ViGSA: Multi-Task Aspect-Based Sentiment Analysis with Auxiliary Embeddings & Global Sentiment Integration for Vietnamese Restaurant Reviews

👉 If you find this project interesting and useful, please give it a star ⭐ to support us!

Authors

Dien X. Tran, Kien-Cao Van, Tinh-Nguyen Huu, Hoang-Tuan Dao-Xuan, Hung-Nguyen Viet, Khanh-Duy Cao-Phan

---

📌 Overview

ViGSA (Vietnamese Global Sentiment Auxiliary) is a multi-task model for Vietnamese Aspect-Based Sentiment Analysis (ABSA). The model jointly performs:

1. Aspect Detection
2. Aspect-level Sentiment Classification
3. Global Sentiment Prediction

Key innovations:

• Auxiliary Aspect Embeddings from natural language descriptions
• Global Sentiment Supervision to improve aspect-level understanding
• Multi-loss Training (Cross-Entropy + Asymmetric + Contrastive Loss)

Results: Achieves state-of-the-art performance on UIT-ABSA and VLSP-2018 datasets, proving the effectiveness of integrating semantic priors with multitask learning in low-resource Vietnamese ABSA.

---

✨ Features

• Multi-task Learning → Aspect detection + Sentiment classification + Global sentiment
• Auxiliary Semantic Embeddings → Improves aspect representation
• Multi-domain Support → Trained and evaluated on Restaurant domain (UIT-ABSA, VLSP-2018)
• Transformer-based Encoder → Supports InfoXLM, PhoBERT, ViT5, RemBERT
• Contrastive Learning → Better representation via supervised contrastive loss
• Attention Mechanisms → Multi-head attention for feature extraction

---

📂 Project Structure

ViGSA/
├── main.py                 # Entry point for training/evaluation
├── requirements.txt        # Python dependencies
├── train.sh                # Training script
├── dataset/                # Dataset processing modules
├── evaluation/             # Evaluation metrics/utilities
├── model/                  # Model architecture & losses
│   ├── train.py             # Model training logic
│   └── loss.py              # Custom loss functions
└── processing/              # Data preprocessing utilities

---

⚙️ Installation

git clone https://github.com/xndien2004/ViGSA
cd ViGSA
pip install -r requirements.txt

Requirements

• Python ≥ 3.7
• PyTorch ≥ 1.10
• PyTorch Lightning ≥ 1.5
• Transformers ≥ 4.10
• CUDA-enabled GPU (recommended)

---

📊 Dataset

• UIT-ABSA (Restaurant domain)

• VLSP-2018 ABSA (Restaurant domain)

• Supports multi-task data format:

  • Train.txt, Dev.txt, Test.txt

---

🚀 Usage

Train with default parameters

bash train.sh

Train with custom parameters

python3 -m absa_project.main \
    --train_file "path/to/train.txt" \
    --val_file "path/to/dev.txt" \
    --test_file "path/to/test.txt" \
    --model_name "microsoft/infoxlm-large" \
    --topk_layer 4 \
    --batch_size 40 \
    --epochs 50 \
    --learning_rate 2e-5 \
    --output_dir "output/experiment_name" \
    --save_top_k 1 \
    --patience 10 \
    --max_length 256

---

🏗 Model Architecture

1. Encoder → Transformer-based (InfoXLM, PhoBERT, ViT5, RemBERT)
2. Multi-head Attention → Enhanced feature representation
3. Aspect Detection → Binary classification per aspect
4. Aspect Sentiment Classification → 4-class per aspect
5. Global Sentiment Classification → Overall sentiment prediction
6. Contrastive Learning → Improves inter-class separation

Loss Functions:

• Aspect Detection → Asymmetric Loss
• Aspect Sentiment → Asymmetric Loss
• Global Sentiment → Cross-Entropy Loss
• Representation Learning → Supervised Contrastive Loss

---

📈 Evaluation

Metrics:

• Aspect Detection → Precision, Recall, F1
• Sentiment Classification → Precision, Recall, F1
• Overall → Combined task metrics

---

🏆 Results

• SOTA performance on UIT-ABSA & VLSP-2018
• Detailed logs & checkpoints stored in output/ directory

---

📚 Citation

@misc{dien2025vigsa,
  title={ViGSA: A Multi-Task Aspect-Based Sentiment Analysis Model with Auxiliary Embedding and Global Sentiment Integration for Vietnamese Restaurant Reviews},
  author={Tran, Dien X. and Van, Kien-Cao and Huu, Tinh-Nguyen and Dao-Xuan, Hoang-Tuan and Viet, Hung-Nguyen and Cao-Phan, Khanh-Duy},
  journal={Information Processing & Management (under review)},
  year={2025}
}