An Empirical Study on Developing Stacking Ensemble Model for Bangla Sports Sentiment Analysis

Abstract

Bangladeshi people have a strong connection with sports, evident in the multitude of comments made before and after games regarding their favorite teams. Sports sentiment analysis is conducted to understand public sentiments in the realm of sports. In this work, we proposed a new sports based sentiment dataset extracted from YouTube comments covering football, cricket, traditional local games and others, totaling 82,612 instances which is publicly available in the repository https://github.com/cseku170202/Bangla-Sports-Sentiment-Analysis. Our empirical study considered the use of various trendy machine learning (ML) and deep learning (DL) methods. Specifically, we utilized Convolutional Neural Network (CNN) and Transformer Neural Network (TNN) as deep learning methods and Multinomial Naive Bayes (MNB), Random Forest (RF), and KNearest Neighbors (KNN) as machine learning techniques. Furthermore, we examined 48 different stacking ensemble models by combining these machine learning and deep learning algorithms. Subsequently, our proposed stacking model achieved the highest accuracy of 81%, demonstrating its capability to identify sports sentiments with remarkable F1-score of 0.82. To explain model’s interpretability, we utilized several Explainable AI techniques to better understand the impact of the base model in the stacking model output.