visobert-spam-multiclass

visobert-spam-multi-class: Spam Review Detection for Vietnamese Text

This model is a fine-tuned version of uitnlp/visobert on the ViSpamReviews dataset for spam review detection in Vietnamese e-commerce reviews.

Model Details

• Base Model: uitnlp/visobert
• Description: ViSoBERT - Vietnamese Social BERT
• Dataset: ViSpamReviews (Vietnamese Spam Review Dataset)
• Fine-tuning Framework: HuggingFace Transformers
• Task: Spam Review Detection (multi-class)
• Number of Classes: 4

Hyperparameters

• Max sequence length: 256
• Learning rate: 5e-5
• Batch size: 32
• Epochs: 100
• Early stopping patience: 5

Dataset

The model was trained on the ViSpamReviews dataset, which contains 19,860 Vietnamese e-commerce review samples. The dataset includes:

• Train set: 14,299 samples (72%)
• Validation set: 1,590 samples (8%)
• Test set: 3,971 samples (20%)

Label Distribution

• NO-SPAM (0): Genuine product reviews
• SPAM-1 (1): Fake review (synthetic/manipulated reviews)
• SPAM-2 (2): Brand-only reviews (only mention brand without product details)
• SPAM-3 (3): Irrelevant reviews (unrelated content)

Results

The model was evaluated on the test set with the following metrics:

• Accuracy: 0.9083
• Macro-F1: 0.7871

Usage

You can use this model for spam review detection in Vietnamese text. Below is an example:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
model_name = "visolex/visobert-spam-multiclass"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Example review text
text = "Sản phẩm này rất tốt, shop giao hàng nhanh!"

# Tokenize
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)

# Predict
with torch.no_grad():
    outputs = model(**inputs)
    predicted_class = outputs.logits.argmax(dim=-1).item()
    probabilities = torch.softmax(outputs.logits, dim=-1)


# Map to label
label_map = {
    0: "NO-SPAM",
    1: "SPAM-1 (fake review)",
    2: "SPAM-2 (brand-only)",
    3: "SPAM-3 (irrelevant)"
}
predicted_label = label_map[predicted_class]
confidence = probabilities[0][predicted_class].item()

print(f"Text: {text}")
print(f"Predicted: {predicted_label} (confidence: {confidence:.2%})")

Citation

If you use this model, please cite:

@misc{{
  {model_key}_spam_detection,
  title={{{description}}},
  author={{ViSoLex Team}},
  year={{2025}},
  howpublished={{\url{{https://huggingface.co/{visolex/visobert-spam-multiclass}}}}}
}}

License

This model is released under the Apache-2.0 license.

Acknowledgments

• Base model: {base_model}
• Dataset: ViSpamReviews (Vietnamese Spam Review Dataset)
• ViSoLex Toolkit for Vietnamese NLP