X Transformer-based pretrained large language models (PLM) such as BERT and GPT have achieved remarkable success in NLP tasks. However, PLMs are prone to encoding stereotypical biases. Although a burgeoning literature has emerged on stereotypical bias mitigation in PLMs, such as work on debiasing gender and racial stereotyping, how such biases manifest and behave internally within PLMs remains largely unknown. Understanding the internal stereotyping mechanisms may allow better assessment of model fairness and guide the development of effective mitigation strategies. In this work, we focus on attention heads, a major component of the Transformer architecture, and propose a bias analysis framework to explore and identify a small set of biased heads that are found to contribute to a PLM’s stereotypical bias. We conduct extensive experiments to validate the existence of these biased heads and to better understand how they behave. We investigate gender and racial bias in the English language in two types of Transformer-based PLMs: the encoder-based BERT model and the decoder-based autoregressive GPT model, LLaMA-2 (7B), and LLaMA-2-Chat (7B). Overall, the results shed light on understanding the bias behavior in pretrained language models.
%R 10.18653/v1/2025.trustnlp-main.18
%U https://aclanthology.org/2025.trustnlp-main.18/
%U https://doi.org/10.18653/v1/2025.trustnlp-main.18
%P 276-290
Markdown (Informal)
[Bias A-head? Analyzing Bias in Transformer-Based Language Model Attention Heads](https://aclanthology.org/2025.trustnlp-main.18/) (Yang et al., TrustNLP 2025)
ACL