🇹🇷 BERTurk for Turkish Sentiment Analysis
This model is a fine-tuned version of BERTurk 32k on the TRSAv1 dataset, a labeled collection of Turkish e-commerce reviews categorized into positive, neutral, and negative sentiments. For more details about the dataset, methodology, and experiments, you can refer to the corresponding research paper.
How to Use
You can use the model directly with 🤗 Transformers:
from transformers import pipeline
classifier = pipeline("text-classification", model="incidelen/bert-base-turkish-sentiment-analysis-cased")
result = classifier("Ürün çok kaliteli, paketleme harikaydı. Kesinlikle tavsiye ederim!")
print(result)
Citation
If you use this model in your research or application, please cite the following paper:
@article{incidelen15sentiment,
title={Sentiment Analysis in Turkish Using Language Models: A Comparative Study},
author={{\.I}ncidelen, Mert and Aydo{\u{g}}an, Murat},
journal={European Journal of Technique (EJT)},
volume={15},
number={1},
pages={68--74},
publisher={Hibetullah KILI{\c{C}}}
}
Dataset Overview
The TRSAv1 dataset includes 150,000 Turkish product reviews from e-commerce platforms. It is balanced across three sentiment classes:
| Sentiment | Count |
|---|---|
| Negative | 50,000 |
| Neutral | 50,000 |
| Positive | 50,000 |
| TOTAL | 150,000 |
Evaluation Results
Overall Performance
| Accuracy (%) | Precision (%) | Recall (%) | F1 Score (%) |
|---|---|---|---|
| 83.69 | 83.68 | 83.69 | 83.65 |
Class-wise Performance
| Sentiment | Precision (%) | Recall (%) | F1 Score (%) |
|---|---|---|---|
| Negative | 88.39 | 85.08 | 86.71 |
| Neutral | 77.29 | 76.05 | 76.67 |
| Positive | 85.35 | 89.93 | 87.58 |