A Comparison of Machine Learning Algorithms and Transformer-based Methods for Multiclass Sentiment Analysis on Twitter

The growing number of tweets being produced on Twitter and other social media platforms has made sentiment analysis a crucial task in comprehending individuals' viewpoints and emotions. We focus on the challenge of Twitter emotion classification in this paper and perform sentiment analysis on six different emotions: joy, anger, sadness, fear, surprise, and disgust. The performance of four distinct models is studied as this topic is addressed as a multiclass classification challenge: Naive Bayes, logistic regression, SVM, and DistilBERT, a transformer-based model. Experiments are run on a dataset of over 16,000 tweets that have been annotated with six emotions: joy, anger, sadness, fear, surprise, and disgust. DistilBERT outperforms the other models with an overall accuracy of 93.5%, according to our results. The accuracy of Naive Bayes is 78%, the accuracy of logistic regression is 89%, and the accuracy of SVM is 88%. Our research demonstrates the utility of transformer-based models, specifically DistilBERT, in sentiment analysis tasks using Twitter data. The findings also highlight the significance of appropriate pre-processing techniques in the performance of traditional machine learning models. Our findings have practical implications for businesses looking to gain insights into customer opinions and preferences via social media platforms such as Twitter.