Saved in:
Bibliographic Details
Main Authors: Udagawa, Takuma, Zhao, Yang, Kanayama, Hiroshi, Bhattacharjee, Bishwaranjan
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2504.14212
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912567889231872
author Udagawa, Takuma
Zhao, Yang
Kanayama, Hiroshi
Bhattacharjee, Bishwaranjan
author_facet Udagawa, Takuma
Zhao, Yang
Kanayama, Hiroshi
Bhattacharjee, Bishwaranjan
contents Large language models (LLMs) acquire general linguistic knowledge from massive-scale pretraining. However, pretraining data mainly comprised of web-crawled texts contain undesirable social biases which can be perpetuated or even amplified by LLMs. In this study, we propose an efficient yet effective annotation pipeline to investigate social biases in the pretraining corpora. Our pipeline consists of protected attribute detection to identify diverse demographics, followed by regard classification to analyze the language polarity towards each attribute. Through our experiments, we demonstrate the effect of our bias analysis and mitigation measures, focusing on Common Crawl as the most representative pretraining corpus.
format Preprint
id arxiv_https___arxiv_org_abs_2504_14212
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Bias Analysis and Mitigation through Protected Attribute Detection and Regard Classification
Udagawa, Takuma
Zhao, Yang
Kanayama, Hiroshi
Bhattacharjee, Bishwaranjan
Computation and Language
Large language models (LLMs) acquire general linguistic knowledge from massive-scale pretraining. However, pretraining data mainly comprised of web-crawled texts contain undesirable social biases which can be perpetuated or even amplified by LLMs. In this study, we propose an efficient yet effective annotation pipeline to investigate social biases in the pretraining corpora. Our pipeline consists of protected attribute detection to identify diverse demographics, followed by regard classification to analyze the language polarity towards each attribute. Through our experiments, we demonstrate the effect of our bias analysis and mitigation measures, focusing on Common Crawl as the most representative pretraining corpus.
title Bias Analysis and Mitigation through Protected Attribute Detection and Regard Classification
topic Computation and Language
url https://arxiv.org/abs/2504.14212