Improving sentiment classification using a RoBERTa-based hybrid model.

Area of Science:

• Natural Language Processing
• Deep Learning
• Machine Learning

Background:

• Traditional sentiment analysis methods face challenges with complex language nuances.
• Existing deep learning models have limitations in capturing contextual semantics.

Purpose of the Study:

• To develop a hybrid deep learning model for improved sentiment classification.
• To leverage the strengths of transformer and sequence models while mitigating their weaknesses.

Main Methods:

• A hybrid model integrating Robustly Optimized BERT (RoBERTa) with Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM).
• Utilized RoBERTa for sentence vector representation and CNN+LSTM for semantic comprehension.
• Employed SMOTE technique with word embeddings to address class imbalance in the Twitter dataset.

Main Results:

• Achieved 96.28% accuracy on the IMDb movie reviews dataset.
• Attained 94.2% accuracy on the US airlines Twitter reviews dataset.
• Demonstrated superior performance compared to standard sentiment analysis techniques.

Conclusions:

• The proposed hybrid RoBERTa-(CNN+LSTM) model is highly effective for sentiment classification.
• This approach successfully enhances the understanding of semantics and context in text data.