Sentiment Analysis using Bidirectional Encoder Representations from Transformers

In the contemporary digital landscape, a significant volume of data is generated through social networks such as Twitter, Facebook, and Instagram. This study presents a method for extracting sentiments from Twitter, focusing on two sentiment-based datasets: the Twitter and emotional sentiments datasets. After extraction and preprocessing, we employed three deep learning models: Recurrent Neural Networks (RNNs), Bidirectional Long Short-Term Memory (BiLSTM), and a pre-trained Bidirectional Encoder Representations from Transformers (BERT) model. We introduced Se-BERT, a model designed for emotional sentiment analysis. Our experiments showed that Se-BERT achieved accuracy levels of 97.29% for tweet sentiments (positive and negative) and 86.77% for emotional sentiments (joy, sadness, love, fear, anger, surprise). These results demonstrate that Se-BERT outperforms RNN and BiLSTM in terms of accuracy for sentiment analysis, thereby significantly enhancing information retrieval and providing a deeper understanding of user behaviour.