Abstract
This research investigates the performance of various machine learningalgorithms (CNN, LSTM, VADER, and RoBERTa) for sentiment analysis of Twitterdata related to imported food items in Trinidad and Tobago. The study addressesthree primary research questions: the comparative accuracy and efficiency ofthe algorithms, the optimal configurations for each model, and the potentialapplications of the optimized models in a live system for monitoring publicsentiment and its impact on the import bill. The dataset comprises tweets from2018 to 2024, divided into imbalanced, balanced, and temporal subsets to assessthe impact of data balancing and the COVID-19 pandemic on sentiment trends. Tenexperiments were conducted to evaluate the models under various configurations.Results indicated that VADER outperformed the other models in both multi-classand binary sentiment classifications. The study highlights significant changesin sentiment trends pre- and post-COVID-19, with implications for importpolicies.