Performance evaluation of machine learning and deep learning approaches for sentiment analysis on COVID-19 sentiments

Abstract

Sentiment Analysis (SA) is a study of people's opinions on products, services, organizations, disastrous events etc. The biggest challenge in SA is the text and the context in which it is written or said, for sentiments vary with context. Recently, existing studies have proposed different ML and DL models to classify the data. However, there are still challenges in dealing with unstructured text, classification; preprocessing, and good accuracy measures are ongoing problems. Researchers in the fields of psychology and sociology have focused a lot on deciphering people's emotional expressions during the pandemic. In this paper, we analyse the people's sentiments posted during the COVID-19 pandemic; a real scenario has been used to validate the effectiveness of the proposed work. The proposed approach uses different feature sets and classifiers to analyse the collected tweets and classify them into positive, negative, neutral, extremely positive, and extremely negative sentiments. The proposed model was trained and tested using ML and DL algorithms like Naive Bayes (NB), Random Forest (RF), Multilayer Perceptron (MLP), Passive Aggressive Classifier (PAC), Support Vector Classifier (SVC), Logistic Regression (LR), Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) for SA.

We use evaluation measures (like accuracy, precision, recall, and F1-score) to rate the success of machine learning (ML) and deep learning (DL) classifiers.  To measure the performance, the data set was split in the ratio of 70:30, 80:20 and 90:10 for training and testing the model and observed that LSTM outperforms with the highest 88% accuracy, 88% precision, 87% recall, and 87% F-1 score. According to the results, the proposed model could detect people’s sentiments and allow domain practitioners to analyze sentiments efficiently in decision-making for public health communication strategies.