USING SYNTHETIC TEXT DATA TO TRAIN SENTIMENT ANALYSIS MODELS

The explosive growth of digital text data presents many opportunities and challenges, especially in the area of sentiment analysis. A common problem is the imbalance of datasets, where one sentiment class significantly outperforms others, resulting in prediction bias. In this study, the class imbalance problem is addressed by creating synthetic data using GPT-2, a state-of-the-art generative language model. Comprehensive experiments were conducted using logistic regression, long term short-term memory (LSTM), and bidirectional encoding representations from transformers (BERT) models. The evaluation showed significant results, in particular BERT, which achieved accuracies up to 74.2% when trained solely on synthetic data. This highlights the value of using synthetic data in NLP applications. The study suggests opportunities for future research using more advanced language models and broader NLP applications.