A deep learning approach in predicting products' sentiment ratings: a comparative analysis.

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Sentiment analysis is crucial for understanding public opinion and user feedback.
• Traditional machine learning methods face challenges in capturing complex linguistic nuances.
• Deep learning models offer advanced capabilities for text representation and analysis.

Purpose of the Study:

• To benchmark deep learning models for text sentiment analysis.
• To compare various word embedding techniques, including BERT variants, FastText, and Word2Vec.
• To evaluate model performance with and without data augmentation and across different classification setups (5-class vs. 3-class).

Main Methods:

• Implemented and compared Convolutional Neural Networks (CNN), Recurrent Neural Network (RNN), and Bi-directional Long Short Term Memory (Bi-LSTM) models.
• Utilized word embedding approaches: Word2Vec, FastText, and Bi-directional Encoder Representations from Transformers (BERT) and its variants (e.g., RoBERTa).
• Applied Easy Data Augmentation and evaluated models on original and augmented datasets in 5-class and 3-class configurations.

Main Results:

• The CNN-RNN-Bi-LSTM model with Word2Vec embeddings achieved the highest accuracy (96%) and F-score (91.1%).
• Individually, RNN demonstrated strong performance with 87.5% accuracy and 83.5% F-score.
• RoBERTa achieved the highest F-score of 73.1% among BERT variants, indicating effectiveness in specific scenarios.

Conclusions:

• Deep learning models, particularly neural networks with Word2Vec, significantly outperform traditional supervised machine learning for text sentiment analysis.
• Data augmentation and model architecture choices impact overall performance.
• The study provides a foundation for future research in advanced sentiment analysis techniques.