cringepnh/koelectra-korean-sentiment

🎬 KoELECTRA Korean Sentiment Analyzer

Fine-tuned Korean sentiment classification model based on KoELECTRA, trained on the NSMC dataset (Naver Sentiment Movie Corpus).

Classifies Korean movie reviews as positive (1) or negative (0).

GitHub: cringepnh/korean-sentiment-analyzer

📊 Results

Evaluated on the full NSMC test set (49,157 samples after cleaning):

Metric	Score
Accuracy	90.2%
Precision	90.2%
Recall	90.3%
F1 Score	90.3%

🚀 How to Use

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "cringepnh/koelectra-korean-sentiment"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
model.eval()

def predict(text):
    inputs = tokenizer(text, return_tensors="pt", padding="max_length",
                       truncation=True, max_length=128)
    with torch.no_grad():
        outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=-1)
    label = torch.argmax(probs).item()
    confidence = probs[0][label].item()
    sentiment = "Positive ✅" if label == 1 else "Negative ❌"
    return {"sentiment": sentiment, "confidence": f"{confidence:.1%}"}

# Examples
print(predict("이 영화 정말 재미있어요! 배우들 연기도 최고!"))
# {'sentiment': 'Positive ✅', 'confidence': '99.4%'}

print(predict("완전 별로... 시간 낭비했다."))
# {'sentiment': 'Negative ❌', 'confidence': '99.5%'}

🏋️ Training Details

• Base model: monologg/koelectra-base-finetuned-sentiment
• Dataset: NSMC (146,182 training samples after cleaning)
• Epochs: Up to 10 with early stopping (patience=2 on eval_loss)
• Batch size: 32
• Learning rate: 2e-5
• Max token length: 128
• Hardware: CPU (laptop)

📁 Dataset

NSMC (Naver Sentiment Movie Corpus)

• 150,000 training / 50,000 test Korean movie reviews
• Binary labels: 0 (negative), 1 (positive)
• Source: github.com/e9t/nsmc

📄 License

MIT