Objectivity-Subjectivity Detection to Boost Sentiment Analysis Accuracy: An Approach using BERT

This study analyzed the impact of preceding objectivity-subjectivity classification on sentiment analysis using BERT embeddings.The English Wikipedia and Shopee Review datasets were utilized to train classification and sentiment analysis models.Multiple machine learning models are implemented, including LinearSVC, Random Forest, AdaBoost, XGBoost, LightGBM, LSTM, and a basic dense layer.A BERT model is used for word embedding's purpose.The research consists of three models: Model 1, Model 2, and Model 3. Model 1 focuses on distinguishing between objective and subjective statements, achieving an accuracy of 99%.Model 2 performs sentiment analysis on the Shopee Review dataset, but its accuracy is observed to be less than 50%.Both models are evaluated using both raw and preprocessed data, with raw data demonstrating better accuracy.In Model 3, the bestperforming models from Model 1 and Model 2 are combined in a sequential process.Results showed incorporating objectivity-subjectivity classification improved sentiment analysis accuracy by 7% for the Shopee Review dataset.The findings demonstrate the value of text classification for enhancing performance of sentiment analysis using contextual embeddings.