ZenyxS/indobert-emotion-emotionclf

🤖 IndoBERT Emotion Classifier

A fine-tuned IndoBERT model for Indonesian-language emotion classification in conversational or chatbot contexts. This model was developed as part of a research project on detecting user satisfaction and emotional states from short-text feedback and responses.

It classifies input text into one of five emotions:

• —😄 Senang — Happy / satisfied
• —😐 Netral — Neutral / indifferent
• —🤔 Bingung — Confused / unsure
• —😤 Frustrasi — Frustrated / unsatisfied
• —😡 Marah — Angry / annoyed

📌 Model Details

• —Model name: username/indobert-emotion-classifier <!-- replace username with your HF username -->
• —Developed by: Fabian Prasetyo
• —Institution: SMKN 21 Jakarta
• —Language: Indonesian (id)
• —License: Apache 2.0
• —Base model: indobenchmark/indobert-base-p1
• —Task: Emotion classification from short conversational text
• —Number of classes: 5

🎯 Use Cases

✅ Direct Use

• —Emotion recognition for Indonesian chatbot conversations.
• —Sentiment/feedback analysis for educational or customer service systems.
• —Emotion-aware dialogue systems and virtual assistants.

🔄 Downstream Use

• —As a module in larger conversational AI systems.
• —As a feature extractor for satisfaction prediction or escalation pipelines.

🚫 Out-of-Scope Use

• —❌ Clinical or psychological diagnosis.
• —❌ Text domains unrelated to conversation or short feedback (e.g., news articles, long essays).

⚠️ Bias, Risks, and Limitations

• —The dataset consists of student chatbot feedback and may not generalize beyond similar conversational contexts.
• —Minority labels (Frustrasi, Marah) are underrepresented — this leads to low recall for those classes.
• —Heavy slang, code-switching (Indonesian-English), typos, and non-standard spelling can reduce model accuracy.
• —Use caution before applying in high-risk or safety-critical settings (no clinical/medical decisions, no legal decisions, etc.).

🚀 Quick Start

from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline

model_name = "username/indobert-emotion-classifier"  # replace 'username' with your HF username
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

classifier = pipeline("text-classification", model=model, tokenizer=tokenizer, return_all_scores=True)

text = "Apaan sih ini, nggak jelas banget!"
print(classifier(text))
✅ Example Output:

python
[
  {'label': 'Frustrasi', 'score': 0.3700},
  {'label': 'Bingung', 'score': 0.3357},
  {'label': 'Netral', 'score': 0.1256},
  {'label': 'Marah', 'score': 0.1058},
  {'label': 'Senang', 'score': 0.0629}
]

📊 IndoBERT Emotion Classifier

A fine-tuned IndoBERT model for classifying emotions in Indonesian short-chat responses. Developed as part of a student research project at SMKN 21 Jakarta.

📁 Dataset

• —Source: 1,896 annotated chatbot responses
• —Participants: 237 students
• —Label distribution:
• —Netral: 1,018
• —Bingung: 442
• —Senang: 249
• —Frustrasi: 98
• —Marah: 89

⚠️ Note: The dataset is imbalanced, particularly for Frustrasi and Marah. This strongly affects per-class performance.

⚙️ Training Procedure

• —Train/Test Split: 80/20
• —Epochs: 3
• —Optimizer: AdamW
• —Loss Function: CrossEntropy with class weights
• —Precision: fp32
• —Hardware: NVIDIA T4 (~2 min/epoch)
• —Framework: 🤗 Transformers + PyTorch

🧠 Model Architecture

• —Base Model: IndoBERT (BERT-base, ~110M parameters)
• —12 layers, 768 hidden size, 12 attention heads
• —Fine-tuning: All layers (full fine-tune)

📈 Evaluation Results

• —Overall Accuracy: ~0.33
• —Macro F1: 0.22

Interpretation: The model captures dominant intents (neutral, confused, positive) reasonably well, but struggles with rare classes (Frustrasi, Marah) due to dataset imbalance.

🧪 Example Predictions

def predict(text, classifier):
    scores = classifier(text)[0]
    return sorted(scores, key=lambda x: x["score"], reverse=True)[0]

print(predict("Apaan sih ini, nggak jelas banget!"))
# → Frustrasi (0.37)

print(predict("Aku sangat senang jawabannya jelas sekali"))
# → Senang (0.42)

print(predict("Hah? Maksudnya gimana?"))
# → Bingung (0.37)

🧠 Technical Specifications

Architecture: IndoBERT (BERT-base: 12 layers, 768 hidden, 12 heads, ~110M parameters) Fine-tuning: Full fine-tuning (all layers) Training time: ~2 minutes/epoch on NVIDIA T4 (reported) Framework: 🤗 Transformers + PyTorch

🛠️ Recommendations (Actionable)

Class imbalance

• —Oversample minority classes.
• —Apply stronger augmentation (back-translation, synonym replacement).
• —Experiment with focal loss or class-balanced loss functions.

Preprocessing

• —Normalize slang and common typos.
• —Apply light text normalization:
• —lowercasing
• —repeated-character normalization
• —token-level normalization for common chat abbreviations

Data collection

• —Collect more labeled samples for Frustrasi and Marah.
• —Use active learning to prioritize annotation of high-uncertainty samples.

Evaluation

• —Report confidence intervals (e.g., via bootstrap).
• —Provide confusion matrix and per-class metrics.

Deployment

• —Apply quantization (e.g., 8-bit) for reduced latency on edge devices.
• —Define fallback policy: escalate or request human review when Frustrasi or Marah prediction exceeds a threshold.

📬 Contact

Author: Fabian Prasetyo Institution: SMKN 21 Jakarta Email: fabianprasetyo133@gmail.com

📚 Citation

If you use this model, please cite it as:

@misc{prasetyo2025indobert,
  title={IndoBERT Emotion Classifier},
  author={Fabian Prasetyo},
  institution={SMKN 21 Jakarta},
  year={2025},
  howpublished={\url{https://huggingface.co/username/indobert-emotion-classifier}}
}