A multi-class classification approach for feminist sentiment analysis in Bangla social media using TF-IDF and ensemble learning

Abstract

Social media has emerged as an important part of societal discourse on feminism and gender equality, especially in Bangladesh. Nevertheless, any feminist debate on social media in Bengali polarizes reactions, highlighting the need for automated sentiment analysis. This paper introduces one of the earliest multi-class feminist sentiment classification schemes of the Bengali social media with a manually annotated dataset of 6,830 comments categorized as positive, neutral, or negative. The framework uses term frequency-inverse document frequency (TF-IDF) based n-gram feature representations utilizing traditional machine learning algorithms, with a majority voting ensemble to determine optimal robust models. The data was divided into 80% and 20% for training and testing, respectively. Models were evaluated on the basis of accuracy, precision, recall, and macro-F1 to correct on imbalance of classes. Multinomial naive bayes (MNB) has the best accuracy of 84.74% and macro-F1 of 84.66, which is 4-7 times higher than other models. The ensemble method improved feature strength. Such results indicate that lightweight machine learning models based on TF-IDF features and ensemble models can be useful to detect feminist sentiment in Bangla social media and serve as a guideline in the field of domain-specific sentiment analysis in low-resource languages and help monitor online feminist discourse.