An Interpretable Sentiment Recognition Method from Political Bangla Texts Using Stacking Ensemble Model

Sentiment analysis is a natural language processing method employed to categorize text according to the sentiment conveyed, which may be positive, negative, or neutral. Political sentiment classification is essential for any language and nation, as a country’s current and future stability is directly contingent upon it. This study uses the Stacking ensemble model to propose an interpretable approach for political sentiment classification in Bangla. A novel text dataset, BangPolSenti, has been developed comprising 34,235 comments from YouTube and it is available in the repository https://github.com/cseku170202/Bangla-Political-Sentiment-Analysis. The suggested Stacking model (Stacked-LCR) integrates Logistic Regression, Capsule Network, and Recurrent Neural Network as the base models, with Random Forest as the meta-model. The Stacked-LCR model demonstrates superior performance, achieving an accuracy of 99.18% in three sentiment categories and 99.47% in two sentiment categories. Experimental findings are analyzed using the Friedman statistical test at a significance threshold of 0.05, with all test outcomes being significant. Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) methodologies are utilized to visualize the principal features of classification.