Work

The Indonesian Religiolect Corpus: Data Curation for Muslim, Protestant, and Catholic Language Varieties

Year: 2026

Type: conference-paper

Abstract: Religiolects-language varieties shaped by religious community identity-are low-resource domains often overlooked within high-resource languages.We present the Indo-Religiolect Corpus, 1 the first large-scale dataset for Indonesian religious language variation, containing 3 million sentences from over 100 institutional websites representing Muslim, Catholic, and Protestant communities.Finetuning IndoBERT demonstrates these religiolects are computationally distinguishable: Islamic Indonesian exhibits high distinctiveness (91.73%), while Catholic and Protestant varieties share substantial lexical overlap yet retain detectable shibboleths (86.41% and 86.64%).Our findings indicate a potential for representation collapse: models trained on majoritynormative data may default to secular or Muslim-dominant Indonesian, blurring distinct minority voices.We hypothesize that these gaps plausibly translate into downstream fairness risks for applications like content moderation and automated hiring.This corpus offers a template for documenting sub-national varieties, advancing linguistic equity beyond "National Language" benchmarks toward "No Language Variety Left Behind."

Source: Workshop on Language Models for Low-Resource Languages (LoResLM)

DOI: 10.18653/v1/2026.loreslm-1.36

Author: Dan Sachs

Institution: Universitas Gadjah Mada

Language: English

Topic: Natural Language Processing Techniques

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

SDG: Reduced inequalities

Open Access status: hybrid

Funders: American Indonesian Exchange Foundation, Universitas Gadjah Mada