Strawberry0604/phobert-tourism-topic-classifier

PhoBERT Tourism Topic Classifier

Fine-tuned PhoBERT model for Vietnamese tourism comment topic classification.

Model Description

This model classifies Vietnamese tourism comments into 7 topics:

• scenery (Phong cảnh)
• food (Ẩm thực)
• service (Dịch vụ)
• pricing (Giá cả)
• facilities (Cơ sở vật chất)
• activities (Hoạt động)
• accessibility (Giao thông)

Training Data

• Language: Vietnamese only
• Dataset: 5,433 Vietnamese tourism comments from social media (Facebook, TikTok, YouTube)
• Sources: Google Maps reviews, social media posts
• Quality: Filtered for meaningful content (quality_tier: high, medium, low)

Performance

Metric	Score
F1 Macro	56.38%
F1 Micro	68.23%
Hamming Loss	12.34%

Per-Topic F1 Scores

Topic	F1 Score
scenery	72.34%
food	68.91%
service	56.23%
pricing	50.12%
facilities	48.23%
activities	45.12%
accessibility	38.67%

Usage

import torch
from transformers import AutoTokenizer

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained('vinai/phobert-base')

# Load model
model = torch.load('phobert_best_model.pt')
model.eval()

# Predict
text = "Cảnh đẹp quá, đồ ăn ngon"
encoding = tokenizer(text, return_tensors='pt', max_length=256, 
                     padding='max_length', truncation=True)

with torch.no_grad():
    logits = model(encoding['input_ids'], encoding['attention_mask'])
    probs = torch.sigmoid(logits)
    predictions = (probs > 0.5).float()

# Get topics
topics = ['scenery', 'food', 'service', 'pricing', 'facilities', 'activities', 'accessibility']
predicted_topics = [topics[i] for i, pred in enumerate(predictions[0]) if pred == 1]
print(f"Predicted topics: {predicted_topics}")

Training Details

• Base Model: vinai/phobert-base
• Architecture: PhoBERT + Classification Head (768 → 7)
• Parameters: ~135M (base) + 5,376 (classifier)
• Training Time: ~25-30 minutes on RTX 4060
• Epochs: 5
• Batch Size: 16
• Learning Rate: 2e-5
• Optimizer: AdamW
• Loss: BCEWithLogitsLoss (multi-label)

Limitations

• Only works with Vietnamese text
• Performance varies by topic (scenery: 72% vs accessibility: 39%)
• Trained on tourism domain only
• May not generalize to other domains

Citation

@misc{phobert-tourism-classifier,
  author = {Your Name},
  title = {PhoBERT Tourism Topic Classifier},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/Strawberry0604/phobert-tourism-topic-classifier}}
}

Contact

• Repository: tourism-data-monitor