Real-time topic-based sentiment analysis for movie tweets using hybrid approach

This paper aims to perform sentiment analysis (SA), which deals with dissecting and then extracting the hidden insights from the sentences said or written by a person. The proposed methodology for SA uses a combined and novel framework involving a Lexicon-based approach (LBA) and a deep learning (DL) technique to predict the overall sentiment of the text. Firstly, the LBA segregates the neutral tweets from the polar tweets. Later, a variant of CNN, namely a BERT-based bidirectional long short-term memory-temporal convolutional network (BiLSTM-TCN) grounded by the residual module and the dilated convolutions, is used to identify the type of polarity of the text. The paper also investigates various other BERT-based models like CNN, LSTM, and BiLSTM on the IMDB Movie Review dataset containing 50k movie reviews to show that the suggested model achieves a mean validation accuracy of 0.932 and a mean validation loss of 0.238 for the last three epochs for the polar reviews. After that, the trained model is used to forecast the sentiment of the live-streaming tweets about a particular movie of interest. The Twitter API, Tweepy, fetches tweets crawling over Twitter. The study obtains test accuracy, F1, and ROC-AUC scores of 92.94%, 0.929, and 0.98, respectively, in the least number of epochs, which is better than other models mentioned above, running in the same environment.