dansachs/indonesian-religious-corpus

Indonesian Religious Corpus

A large-scale corpus of Indonesian religious texts collected from authoritative religious websites, designed for religiolect (religious dialect) classification and analysis.

Dataset Description

The Indonesian Religious Corpus contains approximately 3 million clean sentences scraped from 100+ authoritative religious websites representing three major religious communities in Indonesia: Islam, Catholicism, and Protestantism.

Dataset Summary

• Total Sentences: ~3,000,000
• Languages: Indonesian (Bahasa Indonesia)
• Task: Text Classification (3-class religiolect classification)
• Format: CSV
• License: MIT (for academic research purposes)

Supported Tasks

• Text Classification: Classify Indonesian text into religious denominations (Islam, Catholic, Protestant)
• Religiolect Analysis: Study linguistic patterns and vocabulary differences across religious communities
• NLP Research: Fine-tuning language models for Indonesian religious text understanding

Dataset Structure

Data Fields

The dataset contains the following columns:

• Label (string): The religious denomination label - one of "Islam", "Catholic", or "Protestant"
• Denomination (string): Full denomination name or identifier
• Location (string): Geographic location or region associated with the text (if available)
• Date (string): Publication date of the source text (if available)
• Title (string): Title of the article or page from which the sentence was extracted
• Sentence_Unit (string): The actual sentence text to be classified (main feature)
• Link (string): Source URL where the text was collected from

Data Splits

The dataset is provided as a single train split containing all collected sentences.

Data Sources

Website Distribution

• 30 Catholic websites including:

  • Mirifica (Catholic news and media)
  • KAS (Komisi Awam Sekuler)
  • And other authoritative Catholic sources

• 27 Islamic websites including:

  • NU Online (Nahdlatul Ulama)
  • And other authoritative Islamic sources

• 44 Protestant websites including:

  • PGI (Persekutuan Gereja-gereja di Indonesia)
  • And other authoritative Protestant sources

All sources are publicly accessible, authoritative religious websites that publish content in Indonesian.

Data Collection

Methodology

The dataset was collected using an ethical web crawling pipeline with the following features:

• Rate Limiting: 1 request per domain to respect server resources
• Adaptive Backoff: Automatic retry with exponential backoff on 429/503 errors
• Resumable: Crawling can be stopped and resumed at any time
• Language Detection: Automatic filtering to ensure only Indonesian text is included
• Noise Removal: Cleaning of scripture references, navigation elements, and other non-content text
• Live Dashboard: Real-time progress tracking during collection

Preprocessing

The collected text underwent several preprocessing steps:

1. Language Detection: Non-Indonesian content was filtered out
2. Sentence Segmentation: Text was split into individual sentences
3. Cleaning:
   • Removal of scripture references (e.g., "Matius 5:3")
   • Removal of navigation elements
   • Removal of boilerplate text
   • Normalization of whitespace
4. Deduplication: Duplicate sentences were removed based on content hashing
5. Metadata Extraction: Title, date, location, and source URL were preserved

Dataset Statistics

• Total Sentences: ~3,000,000
• Class Distribution:
  • Islam: ~1,000,000 sentences
  • Catholic: ~1,000,000 sentences
  • Protestant: ~1,000,000 sentences
• Average Sentence Length: ~20-30 words
• Time Period: Collected from contemporary religious websites
• Geographic Coverage: Primarily Indonesia, with some international sources

Usage

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("dansachs/indonesian-religious-corpus")

# Access the data
train_data = dataset["train"]
print(f"Total examples: {len(train_data)}")
print(f"Features: {train_data.column_names}")

# Example access
example = train_data[0]
print(f"Label: {example['Label']}")
print(f"Sentence: {example['Sentence_Unit']}")

Using with Pandas

from datasets import load_dataset
import pandas as pd

# Load and convert to pandas
dataset = load_dataset("dansachs/indonesian-religious-corpus")
df = dataset["train"].to_pandas()

# Filter by denomination
catholic_texts = df[df['Label'] == 'Catholic']
print(f"Catholic sentences: {len(catholic_texts)}")

Training a Model

from datasets import load_dataset
from transformers import AutoTokenizer

# Load dataset
dataset = load_dataset("dansachs/indonesian-religious-corpus")

# Tokenize for training
tokenizer = AutoTokenizer.from_pretrained("indolem/indobert-base-uncased")

def tokenize_function(examples):
    return tokenizer(
        examples["Sentence_Unit"],
        padding="max_length",
        truncation=True,
        max_length=128
    )

tokenized_dataset = dataset.map(tokenize_function, batched=True)

Related Models

This dataset was used to train the following model:

🔗 Model: dansachs/indo-religiolect-bert-v2

A fine-tuned IndoBERT model for Indonesian religious text classification.

Repository

🔗 GitHub Repository: dansachs/indo-religiolects

The repository includes:

• Data collection pipeline and crawler scripts
• Training scripts and notebooks
• Model inference tools
• Full documentation

Limitations and Considerations

Ethical Considerations

• Religious Sensitivity: This dataset deals with religious content and should be used responsibly
• Bias: The dataset reflects the content available on the web and may contain biases present in online religious discourse
• Representation: The dataset may not represent all perspectives within each religious community
• Temporal Bias: Content reflects the time period when it was collected

Technical Limitations

• Source Dependency: Data quality depends on the quality and availability of source websites
• Language: Dataset is specifically for Indonesian text; performance on other languages is not guaranteed
• Domain Specificity: Model trained on this data may not generalize well to other domains
• Preprocessing Artifacts: Some sentences may contain remnants of HTML or formatting

Data Quality

• Some metadata fields (Location, Date) may be missing for certain entries
• Sentence segmentation may occasionally split or merge sentences incorrectly
• Some sentences may be very short or very long due to the nature of web content

Citation

If you use this dataset in your research, please cite:

@misc{indonesian-religious-corpus,
  title={Indonesian Religious Corpus: A Large-Scale Dataset for Religiolect Classification},
  author={Sachs, Dan},
  year={2025},
  howpublished={\url{https://huggingface.co/datasets/dansachs/indonesian-religious-corpus}}
}

Acknowledgments

• Data Sources: All religious websites that provided publicly accessible content
• Base Infrastructure: Built with Hugging Face Datasets
• Language Model: Designed for use with IndoBERT

License

MIT License - For academic research purposes.

Please use this dataset responsibly and in accordance with the terms of service of the source websites.