A Comparative Study of Transformer-Based Models and Machine Learning Techniques for Enhancing Bangla Sentiment Analysis
Md. Saimim Islam Khan Hamim, Ishtiaque Ahmed, Tawhidul Islam Sazid, Faria Haque, Md. Najimul Hossain, Suborna Rani Pal, Mst. Sazia Tahosin, Md. Alif Sheakh · 2025
Social media monitoring alongside business intelligence needs accurate sentiment analysis because the digital activity of Bangla-speaking communities keeps expanding. The current classification models do not deliver effective Bangla context analysis mainly because of insufficient annotated datasets when processing rare languages. The research examines sentiment classification through the combination of deep learning transformers, Flair, TARS (Task-Aware Representation of Sentences), and GPT-2 (Generative Pre-training Transformer) with traditional machine learning algorithms such as LR (Logistic Regression), SVC (Support Vector Classifier), RF (Random Forest), XGB (XGBoost), DT (Decision Tree), KNN (K-Nearest Neighbors), and NB (Naive Bayes). During tests on our Bangla dataset, the Flair model implementing Bangla BERT Base embeddings produced the best performance of$\mathbf{9 4. 2 1 \%}$accuracy, which surpassed both TARS at$\mathbf{8 8. 5 \%}$and traditional classifiers. Embedded into contexts selects complex sentiment patterns efficiently, which proves integral to boosting low-resource NLP (Natural Language Processing) task performance. For better sentiment classification in practical settings, researchers will concentrate on enhancing the efficiency and diversity of datasets and implementing crosslingual capabilities.