Analyzing the Performance of BERT for the Sentiment Classification Task in Bengali Text

The recent era has seen significant growth of technologies in the field of Natural Language Processing (NLP). But the scarce resource languages like Bengali have not got much attention from the research community. The BERT language model has laid a very positive impact on the performance of the NLP tasks. Although several others language models came into the scenario, we investigate the performance of BERT model and other conventional methods for the sentiment classification task in Bengali text. The obtained result shows that BERT overperformed other conventional machine learning and lexicon-based methods in all aspects of the performance metrics. Along with BERT, conventional methods namely Logistic Regression, Decision Tree, SVM, Random Forest, Naive Bayes and Neural Network were implemented. Besides these methods a lexicon-based approach was used to see the overall variation in the results. The lexicon resource for Benali was created for this implementation.