YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
LRSA-XLMR-Sentiment-ID
Model Description
LRSA-XLMR-Sentiment-ID is a fine-tuned XLM-RoBERTa (XLM-R) model for Indonesian sentiment analysis developed as part of the study:
Stress-Testing Large Language Models (LLMs) against Code-Mixing and Distributional Shifts in Low-Resource NLP
The model was trained and evaluated on a large Indonesian sentiment corpus containing approximately 75,000 samples collected from social media and news domains.
Task
Sentiment Classification
Labels:
- Negative
- Neutral
- Positive
Dataset
Sources:
- YouTube comments
- Indonesian news articles and comments
Domains:
- Politics
- Economics
- Social issues
- Public policy
Performance
| Metric | Score |
|---|---|
| Accuracy | 0.8610 |
| Macro Precision | 0.8409 |
| Macro Recall | 0.8532 |
| Macro F1 | 0.8466 |
Robustness Evaluation
| Dataset | Macro F1 |
|---|---|
| Clean | 0.8466 |
| Noise p=0.1 | 0.8352 |
| Noise p=0.2 | 0.8175 |
| Noise p=0.3 | 0.7982 |
Cross-Domain Generalization
| Domain | Macro F1 |
|---|---|
| YouTube | 0.7872 |
| News | 0.8614 |
Generalization Score: 0.8243
Statistical Significance
McNemar testing demonstrated statistically significant differences between XLM-R and the baseline SVM model (p < 0.001), confirming the effectiveness of transformer-based architectures under low-resource Indonesian sentiment classification settings.
Repository
Code, experimental results, figures, and supplementary materials:
https://github.com/mziarehman4353/LRSA-LLM
Authors
Zia Ul Rehman Zafar
Dedi Gunawan
Endang Wahyu Pamungkas
Widi Widayat
Helmi Imaduddin
Department of Informatics Engineering
Universitas Muhammadiyah Surakarta, Indonesia
Citation
If you use this model, please cite:
Zafar, Z. U. R., Gunawan, D., Pamungkas, E. W., Widayat, W., & Imaduddin, H.
Stress-Testing Large Language Models (LLMs) against Code-Mixing and Distributional Shifts in Low-Resource NLP.
License
MIT License
- Downloads last month
- 38