catalina-titoveanu/review-bert-romanian

ReviewBERT: Romanian Product Review Classifier

🧠 Model Overview

ReviewBERT is a fine-tuned Romanian language BERT model designed to classify product reviews on a scale from 1 to 5. It is based on the readerbench/RoBERT-base checkpoint and was fine-tuned using a dataset of 15826 Romanian-language product reviews on cosmetics. The task is multiclass sentiment classification, where:

• 1 = Very Negative
• 2 = Negative
• 3 = Neutral
• 4 = Positive
• 5 = Very Positive

📊 Evaluation Metrics

Classification Report

Label	Precision	Recall	F1-score	Support
1	0.92	0.96	0.94	746
2	0.96	0.90	0.93	775
3	0.91	0.90	0.90	771
4	0.87	0.85	0.86	789
5	0.86	0.90	0.88	768
Accuracy			0.90	3849
Macro avg	0.90	0.90	0.90	3849
Weighted avg	0.90	0.90	0.90	3849

🛠️ Training Details

• Base model: readerbench/RoBERT-base
• Tokenizer: BertTokenizerFast
• Max sequence length: 128
• Batch size: 16
• Learning rate: 2e-5
• Epochs: 4
• Optimizer: AdamW
• Class balancing with: RandomOverSampler
• Trained on: Google Colab

💡 How to Use

from transformers import AutoTokenizer, AutoModelForSequenceClassification
from transformers import pipeline

model = AutoModelForSequenceClassification.from_pretrained("catalina-titoveanu/romanian-product-reviews")
tokenizer = AutoTokenizer.from_pretrained("catalina-titoveanu/romanian-product-reviews")

clf = pipeline("text-classification", model=model, tokenizer=tokenizer)
clf("Produsul a fost groaznic, nu recomand.")

⚠️ Limitations & Biases

• The model was trained exclusively on cosmetics product reviews and may not generalize well to other Romanian-language domains.
• Sentiment labels (especially mid-range scores like 2–4) are inherently subjective and could reflect user bias.

👋 Author

Created by Catalina Titoveanu
Feel free to open an issue for bugs, feedback, or ideas.

---

Built with care and too much coffee. ☕🐈