Analyzing Bangla Drama and TV Series Reviews Using Stacking Method

Abstract

Sentiment analysis (SA) of drama and TV series evaluations is currently popular in many languages due to its various innovative applications and uses. However, it is unfortunate that there has been no progress in the field of sentiment analysis in the Bengali language so far. In this study, we have created a new extensive collection of 20,000 reviews from 93 dramas and 28 TV series on YouTube, which is publicly available in (https://github.com/cseku170202/Bangla-Drama-and-TV-Series-Sentiment-Analysis).The primary aim of this study is to utilize diverse models to precisely detect the sentiments conveyed in evaluations of Bangla dramas and TV series. Additionally, we will employ explainable AI techniques to provide insights into the reasons behind the high or low performance of these models. Out of the implemented algorithms, Support Vector Machine (SVM) achieved the maximum accuracy of 73.28% in machine learning. Convolutional Neural Network - Gated Recurrent Unit (CNN-GRU) obtained an accuracy of 79.83% in deep learning. The stacking ensemble model achieved the highest accuracy of 82.95%. The Friedman statistical test is run to ascertain the statistical relevance of the obtained results; the test results show to be statistically significant at a significance level of 0.05. This work analyzes the performance of models at both local and global levels using Local Interpretable Model-agnostic Explanation (LIME) and SHapley Additive exPlanation (SHapley Addition from explainable AI).