A comparative approach on multiple neural network models aiming to perform a sentiment classification on tweets related to COVID-19 pandemic

This study uses neural networks to analyze COVID-19 tweets, classifying sentiment as positive, neutral, or negative to reveal insights. The dataset of 40,000 tweets taken from Twitter in the period from March to April 2020 was subjected to preprocessing wherein text-cleaning and word-embedding methods were applied. Six deep learning models consisting of one- and two-layer Dense Neural Networks, one- and two-layer CNNs, a Convolutional Long Short-Term Memory (ConvLSTM) scheme and a Dense LSTM model were trained and tested by categorical cross-entropy loss and the RMSProp optimizer. Besides using a single measure, the authors resorted to a battery of metrics accuracy, precision, recall, and AUC to get a fuller picture of their models’ behavior. In comparison to the baseline and other models, the ConvLSTM model was leading its performance with an accuracy of 0.8572, a precision of 0.8689, a recall of 0.8424, and an AUC of 0.8894. The paper’s empirical evidence serves as a demonstration of the deep-learning paradigm’s efficacy, especially the hybrid convolutional-recurrent architectures, for sentiment analysis. The suggested model has definite merits in helping to gauge the public’s mindset during a health crisis and can be utilized to look at different areas of social sentiment monitoring and policy evaluation.