Aspect-Based Sentiment Analysis in Vietnamese - Group 12

Overview

This project focuses on Aspect-Based Sentiment Analysis (ABSA) for Vietnamese hotel and restaurant reviews, leveraging the VLSP 2018 dataset. The goal is to predict sentiments for specific aspect categories (e.g., "cleanliness", "service quality") and provide actionable insights for businesses in the Vietnamese market.

We explore and evaluate multiple approaches, including:

• PhoBERT-BiLSTM Single-task
• PhoBERT-Multi-task
• PhoBART-Generation
• Hierarchical Graph Convolutional Network (Hier-GCN)

This project is part of the Natural Language Processing course at Hanoi University of Science and Technology.

Key Features

• Aspect Detection: Identifies service aspects in reviews (e.g., ROOMS#COMFORT).
• Sentiment Classification: Classifies sentiments for each aspect as Positive, Negative, or Neutral.
• Advanced NLP Models:
  • Pretrained Vietnamese language models such as PhoBERT and BARTPho.
  • Hierarchical Graph Convolutional Network for modeling inter-category and category-sentiment relationships.
• Dataset: Utilizes the VLSP 2018 dataset, containing annotated hotel and restaurant reviews in Vietnamese.

Models and Methods

1. PhoBERT-BiLSTM Single-task

• Combines PhoBERT embeddings with a BiLSTM-CNN architecture.
• Optimized using Focal Loss and Cross-Entropy Loss.
• Implements separate models for aspect detection and sentiment classification.

2. PhoBERT-Multi-task

• Employs Multitask Learning (MTL) to simultaneously predict sentiments for multiple aspects.
• Uses a shared representation of contextual embeddings from PhoBERT.

3. PhoBART-Generation

• Treats ABSA as a text generation task using BARTPho.
• Outputs natural language sentences, e.g., "The sentiment polarity of ROOMS#COMFORT is Positive".
• Highly adaptable to new domains and unseen aspect categories.

4. Hier-GCN

• Models ABSA as a hierarchical prediction task.
• Captures relationships between categories and sentiments using graph convolutional layers.

Installation

Prerequisites

• Python 3.8+
• PyTorch
• Hugging Face Transformers
• Additional libraries:
  • TensorFlow
  • scikit-learn
  • PyVnCoreNLP