A semi-supervised approach in detecting sentiment and emotion based on digital payment reviews

This paper investigates the sentiment and emotion of digital payment application consumers using a hybrid approach consisting of both supervised and unsupervised machine learning techniques. Support vector machine, random forest and Naive Bayes were modeled for sentiment and emotion analyses, whereas latent Dirichlet allocation was administered to identify top emerging topics based on English textual reviews from three digital payment applications. Random forest produced the best results for sentiment (F1 score = 73.8%; Cohen's Kappa = 52.2%) and emotion (F1 score = 58.8%; Cohen's Kappa = 44.7%) analyses based on a tenfold cross-validation. Latent Dirichlet allocation revealed best clusters atk = 5 and items = 25, with the top topics being App Service, Transaction, Reload Features, Connectivity and Reward. Findings are presented and discussed in general and also based on each application.