visolex / mbert-absa-hotel

mbert-absa-hotel: Aspect-based Sentiment Analysis for Vietnamese Reviews

This model is a fine-tuned version of bert-base-multilingual-cased on the VLSP2018-ABSA-Hotel dataset for aspect-based sentiment analysis in Vietnamese reviews.

Model Details

• Base Model : bert-base-multilingual-cased
• Description : mBERT (bert-base-multilingual-cased) fine-tuned
• Dataset : VLSP2018-ABSA-Hotel
• Fine-tuning Framework : HuggingFace Transformers
• Task : Aspect-based Sentiment Classification (3 classes)

Hyperparameters

• Batch size: 32
• Learning rate: 3e-5
• Epochs: 100
• Max sequence length: 256
• Weight decay: 0.01
• Warmup steps: 500
• Optimizer: AdamW

Dataset

Model was trained on VLSP2018 ABSA Hotel dataset for aspect-based sentiment analysis.

Sentiment Labels:

• 0 - Negative (Tiêu cực): Negative opinions
• 1 - Neutral (Trung lập): Neutral, objective opinions
• 2 - Positive (Tích cực): Positive opinions

Aspect Categories:

Model được train để phân tích sentiment cho các aspects sau:

• FACILITIES#CLEANLINESS
• FACILITIES#COMFORT
• FACILITIES#DESIGN&FEATURES
• FACILITIES#GENERAL
• FACILITIES#MISCELLANEOUS
• FACILITIES#PRICES
• FACILITIES#QUALITY
• FOOD&DRINKS#MISCELLANEOUS
• FOOD&DRINKS#PRICES
• FOOD&DRINKS#QUALITY
• FOOD&DRINKS#STYLE&OPTIONS
• HOTEL#CLEANLINESS
• HOTEL#COMFORT
• HOTEL#DESIGN&FEATURES
• HOTEL#GENERAL
• HOTEL#MISCELLANEOUS
• HOTEL#PRICES
• HOTEL#QUALITY
• LOCATION#GENERAL
• ROOMS#CLEANLINESS
• ROOMS#COMFORT
• ROOMS#DESIGN&FEATURES
• ROOMS#GENERAL
• ROOMS#MISCELLANEOUS
• ROOMS#PRICES
• ROOMS#QUALITY
• ROOM_AMENITIES#CLEANLINESS
• ROOM_AMENITIES#COMFORT
• ROOM_AMENITIES#DESIGN&FEATURES
• ROOM_AMENITIES#GENERAL
• ROOM_AMENITIES#MISCELLANEOUS
• ROOM_AMENITIES#PRICES
• ROOM_AMENITIES#QUALITY
• SERVICE#GENERAL

Evaluation Results

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

• Accuracy : 0.9456
• Macro-F1 : 0.4630
• Weighted-F1 : 0.7258
• Macro-Precision : 0.7090
• Macro-Recall : 0.3792

Usage Example

import torch
from transformers import AutoTokenizer, AutoModel

# Load model and tokenizer
repo = "visolex/mbert-absa-hotel"
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModel.from_pretrained(repo, trust_remote_code=True)
model.eval()

# Aspect labels for VLSP2018-ABSA-Hotel
aspect_labels = [
    "FACILITIES#CLEANLINESS",
    "FACILITIES#COMFORT",
    "FACILITIES#DESIGN&FEATURES",
    "FACILITIES#GENERAL",
    "FACILITIES#MISCELLANEOUS",
    "FACILITIES#PRICES",
    "FACILITIES#QUALITY",
    "FOOD&DRINKS#MISCELLANEOUS",
    "FOOD&DRINKS#PRICES",
    "FOOD&DRINKS#QUALITY",
    "FOOD&DRINKS#STYLE&OPTIONS",
    "HOTEL#CLEANLINESS",
    "HOTEL#COMFORT",
    "HOTEL#DESIGN&FEATURES",
    "HOTEL#GENERAL",
    "HOTEL#MISCELLANEOUS",
    "HOTEL#PRICES",
    "HOTEL#QUALITY",
    "LOCATION#GENERAL",
    "ROOMS#CLEANLINESS",
    "ROOMS#COMFORT",
    "ROOMS#DESIGN&FEATURES",
    "ROOMS#GENERAL",
    "ROOMS#MISCELLANEOUS",
    "ROOMS#PRICES",
    "ROOMS#QUALITY",
    "ROOM_AMENITIES#CLEANLINESS",
    "ROOM_AMENITIES#COMFORT",
    "ROOM_AMENITIES#DESIGN&FEATURES",
    "ROOM_AMENITIES#GENERAL",
    "ROOM_AMENITIES#MISCELLANEOUS",
    "ROOM_AMENITIES#PRICES",
    "ROOM_AMENITIES#QUALITY",
    "SERVICE#GENERAL"
]

# Sentiment labels
sentiment_labels = ["POSITIVE", "NEGATIVE", "NEUTRAL"]

# Example review text
text = "Khách sạn rất sạch sẽ, phòng ốc thoải mái nhưng giá hơi cao."

# Tokenize
inputs = tokenizer(
    text,
    return_tensors="pt",
    padding=True,
    truncation=True,
    max_length=256
)
inputs.pop("token_type_ids", None)

# Predict
with torch.no_grad():
    outputs = model(**inputs)

# Get logits: shape [1, num_aspects, num_sentiments + 1]
logits = outputs.logits.squeeze(0)  # [num_aspects, num_sentiments + 1]
probs = torch.softmax(logits, dim=-1)

# Predict for each aspect
none_id = probs.size(-1) - 1  # Index of "none" class
results = []

for i, aspect in enumerate(aspect_labels):
    prob_i = probs[i]
    pred_id = int(prob_i.argmax().item())
    
    if pred_id != none_id and pred_id < len(sentiment_labels):
        score = prob_i[pred_id].item()
        if score >= 0.5:  # threshold
            results.append((aspect, sentiment_labels[pred_id].lower()))

print(f"Text: {text}")
print(f"Predicted aspects: {results}")
# Output example: [('aspects', 'positive'), ('aspects', 'positive'), ('aspects', 'negative')]

Citation

If you use this model, please cite:

@misc{visolex_absa_mbert_absa_hotel,
  title={mBERT (bert-base-multilingual-cased) fine-tuned for Vietnamese Aspect-based Sentiment Analysis},
  author={ViSoLex Team},
  year={2025},
  url={https://huggingface.co/visolex/mbert-absa-hotel}
}

License

This model is released under the Apache-2.0 license.

Acknowledgments

• Base model: bert-base-multilingual-cased
• Dataset: VLSP2018-ABSA-Hotel
• ViSoLex Toolkit