BERTurk Turkish Sentiment Analysis
A fine-tuned version of dbmdz/bert-base-turkish-cased (BERTurk) for sentiment classification of Turkish text. Given a Turkish sentence, it predicts whether the sentiment is Positive, Negative, or Notr (neutral).
Results
Evaluated on the held-out test split of the winvoker/turkish-sentiment-analysis-dataset:
| Metric | Score |
|---|---|
| Accuracy | 0.9669 |
| Macro F1 | 0.9414 |
Usage
from transformers import pipeline
clf = pipeline("text-classification", model="Omar1010/bert-turkish-sentiment")
print(clf("Bu ürün gerçekten harika, çok memnun kaldım."))
Labels
Positive— olumluNegative— olumsuzNotr— nötr / neutral
Training
- Base model: BERTurk (
dbmdz/bert-base-turkish-cased) - Dataset: winvoker/turkish-sentiment-analysis-dataset (subsampled to 60,000 train / 8,000 test)
- Epochs: 2, learning rate: 2e-5, max length: 128, fp16
Limitations
Trained mainly on product-review-style Turkish text, so it performs best on similar language. It may be less reliable on sarcasm, very informal slang, or highly domain-specific text.