codealchemist01/turkish-sentiment-analysis

Turkish Sentiment Analysis Model

A fine-tuned BERT model for Turkish sentiment analysis, trained on a combined dataset of 439,384 labeled Turkish sentences.

Model Details

• Base Model: dbmdz/bert-base-turkish-cased
• Task: Text Classification (Sentiment Analysis)
• Language: Turkish
• Labels: positive, negative, neutral

Training Data

The model was trained on a combination of two high-quality Turkish sentiment datasets:

• winvoker/turkish-sentiment-analysis-dataset (440,641 samples)
• WhiteAngelss/Turkce-Duygu-Analizi-Dataset (440,641 samples)

After deduplication and preprocessing, the final training set consisted of:

• Training: 351,507 samples
• Validation: 43,938 samples
• Test: 43,939 samples

Label Distribution

• Positive: 234,957 (53.5%)
• Neutral: 153,809 (35.0%)
• Negative: 50,618 (11.5%)

Training

• Epochs: 3
• Learning Rate: 2e-5
• Batch Size: 16
• Max Length: 128
• Optimizer: AdamW

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
model_name = "codealchemist01/turkish-sentiment-analysis"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Example text
text = "Bu ürün gerçekten harika!"

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

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

# Map to label
id2label = {0: "negative", 1: "neutral", 2: "positive"}
predicted_label = id2label[predicted_label_id]
confidence = predictions[0][predicted_label_id].item()

print(f"Label: {predicted_label}")
print(f"Confidence: {confidence:.4f}")

Performance

Evaluation metrics on the test set (43,939 samples):

• Accuracy: 97.45%
• Weighted F1: 97.42%
• Weighted Precision: 97.41%
• Weighted Recall: 97.45%

Per-Class Performance

Class	Precision	Recall	F1-Score	Support
Negative	91.42%	86.69%	88.99%	5,062
Neutral	99.79%	99.96%	99.87%	15,381
Positive	97.15%	98.12%	97.63%	23,496

Note: Negative class has lower performance due to class imbalance (only 11.5% of the dataset). The model performs excellently on neutral and positive classes.

Limitations

• The model may not perform well on very short texts (< 3 words)
• Performance may vary across different domains (social media, news, reviews)
• Class imbalance may affect performance on minority classes (negative)

Citation

If you use this model, please cite:

@misc{turkish-sentiment-analysis,
  title={Turkish Sentiment Analysis Model},
  author={codealchemist01},
  year={2024},
  howpublished={\url{https://huggingface.co/codealchemist01/turkish-sentiment-analysis}}
}

License

Apache 2.0