Emotions in Text - Enhanced Sentiment Analysis Using Stacked Ensemble and Custom Threshold

In recent years, sentiment analysis has gained predominance due to its applications in various domains, including social media monitoring, customer feedback analysis, market research, and reputation management. As a result, many tools have been developed using lexicon-based approaches, machine learning techniques, and deep learning models each of which has been structured to a particular type of textual content. This study is based on the presumption that sentiments are related with the underlying emotional intensity reflected in polarity of an utterance which is the basic building block of textual sentiment analysis. In this study we compare the performance of existing sentiment analysis tools and models on established data-sets. We find that existing sentiment-analysis tools are subjective, poorly correlated and not very accurate. We therefore propose a stacked ensemble architecture and demonstrate that it outperforms existing state-of-the-art models for predicting the polarity of an utterance by an accuracy score of 1% to 12% depending upon the dataset and configuration. We also demonstrate that by reducing the sensitivity of our model, we can pick up highly polarised sentiments with greater accuracy for extracting intense emotions. Our approach requires minimal training and computing resource, and also does not require any pre-processing, and can be used in modified forms for many relevant use cases.