The Feasibility of Using Synthetic-Generated Dataset for Training Sentiment Analysis Model

New features in machine learning applications are constantly being explored demanding the availability and quality of data. To address this, the exploration of synthetic data emerges as a promising solution. This paper intends to solve the problem of data availability by finding out the feasibility of using synthetic data as a solution. Many researchers tried to use algorithms such as SMOTE and ADASYN, but encountered challenges in producing semantically and grammatically acceptable material. One of the few front-runners in synthetic text data generation is GPT 3.5. A trained sentiment analysis model would benefit users greatly by providing valuable insights into given text data, helping to understand emotions, opinions, and trends. The criteria of a high-quality dataset is providing various examples for each class. High-quality training data is crucial for accurate sentiment analysis, ensuring correct patterns, bias minimization, and reliable predictions. This paper utilizes GPT 3.5 to generate the synthetic data, with BERT and CNN-LSTM as the model architectures. Each architecture employs three models, each with a unique train-test combination that serves its own purpose. The performance of the six trained models is compared based on accuracy, precision, recall, and F1 score to evaluate the quality and efficacy of synthetic data. The conducted evaluation shows that data augmentation using synthetic data is feasible. This finding is presented as the main contribution of the study. However, fully synthetic data-trained models require more refining.