Stacked-KBR: An Ensemble Method for Analyzing YouTube Bangla Song Reviews Using a Novel Dataset

Song review analysis is very crucial in every language due to its widespread use in the cinema and music industry. The success of many movies depends on having some hit and blockbuster songs in the movie. From this one can understand the importance of developing automated methods for analyzing song reviews. This work is devoted to deploying a method for YouTube Bangla song review analysis by developing a new dataset of 43,264 comments annotated as either positive or negative, covering over 100 songs and 15 music genres which is publicly available in the repository https://github.com/cseku170202/Bangla-Song-Sentiment-Analysis. The prime goal of this investigation is to employ various machine learning algorithms to precisely recognize the sentiments expressed in Bangla song reviews and apply explainable AI to provide explanations. Our empirical study considered several algorithms including the K-Nearest Neighbour (KNN), Bidirectional Long Short-Term Memory (Bi-LSTM), Recurrent Convolutional Network (RCN), and a Stacking ensemble model (Stacked-KBR) which utilizes the previous three models as base classifiers. Among these models, the Stacked-KBR outperformed others with an accuracy of 83.98%. Local Interpretable Model-agnostic Explanation (LIME) and SHapley Additive exPlanation (SHAP) from explainable AI are also considered in this work to observe the reason behind the model’s performance.