A GAN-Driven Transformer-based Approach for Sentiment Analysis of Social Media Posts

Recently, there has been an influx of interest in sentiment analysis because of its many uses in various sectors and research domains. Understanding sentiment is vital for market research, public opinion analysis, and social impact assessment. However, existing sentiment analysis tools are primarily designed for high-resource languages (HRLs), limiting their efficacy in low-resource languages (LRLs), including Bengali. This work proposes a comprehensive framework that leverages the Generative Adversarial Network (GAN) to handle dataset imbalance, along with transformer-based models, to classify sentiment into positive, negative, and neutral categories. Machine learning (ML), deep learning (DL), and transformer-based baselines with GAN models are exploited to classify sentiment in Bengali social media posts. The evaluation results reveal that the GAN+Bangla-BERT-1 model achieved the highest micro F1-score of 70.52.