Abstract:
Sentiment analysis represents a fundamental component of natural language processing, playing
a pivotal role in understanding user sentiments and public opinion across diverse domains. This
research undertakes a comparative analysis of various machine learning methodologies for
sentiment analysis, focusing specifically on polarity. The models evaluated include SVM with
both Linear and Non-Linear (RBF) kernels, as well as BERT and ROBERTA. The primary
metric employed to gauge the performance of each model is accuracy. The findings reveal that
SVM with a Linear kernel achieved an accuracy of 84%, while the Non-Linear SVM (RBF)
recorded an accuracy of 78%. BERT, when not fine-tuned, demonstrated an accuracy of 89%,
which improved to 90% following fine-tuning. ROBERTA also achieved an accuracy of 90%.
The insights derived from this study offer valuable guidance regarding the most effective
machine learning models for sentiment analysis tasks. Researchers and practitioners can leverage
these findings to select appropriate sentiment analysis models tailored to their specific
application requirements. The research highlights the relative ease of enhancing accuracy
through fine-tuning and the application of various pre-processing techniques. Furthermore, it
underscores the importance of considering the wide range and scope of machine learning
approaches to enhance and increase the accuracy and reliability of opinion mining and sentiment
analysis systems in real-world applications.
