Abstract: Customers' reviews are laden with emotions and therefore present a challenge from data analytical perspectives. When looking at more than three data categories, this analysis becomes even more complicated. The English language does have access to a wide variety of data and advanced pre-trained models. However, the same cannot be said for the Bangla language, which does not have large documented datasets nor accurate methods of data labelling due to being a low-resource language. The goal of this study is to fulfil this gap by proposing a novel model for the classification of Bangla reviews into the following five categories concerning emotions: Negative(0), Positive(1), Neutral(2), Slightly Negative(3), and Slightly Positive(4). 26,028 Bangla reviews were collected from famous e-commerce platforms, annotated by 3 annotators. Inter-annotator reliability was strong (Cohen's κ=0.8; Fleiss' κ=0.79)and processed through a Bangla-focused normalization, which included language filtering, script correction, tokenization, removal of stop-words, and light morphological reduction. Classical TF-IDF features and transformer-based subword embeddings used to represent the text. Six classical ML algorithms (MNB, LR, SVM, RF, KNN, and Decision Tree) and three transformer models (BanglaBERT, RoBERTa, Sentence-BERT) were evaluated under the same setup. Among all models, SVM and Random Forest reached highest 95% accuracy, while RoBERTa achieved 84%. Model outcomes indicate that, even with modern architectures available, optimized classical models still offer strong performance. For evaluation transparency, we report also precision, recall, and F 1, and we describe an additional robustness protocol beyond a single random split to reduce the risk of split-dependent conclusions.
