AcehX

AcehX is a high-quality Acehnese language resource designed to support NLP research, including Masked Language Modeling (MLM) and sentiment analysis for low-resource Indonesian regional languages.

This repository contains:

• AcehX Corpus (unlabeled dataset)
• AcehX Sentiment Dataset (labeled dataset)
• AcehX Tokenizer Vocabulary
• AcehXBERT-base (MLM Model)
• AcehXBERT-sentiment (Classification Model)

AcehXBERT is introduced in the paper:

“Development of AcehX for Sentiment Analysis Using a BERT-Based Model” (2025)

📚 Dataset Overview

1. AcehX Corpus for MLM

The full AcehX corpus contains 63,367 sentences.

In this study, 41,627 sentences were used for MLM pretraining of AcehXBERT.

A large unlabeled Acehnese corpus collected from:

• literature, folklore, oral traditions
• traditional customs and customary law texts
• Wikipedia
• YouTube comments

Dataset	Type	Sentences
AcehX Corpus	Unlabeled (MLM)	41,627

Used for MLM pretraining of AcehXBERT.

2. AcehX Sentiment for Sentiment Analysis

Manually labeled by native Acehnese speakers into:

• Positive
• Negative
• Neutral

Dataset	Type	Sentences
AcehX Sentiment	Labeled	3,725

Used for fine-tuning AcehXBERT classification model.

3. Distribution of sentiment dataset labels

Sub Dataset	Positive	Negative	Neutral	Total Sentences
Training	884	931	1,537	3,352
Validation	75	76	111	262
Test	29	34	48	111
Total	998	1,041	1,696	3,725

Hyperparameters classification model

Hyperparameters	Value
Sequence length	128
Batch size	8
Epoch	10
Learning rate	5e-5

🔤 AcehX Tokenizer

The IndoBERT tokenizer is extended with Acehnese vocabulary.

• Extracted vocabulary: 8,961 tokens
• New tokens added: 4,765
• Final vocabulary size: 35,286 tokens

This significantly reduces OOV (out-of-vocabulary) tokens and improves Acehnese representation.

🧠 AcehXBERT Model

AcehXBERT is created by:

1. Extending the IndoBERT tokenizer with Acehnese tokens
2. Pretraining with Masked Language Modeling (MLM) on AcehX Corpus
3. Fine-tuning for sentiment classification (3 classes)

Performance

📌 AcehX MLM Results

Model	Perplexity
AcehXBERT	78.3575
NusaBERT Base	140.7514
IndoBERT Base	76479.9258

📌 AcehX Sentiment Results

Model	F1-macro
AcehXBERT	82.50%
NusaBERT Base	75.36%
BERT Base	73.78%
IndoBERT Base	73.00%
RoBERTa Base	72.77%

AcehXBERT outperforms existing Indonesian and multilingual BERT models.

Authors

• Doni Sumito Sukiswo (Universitas Syiah Kuala (USK))
• Hammam Riza* (National Research and Innovation Agency (BRIN))
• Muhammad Subianto (Universitas Syiah Kuala (USK))
• Taufik Fuadi Abidin (Universitas Syiah Kuala (USK))
• Afnan (Universitas Syiah Kuala (USK))

* Corresponding Author

Acknowledgement

The AcehXBERT dataset was collaboratively collected by the author and the following contributors:

Name	Affiliation
Doni Sumito Sukiswo	Universitas Syiah Kuala (USK)
Muhammad Ruski	Universitas Syiah Kuala (USK)
Habib Asy Muhyi	Universitas Syiah Kuala (USK)
Muttaqin	Universitas Syiah Kuala (USK)
Iskandar	Universitas Syiah Kuala (USK)
Saryulis	Universitas Syiah Kuala (USK)
Maria Ulfa	Universitas Syiah Kuala (USK)
Fadhilah Syafa	Universitas Syiah Kuala (USK)
Nurul Asmi Amalia	Universitas Syiah Kuala (USK)
Aida Afira	Universitas Syiah Kuala (USK)
Hafizha Dini Giandra	Universitas Syiah Kuala (USK)

Citation

You can find the full details in our paper:
Development of AcehX for Sentiment Analysis Using a BERT-Based Model

Proceedings of the 28th O-COCOSDA Conference, IEEE, 2025

https://doi.org/10.1109/O-COCOSDA68185.2025.11384824

If you use AcehX or AcehXBERT, please cite:

@INPROCEEDINGS{11384824,
  author={Sukiswo, Doni Sumito and Riza, Hammam and Subianto, Muhammad and Abidin, Taufik Fuadi and Afnan},
  booktitle={2025 28th Conference of the Oriental COCOSDA International Committee for the Co-ordination and Standardisation of Speech Databases and Assessment Techniques (O-COCOSDA)}, 
  title={Development of Acehx for Sentiment Analysis Using a Bert-Based Model}, 
  year={2025},
  volume={},
  number={},
  pages={1-6},
  keywords={Training;Adaptation models;Sentiment analysis;Analytical models;Vocabulary;Technological innovation;Linguistics;Transformers;Cultural differences;Artificial intelligence;AcehX;sentiment analysis;NLP;transformers;BERT;deep learning},
  doi={10.1109/O-COCOSDA68185.2025.11384824}}