experiment_result / README.md
michaelcs564's picture
Update README.md
5d6f7a1 verified
|
Raw History Blame Contribute Delete
4.28 kB
---
license: apache-2.0
language:
- id
- vi
- min
- ace
- htu
- jav
task_categories:
- text-generation
tags:
- natural-language-processing
- low-resource-languages
- multilingual
- text-generation
- synthetic-data
- data-generation
- southeast-asia
- computational-linguistics
- underrepresented-languages
- language-models
size_categories:
- 10K<n<100K
---
# Synthetic Text Generation in Low-Resource Southeast Asian Languages
## Overview
This dataset contains **54,000 synthetically generated text samples** across six Southeast Asian languages, produced through a systematic evaluation of large language models (LLMs) for low-resource language text generation.
The dataset accompanies the research paper:
**Text Generation in Low-Resource Languages Using Human-Curated Vocabularies**
The research investigates whether LLMs can generate coherent text in low-resource and extremely low-resource languages by leveraging human-curated dictionaries, word candidates, and human-written exemplars.
Our approach combines human-curated vocabulary resources with few-shot prompting and Zipfian word sampling to improve lexical coverage and text generation quality.
## Research Paper
- **Paper:** [Text Generation in Low-Resource Languages Using Human-Curated Vocabularies](https://arxiv.org/abs/2610.37105v1)
- **Implementation:** [vocab2text (GitHub)](https://github.com/michaelsinanta/vocab2text)
Please refer to the paper for the complete methodology, experimental analysis, and evaluation results.
## Dataset Composition
The dataset is generated from an experimental setup consisting of five dimensions:
| Dimension | Configurations |
|---|---|
| Languages | 6 |
| Language Models | 3 |
| Thinking Modes | 2 |
| Prompting Strategies | 3 |
| Text Lengths | 5 |
| **Total Configurations** | **540** |
| Outputs per Configuration | 100 |
| **Total Generated Samples** | **54,000** |
### Languages
We experiment with six Southeast Asian languages spanning different levels of resource availability.
| Language | ISO 639-3 Code | Resource Availability |
|---|---|---|
| Indonesian | `ind` | Medium-resource |
| Vietnamese | `vie` | Medium-resource |
| Javanese | `jav` | Low-resource |
| Minangkabau | `min` | Low-resource |
| Acehnese | `ace` | Extremely low-resource |
| Hitu | `htu` | Extremely low-resource |
### Language Models and Thinking Modes
Three LLMs are evaluated under two reasoning configurations:
| Model | Thinking Enabled | Thinking Disabled |
|---|---|---|
| Gemini 3.5 Flash | ✓ | ✓ |
| Claude Haiku 4.5 | ✓ | ✓ |
| GPT-5.4 Mini | ✓ | ✓ |
This results in six model configurations.
### Prompting Strategies
We evaluate three prompting strategies that differ in the amount of linguistic information provided to the LLM.
1. **Topic-Conditioned Zero-Shot Prompt (Zero-Shot)**
Instructs the model to generate text in the target language about a specified topic without providing additional linguistic examples or vocabulary candidates.
2. **Few-Shot Exemplars Only Prompt (Few-Shot)**
Provides human-written exemplars of the target language and requested text length to guide generation.
3. **Few-Shot Exemplars and Word Candidates Prompt (Ours)**
Combines human-written exemplars with candidate words sampled from human-curated dictionaries using a Zipfian distribution. This is the proposed vocabulary-guided generation strategy.
### Text Lengths
Five target text lengths are considered:
| Category | Target Length |
|---|---|
| Short sentence | 5 words |
| Sentence | 10 words |
| Sentence | 15 words |
| Long sentence | 20 words |
| Paragraph | 3 sentences and 25 words |
## Citation
If you use this dataset in your research, please cite the accompanying paper:
```bibtex
@misc{sinanta2026textgenerationlowresourcelanguages,
title={Text Generation in Low-Resource Languages Using Human-Curated Vocabularies},
author={Michael Christlambert Sinanta and Muhammad Dehan Al Kautsar and Amanda Nurul Izzah and Thanh-Nhi Nguyen and Phan Phuong Mai Chau and Zheng Wei Lim and Fajri Koto},
year={2026},
eprint={2610.37105},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2610.37105},
}
```
## License
This dataset is released under the Apache License 2.0.