Developed a sophisticated multi-class sentiment analysis system using BERT, achieving 80.69% accuracy on aspect-based opinion classification. The project successfully addressed class imbalance challenges through weighted loss functions and optimised model architecture, demonstrating robust performance in categorising opinions as positive, negative, or neutral.

Overview

The project tackles the complex challenge of aspect-based sentiment analysis through implementation of advanced transformer architectures. Using a dataset of 1,503 training samples and 376 development samples, the system analyses sentence-aspect-category triplets to determine sentiment polarity. The research particularly focused on addressing significant class imbalance issues, with neutral sentiments comprising only 4% of the dataset, through innovative weighting strategies and architectural optimisations.

Technologies

• Development Environment: Python 3.9
• Data Processing: Pandas, NumPy
• Machine Learning: Scikit-learn
• Deep Learning Framework: PyTorch, PyTorch Lightning
• NLP Libraries: Transformers, NLTK, Stanza
• Pre-trained Models: BERT (bert-base-uncased)
• Dataset Management: Hugging Face Datasets
• GPU Acceleration: CUDA
• Tokenisation: SentencePiece
• Performance Monitoring: TensorBoard

Implementation Details

1. Model Architecture

• Implemented BERT-based classifier with custom layers
• Developed weighted cross-entropy loss for class imbalance
• Created dropout layers for regularisation
• Implemented softmax-based classification head
• Optimised model for multi-class prediction

2. Data Processing

• Created efficient data loading pipeline
• Implemented aspect-category concatenation with [SEP] tokens
• Developed tokenisation framework for input processing
• Created balanced batch sampling strategy
• Implemented dynamic padding for efficiency

3. Training Framework

• Implemented AdamW optimiser with learning rate 5e-5
• Developed epoch-based training schedule
• Created comprehensive evaluation pipeline
• Implemented GPU acceleration support
• Developed model checkpointing system

4. Performance Optimisation

• Implemented class weighting (1:3:20 ratio)
• Created efficient batch processing system
• Developed early stopping mechanism
• Implemented learning rate scheduling
• Created robust evaluation metrics

Analysis Features

• Multi-class sentiment distribution analysis
• Model performance tracking
• Class-wise accuracy assessment
• Training convergence monitoring
• Aspect category impact analysis
• Prediction confidence scoring
• Error pattern investigation
• Performance visualisation
• Cross-validation results
• Resource utilisation monitoring

Key Results

• Achieved 80.69% accuracy on development dataset
• Successfully handled severe class imbalance
• Optimised training time to 9 minutes with GPU
• Processed 1,503 training samples effectively
• Implemented robust 3-class classification
• Established balanced performance metrics
• Demonstrated aspect-based analysis capability
• Achieved consistent cross-run performance

Skills Gained

• Advanced NLP model implementation
• Deep learning architecture design
• Class imbalance handling techniques
• GPU computing optimisation
• Machine learning pipeline development
• Model evaluation and metrics analysis
• Transfer learning implementation
• Performance optimisation strategies
• Multi-class classification techniques

Impact

The system provides robust sentiment analysis capabilities for aspect-based opinion mining, particularly valuable for customer feedback analysis and market research. The methodology demonstrates effective handling of class imbalance in natural language processing tasks, offering insights for similar multi-class classification challenges. The project’s approach to combining aspect categories with textual content provides a foundation for more nuanced sentiment analysis applications.