Text Sentiment Analysis Based on Transformer and Augmentation.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Social Media Analysis

Background:

• Social media platforms are integral to modern life, generating vast amounts of public opinion data.
• Sentiment analysis of this data is crucial for understanding and maintaining societal stability.
• Existing advanced models like BERT are computationally expensive, limiting their widespread application.

Purpose of the Study:

• To develop an efficient sentiment analysis model suitable for large-scale public opinion analysis.
• To address the computational resource limitations of current state-of-the-art natural language processing models.
• To enhance the performance of sentiment analysis, particularly in few-sample learning scenarios.

Main Methods:

• A novel transformer-based model incorporating knowledge distillation to reduce model size and computational requirements.
• Text augmentation techniques to expand the training dataset and improve generalization.
• Implementation of knowledge distillation to decrease model parameters, computational cost, and training time.

Main Results:

• The proposed model achieves competitive results in sentiment analysis tasks.
• Demonstrated effectiveness in few-sample sentiment analysis due to text augmentation.
• Significant reduction in computational cost and training time compared to large models.

Conclusions:

• The developed transformer-based model offers an efficient and effective solution for sentiment analysis.
• Knowledge distillation and text augmentation are key components for improving model efficiency and performance.
• This approach makes advanced sentiment analysis more feasible for real-world applications with limited resources.