Towards Multi-class Sentiment Analysis With Limited Labeled Data

Analyzing public sentiment about an entity or issue can be of interest to governments and businesses alike. There is a growing body of research that attempt to devise new sentiment analysis techniques, especially techniques based on machine learning. These machine learning-based techniques typically require large, labeled training data with a large number of instances for training in order to provide reasonable accuracy in sentiment analysis. However, labelling large volumes of data is tedious and expensive. In this paper, we propose a multi-class sentiment analysis technique, named SG-Elect, utilizing cutting-edge transformer based pre-trained models along with more traditional machine learning based approaches in a semi-supervised setting. Our experiments demonstrate that SG-Elect outperforms a recent state-of-the-art baseline for all three datasets.