A Comprehensive Analysis of Bengali Sentiment Exploration Using Transformer Block I and Transformer Block II

The internet is responsible for the constant modernization of the planet. Anything is visible to everyone on social media. We work with sentiment or opinion mining in this study. Sentiment comes in two flavors: good sentiment and bad emotion. Moreover, there are numerous works in several languages, but the amount of work in Bengali emotion is restricted by the use of transformers. As a result, we concentrate on transformer blocks I and II's Bengali sentiment investigation. We gather information from Kaggle using the following datasets: Bengali Review (dataset 3), Conversion (dataset 2), and Predict Unsupervised (dataset 1). There are 19600 Bengali feelings in total. Various transformer techniques, include a lite Bert for self-supervised learning, a distilled version of BERT (DistilBERT), and bidirectional encoder representation from transformer (BERT). We employ long-short-term memory (LSTM), gated recurrent units (GRU), and recurrent neural networks (RNN) as examples of deep neural networks. While all the datasets function well and efficiently, the Bengali Review dataset with Transformer II operates at a very high speed. This is demonstrated by the following metrics: accuracy, precision, recall, F1-score, and Roc score are 93.91% in LSTM with DistilBERT, 97.43% in LSTM with ALBERT, 97.15% in LSTM with BERT, 95.77% in LSTM with DistilBERT, and 93.41% in LSTM with ALBERT, respectively. Sentiment analysis greatly benefits from error exploration. The model functions more effectively when there are fewer mistakes. The LSTM using DISTILBERT yields the following error values: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Square Log Error (MSLE), respectively: 0.060, 0.246, and 0.029." All things considered; transformer block II offers the finest analysis.