Analysis and Mitigation of Religion Bias in Indonesian Natural Language Processing Datasets
Dublin Core
Title
Analysis and Mitigation of Religion Bias in Indonesian Natural Language Processing Datasets
Subject
natural language processing; Indonesian NLP; social bias; debiasing
Description
Previous studies have shown the existence of misrepresentation regarding various religious identities in Indonesian media.
Misrepresentations of other marginalized identities in natural language processing (NLP) datasets have been recorded to
inflict harm against such marginalized identities, in cases such as automated content moderation, and as such must be
mitigated. In this paper, we analyze, for the first time, several Indonesian NLP datasets to see whether they contain unwanted
bias and the effects of debiasing on them. We find that two, out of three, datasets analyzed in this study contain unwanted bias,
whose effects trickle down to downstream performance under the form of allocation and representation harm. The results of
debiasing at the dataset level, as a response to the biases previously discovered, are consistently positive for the respective
dataset. Nevertheless, depending on the dataset and embedding used to train the model, they vary highly at the downstream
performance level. In particular, the same debiasing technique can decrease bias on a combination of datasets and embedding,
yet increase bias on another, particularly in the case of representation harm.
Misrepresentations of other marginalized identities in natural language processing (NLP) datasets have been recorded to
inflict harm against such marginalized identities, in cases such as automated content moderation, and as such must be
mitigated. In this paper, we analyze, for the first time, several Indonesian NLP datasets to see whether they contain unwanted
bias and the effects of debiasing on them. We find that two, out of three, datasets analyzed in this study contain unwanted bias,
whose effects trickle down to downstream performance under the form of allocation and representation harm. The results of
debiasing at the dataset level, as a response to the biases previously discovered, are consistently positive for the respective
dataset. Nevertheless, depending on the dataset and embedding used to train the model, they vary highly at the downstream
performance level. In particular, the same debiasing technique can decrease bias on a combination of datasets and embedding,
yet increase bias on another, particularly in the case of representation harm.
Creator
Muhammad Arief Fauzan, Ari Saptawijaya
Source
http://jurnal.iaii.or.id
Publisher
Professional Organization Ikatan Ahli Informatika Indonesia (IAII)/Indonesian Informatics Experts Association
Date
August 2023
Contributor
Sri Wahyuni
Rights
ISSN Media Electronic: 2580-0760
Format
PDF
Language
English
Type
Text
Files
Collection
Citation
Muhammad Arief Fauzan, Ari Saptawijaya, “Analysis and Mitigation of Religion Bias in Indonesian Natural Language Processing Datasets,” Repository Horizon University Indonesia, accessed February 3, 2026, https://repository.horizon.ac.id/items/show/10030.