indobert-sentiment-indonesian-playstore

sentiment-indonesian-playstore

Fine-tuned RoBERTa model for Indonesian sentiment analysis on Google Play Store reviews.

Model Description

This model performs 3-class sentiment classification:

• Positive (label: 2) 😊
• Neutral (label: 1) 😐
• Negative (label: 0) 😞

Training Data

• Dataset: SMSA (Sentiment Analysis on Indonesian Movie Reviews)
• Language: Indonesian (Bahasa Indonesia)
• Size: 11,000 reviews
  • Positive: 6,416 reviews (58.3%)
  • Negative: 3,436 reviews (31.2%)
  • Neutral: 1,148 reviews (10.4%)
• Domain: App reviews (Google Play Store)

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "rkkzone/roberta-sentiment-indonesian-playstore"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Example Indonesian review
text = "Aplikasi bagus sekali! Sangat direkomendasikan."

inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True, max_length=128)

with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
    sentiment = torch.argmax(predictions, dim=-1).item()

sentiment_labels = {0: "Negative", 1: "Neutral", 2: "Positive"}
print(f"Sentiment: {sentiment_labels[sentiment]}")
print(f"Confidence: {predictions[0][sentiment].item():.4f}")

Batch Processing

reviews = [
    "Aplikasi bagus, mudah digunakan!",
    "Tidak bisa login, aplikasi error terus",
    "Biasa aja sih"
]

inputs = tokenizer(reviews, return_tensors="pt", padding=True, truncation=True, max_length=128)

with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
    sentiments = torch.argmax(predictions, dim=-1)

for review, sentiment, probs in zip(reviews, sentiments, predictions):
    label = sentiment_labels[sentiment.item()]
    conf = probs[sentiment].item()
    print(f"Review: {review}")
    print(f"Sentiment: {label} ({conf:.2%} confident)\n")

Model Performance

Training Configuration:

• Base Model: IndoBERT (indobenchmark/indobert-base-p1)
• Training Epochs: 3
• Batch Size: 16
• Learning Rate: 2e-5
• Max Length: 128 tokens
• Optimizer: AdamW
• Weight Decay: 0.01

Evaluation Results:

• Check model card for detailed metrics on test set
• Includes per-class precision, recall, and F1 scores

Dataset Preprocessing

The model was trained with the following preprocessing:

1. Indonesian Slang Normalization: Common abbreviations expanded
2. Text Cleaning: Remove special characters, URLs, excessive whitespace
3. Tokenization: BertTokenizer with max 128 tokens
4. Label Mapping:
   • negative → 0
   • neutral → 1
   • positive → 2

Intended Use

This model is optimized for:

• ✅ Indonesian language app reviews
• ✅ Google Play Store sentiment analysis
• ✅ Customer feedback classification
• ✅ Review monitoring and analytics
• ✅ E-commerce product reviews (Indonesian)

Limitations

• Optimized for informal Indonesian text (app reviews)
• May not generalize well to formal Indonesian literature
• Best performance on short texts (< 128 tokens)
• Sensitive to slang variations across Indonesian regions
• Trained primarily on app domain data

Applications

1. App Analytics Dashboard

Analyze user sentiment in Indonesian app reviews to track satisfaction over time.

2. Customer Support Prioritization

Automatically flag negative reviews for urgent response.

3. Market Research

Understand Indonesian user preferences and pain points.

4. Review Filtering

Filter out spam or irrelevant reviews based on sentiment patterns.

Example Predictions

Citation

@misc{roberta_sentiment_indonesian_playstore,
  author = {rkkzone},
  title = {Indonesian Sentiment Analysis for Google Play Reviews},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/rkkzone/roberta-sentiment-indonesian-playstore}
}

License

MIT License - Free for commercial and non-commercial use.

Related Resources

• English version: rkkzone/roberta-sentiment-playstore
• Dataset: SMSA Indonesian Sentiment
• Live Demo: Google Play Review Analyzer

Acknowledgments

• Dataset: Indonesian NLP research community (SMSA dataset)
• Base Model: IndoNLP team for IndoBERT
• Framework: Hugging Face Transformers

---

Built with ❤️ for Indonesian app developers

For issues or questions, please open an issue on the model repository.