Thai Sentiment Analysis with PhayaThaiBERT

Production-ready Thai sentiment classifier with 82% accuracy. Fine-tuned PhayaThaiBERT on 21k social media messages.

Quick Demo

Try it now: Live Demo

Or run locally:

pip install transformers torch gradio
python app.py

Features

• Multiple Interfaces: CLI, Web UI, REST API
• Production Ready: Docker, FastAPI, HuggingFace hosting
• High Performance: 82% accuracy, ONNX optimization
• Explainable: LIME interpretability
• Well Tested: Unit tests, CI/CD
• Hardware Agnostic: GPU/Mac/CPU support

Performance

Metric	Score
Overall Accuracy	82%
Positive F1	0.72
Neutral F1	0.85
Negative F1	0.85

Quick Start

Use Pre-trained Model (No Training)

from transformers import pipeline

classifier = pipeline("sentiment-analysis",
                     model="SiemonCha/thai-sentiment-phayabert")

result = classifier("อาหารอร่อยมาก")
print(result)
# [{'label': 'POSITIVE', 'score': 0.98}]

Run Web UI

git clone https://github.com/SiemonCha/thai-sentiment
cd thai-sentiment
pip install -r requirements.txt
python app.py

REST API

python api.py
# API docs: http://localhost:8000/docs

curl -X POST "http://localhost:8000/predict" \
  -H "Content-Type: application/json" \
  -d '{"text": "อาหารอร่อยมาก"}'

Docker

docker build -t thai-sentiment .
docker run -p 7860:7860 thai-sentiment

Training from Scratch

# 1. Install dependencies
pip install -r requirements.txt

# 2. Download dataset
python data_download.py

# 3. Train model (~40 min Mac M1 / ~15 min GPU)
python train.py

# 4. Evaluate
python evaluate.py

Project Structure

thai-sentiment/
├── app.py                 # Gradio web UI
├── api.py                 # FastAPI REST API
├── demo.py                # Interactive CLI
├── predict.py             # Batch predictions
├── train.py               # Training script
├── evaluate.py            # Evaluation with metrics
├── explain.py             # LIME explainability
├── optimize.py            # ONNX conversion
├── benchmark.py           # Model comparison
├── test_model.py          # Unit tests
├── Dockerfile             # Container
└── requirements.txt       # Dependencies

Hardware Support

Platform	Training Time	Status
Mac M1/M2	~40 min	>> Tested
NVIDIA GPU	~15 min	>> Tested
AMD GPU (Linux)	~15 min	>> Tested
CPU	~4-5 hours	>> Works
AMD GPU (Windows)	~4-5 hours	XX CPU fallback

Testing

# Run tests
pytest test_model.py -v
pytest test_model.py test_model_advanced.py -v

# Benchmark
python benchmark.py

# Explain predictions
python explain.py

Advanced Features

Model Explainability

python explain.py
# Generates explanation.html showing word contributions

ONNX Optimization (3x faster)

python optimize.py
# Creates optimized model in ./model_onnx/

Batch Processing

python batch_predict.py
# Processes input.csv → output.csv

Technical Details

• Base Model: PhayaThaiBERT (110M parameters)
• Dataset: Wisesight Sentiment (21k messages)
• Training: 5 epochs, class-weighted loss, AdamW
• Architecture: BERT with classification head
• Max Length: 128 tokens
• Classes: Positive, Neutral, Negative

Citation

@misc{thai-sentiment-2025,
  author = {SiemonCha},
  title = {Thai Sentiment Analysis with PhayaThaiBERT},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/SiemonCha/thai-sentiment}
}

Dataset:

@software{wisesight_sentiment,
  author = {Suriyawongkul, Arthit and Chuangsuwanich, Ekapol},
  title = {PyThaiNLP/wisesight-sentiment},
  year = 2019,
  doi = {10.5281/zenodo.3457447}
}

License

MIT License

Links

• Live Demo: https://huggingface.co/spaces/SiemonCha/thai-sentiment-demo
• Model Hub: https://huggingface.co/SiemonCha/thai-sentiment-phayabert
• GitHub: https://github.com/SiemonCha/thai-sentiment