Abstract:
Natural language processing's subfield of sentiment analysis involves locating and categorizing the feelings, viewpoints,
and attitudes expressed in text. Because it enables us to understand public opinion on a variety of topics, sentiment analysis has grown
in importance as social media platforms become more widely used. In this research paper, we used two deep learning models, BERT
and RoBERTa, and their fusion of both architectures to perform sentiment analysis on a dataset of tweets related to the dataset of
COVID-19 pandemic. To eliminate noise and unrelated data, the dataset underwent pre-processing and cleaning. Then, using the
dataset, we trained the BERT and RoBERTa models and assessed their performance. Both models achieved high F1 scores, recall, and
accuracy for all three sentiment classes (negative, neutral, and positive) for sentiment analysis. While there were some differences in
how well these models performed across these metrics, both models did well and classified the sentiment of tweets in the dataset with
high accuracy. Our study's findings show how well BERT and RoBERTa perform sentiment analysis on tweets about the COVID-19
pandemic. Our study also emphasizes how crucial it is to clean up and pre-process the dataset to get rid of extraneous data and noise
that can harm the models' performance. The effectiveness of these models on datasets from other domains and topics can be investigated
in further research. Future studies should also look into the models' interpretability and comprehend the features and patterns crucial
to sentiment analysis. This paper emphasizes how we can avoid disaster tweets and be cautious to identify hate speech that disturbs the
harmony in society.
