From Trees to Transformers: Advancement in Bengali Sentiment Classification on BANEMO with mBERT and Deep Learning

In this study, we present BANEmo, a novel Bangla language dataset designed specifically for sentiment analysis tasks which is derived from a diverse range of Bangla text/comments, harvested from both social media and newspapers, encompassing a total of 15,000 individual comments. Each comment in the data set is meticulously labeled for various sentiments, such as happiness, sadness, disgust, anger, and more. In our investigation, we initially focus on a binary classification of sentiments, specifically happiness and sadness, where two classical machine learning algorithms, called Support Vector Machine (SVM) and Decision Tree (DT), were employed to establish a baseline performance on the Banemo dataset. In particular, these conventional models achieved state-of-the-art accuracy levels for binary Bangla sentiment analysis. To further push the boundaries, we employed mBERT, a transformer-based pre-training method, for sentiment analysis which exceeded the benchmarks set by the traditional models, underlining the potential of transformer-based models in Bangla sentiment analysis. Finally, this study aims to stimulate further advancements in the field of Bangla natural language processing and sentiment analysis.