IGED: Indonesian Grammar Error correction Dataset

⚠️ Citation Required: If you use this dataset in any publication, you must cite the CASTLE paper (see Citation section below).

---

Dataset Summary

IGED is a large-scale Indonesian Grammar Error Correction (GEC) dataset containing 1,345,096 sentence pairs — the largest publicly available GEC dataset for Indonesian. Each pair consists of an ungrammatical source sentence and its grammatically correct target.

The dataset covers three major error categories in Indonesian: morphological (affixation, word formation, reduplication), syntactic (phrase structure, preposition, sentence completeness), and semantic errors (diction, pleonasm, ambiguity).

IGED was introduced as part of the CASTLE system — a Context-Aware Semantic Transformer with Knowledge Graph Enhancement for low-resource grammar correction, published in Expert Systems With Applications (2026).

---

Supported Tasks

• Grammatical Error Correction (GEC): Given an ungrammatical Indonesian sentence, generate the corrected version.
• Error Type Classification: Each sample includes a fine-grained error category label for research on targeted correction.

---

Languages

Indonesian (id) — monolingual.

---

Dataset Structure

Data Fields

Field	Type	Description
src	string	Ungrammatical / error-containing source sentence
trg	string	Grammatically corrected target sentence
category	string	Fine-grained error category (see below)
split	string	Dataset split: train, val, or test

Error Categories

Category	Type	Proportion
afiksasi (Affixation)	Morphological	19.8%
bentuk_kata (Word Formation)	Morphological	12.3%
reduplikasi (Reduplication)	Morphological	6.7%
struktur_frasa / sintaksis_frasa (Phrase Structure)	Syntactic	19.0%
preposisi (Preposition)	Syntactic	12.6%
kelengkapan_kalimat (Sentence Completeness)	Syntactic	11.9%
diksi (Diction)	Semantic	10.6%
ambiguitas (Ambiguity)	Semantic	5.0%
pleonasme (Pleonasm)	Semantic	2.0%

Data Splits

Split	Samples
Train	1,072,192
Validation	134,024
Test	134,025
Total	1,345,096

Splits are pre-stratified by error category. Use the split column directly — do not re-split.

---

Dataset Creation

Source Data

IGED was constructed from two Indonesian text corpora:

1. Journalistic corpus — 150,466 articles from 6 major Indonesian news websites (Muzad & Rahutomo, 2016), providing naturalistic sentence structures.
2. Academic corpus — 800 undergraduate theses (skripsi S1) from Universitas Nahdlatul Ulama Yogyakarta, covering IT, Economics, and Halal Industry departments.

Error Generation

A hybrid three-stage pipeline was used to inject grammatical errors into correct sentences:

1. Rule-based (60%) — Morphological errors (affixation, word formation, reduplication) generated using the Sastrawi stemmer, POS tagging, and a whitelist from Wiktionary.
2. LLM-driven (40%) — Semantic errors (ambiguity, pleonasm) generated using DeepSeek R1 with constrained prompting to produce plausible but incorrect variants.
3. Statistical collocation — PMI-based co-occurrence patterns (window=3) to ensure naturalness.

Validation

All generated pairs were validated through:

• Automated filtering: Levenshtein similarity threshold (θ=0.8) + IndoBERT cosine similarity ≥ 0.6
• Manual review: 500 random samples (0.2% sample) reviewed by 1 linguistics lecturer + 4 linguistics graduate students
• Inter-annotator agreement: Cohen's κ = 0.82

Statistics

Metric	Value
Average sentence length	21.3 words
Average character length	154.3 characters
Average edit distance (src→trg)	12.10

The average edit distance of 12.10 is notably higher than comparable GEC datasets, reflecting the complexity and diversity of Indonesian morphological and syntactic errors.

---

Usage

With datasets library

from datasets import load_dataset

dataset = load_dataset("syauqie/IGED")

# Pre-split into train/val/test
train_ds = dataset["train"]     # 1,072,192 samples
val_ds   = dataset["validation"]  # 134,024 samples
test_ds  = dataset["test"]      # 134,025 samples

# Access fields
example = train_ds[0]
print(example["src"])       # Ungrammatical sentence
print(example["trg"])       # Corrected sentence
print(example["category"])  # Error category

Filter by error category

# Get only semantic error samples for training
semantic_cats = ["diksi", "ambiguitas", "pleonasme"]
semantic_train = train_ds.filter(lambda x: x["category"] in semantic_cats)
print(f"Semantic error samples: {len(semantic_train)}")

With the CASTLE model

This dataset is the official training data for the CASTLE model. To reproduce results from the paper, use the official CASTLE code:

git clone https://github.com/syauqiezjut/CASTLE-GEC.git
cd CASTLE-GEC
pip install -r requirements.txt
python scripts/train_castle.py --config configs/castle_base.yaml

Expected results on IGED test set:

• F1 Score: 0.9629
• BLEU: 92.72

---

Limitations and Biases

• Domain: The dataset is drawn from news articles and undergraduate theses. Performance may vary on other domains (social media, formal government documents, etc.).
• Error types: IGED does not currently cover punctuation errors or code-switching (campur kode), which are common in informal Indonesian text.
• Error injection: Errors are synthetically generated. While expert-validated, the distribution may not perfectly reflect naturally occurring errors in real-world writing.
• Language variety: IGED covers standard Indonesian (Bahasa Indonesia baku). Regional dialect variations are not represented.

---

Citation

If you use IGED in your research or applications, you must cite the following paper:

@article{marier2026castle,
  title     = {{CASTLE}: Context-Aware Semantic Transformer with Knowledge Graph Enhancement 
               for Low-Resource Grammar Correction},
  author    = {Marier, Syauqie Muhammad and {[co-authors]}},
  journal   = {Expert Systems With Applications},
  volume    = {299},
  pages     = {130233},
  year      = {2026},
  publisher = {Elsevier},
  doi       = {10.1016/j.eswa.2025.130233},
  url       = {https://doi.org/10.1016/j.eswa.2025.130233}
}

Failure to cite the paper when using IGED in published work violates the dataset license (CC BY 4.0), which requires attribution.

---

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

You are free to:

• Share — copy and redistribute the material in any medium or format
• Adapt — remix, transform, and build upon the material for any purpose, even commercially

Under the following terms:

• Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made. Citation of the CASTLE paper is required.

See LICENSE for full details.

---

Contact

For questions about the dataset or the CASTLE model, contact:

Syauqie Muhammad Marier
Email: syauqie.mm@gmail.com
Affiliation: Zhejiang University of Technology / Universitas Nahdlatul Ulama Yogyakarta
GitHub: syauqiezjut/CASTLE-GEC

---

Related Resources

• 📄 Paper: CASTLE (ESWA 2026)
• 💻 Code: syauqiezjut/CASTLE-GEC
• 🤗 Model: syauqiezjut/CASTLE (uploading after training)