PhoBERT V2: Hate Speech Detection for Vietnamese Text

This model is a fine-tuned version of vinai/phobert-base-v2 on the ViHSD (Vietnamese Hate Speech Detection Dataset) for classifying Vietnamese text into three categories: CLEAN, OFFENSIVE, and HATE.

Model Details

• Base Model: vinai/phobert-base-v2
• Description: PhoBERT V2 fine-tuned cho bài toán phân loại Hate Speech tiếng Việt
• Architecture: PhoBERT-V2 (Phiên bản cải tiến của PhoBERT với tokenizer syllable-based)
• Dataset: ViHSD (Vietnamese Hate Speech Detection Dataset)
• Fine-tuning Framework: HuggingFace Transformers + PyTorch
• Task: Hate Speech Classification (3 classes)

Hyperparameters

• Batch size: 32
• Learning rate: 2e-5
• Epochs: 100
• Max sequence length: 256
• Weight decay: 0.01
• Warmup steps: 500
• Early stopping patience: 5
• Optimizer: AdamW
• Learning rate scheduler: Cosine with warmup

Dataset

Model was trained on ViHSD (Vietnamese Hate Speech Detection Dataset) containing ~10,000 Vietnamese comments from social media.

Label Descriptions:

• CLEAN (0): Normal content without offensive language
• OFFENSIVE (1): Mildly offensive or inappropriate content
• HATE (2): Hate speech, extremist language, severe threats

Evaluation Results

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

• Accuracy: 0.9341
• Macro-F1: 0.8048
• Weighted-F1: 0.9326
• Macro-Precision: 0.8306
• Macro-Recall: 0.7854

Basic Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
model_name = "visolex/phobert-v2-hsd"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(
    model_name
)

# Classify text
text = "Văn bản tiếng Việt cần phân loại"
inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True)

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

# Label mapping
label_names = {
    0: "CLEAN",
    1: "OFFENSIVE",
    2: "HATE"
}

print(f"Predicted label: {label_names[predicted_label]}")
print(f"Confidence scores: {predictions[0].tolist()}")

Training Details

Training Data

• Dataset: ViHSD (Vietnamese Hate Speech Detection Dataset)
• Total samples: ~10,000 Vietnamese comments from social media
• Training split: ~70%
• Validation split: ~15%
• Test split: ~15%

Training Configuration

• Framework: PyTorch + HuggingFace Transformers
• Optimizer: AdamW
• Learning Rate: 2e-5
• Batch Size: 32
• Max Length: 256 tokens
• Epochs: 100 (with early stopping patience: 5)
• Weight Decay: 0.01
• Warmup Steps: 500

Contact & Support

• GitHub: ViSoLex Hate Speech Detection
• Issues: Report Issues
• Questions: Open a discussion on the model's Hugging Face page

License

This model is distributed under the MIT License.

Acknowledgments

• Base model: vinai/phobert-base-v2
• Dataset: ViHSD (Vietnamese Hate Speech Detection Dataset)
• Framework: Hugging Face Transformers
• ViSoLex Toolkit