A Hybrid Classification Model Based on BERT for Multi-Class Sentiment Analysis on Twitter

Social media is one of the media to convey opinions and sentiments. Sentiment analysis is an important tool for researchers and business people to understand user emotions efficiently and accurately. Choosing the right classification model has a significant impact on sentiment classification performance. However, the diversity of model architectures and training techniques poses its own challenges. In addition, relying on a single classification model often causes noise, bias, data imbalance, and limitations in handling data variations effectively. This study proposes a hybrid classification model where BERT is the baseline. Furthermore, BERT will be hybridized using LSTM, and BERT is hybridized with CNN to improve sentiment analysis on Twitter social media data. The hybrid approach aims to reduce the limitations of a single model classifier by increasing model effectiveness, reducing bias, and optimizing the model on imbalanced data. The following are the steps in this study, data preprocessing, data balancing, tokenization, model training, and performance evaluation. Three models were trained: the baseline BERT model, the BERT-CNN hybrid, and the BERT-LSTM hybrid. Model performance was assessed using accuracy, precision, recall, and F1 score. Experimental results show that the baseline BERT model achieves an accuracy of 91.45%, while BERT-LSTM achieves 91.60%, and BERT-CNN achieves the highest accuracy of 91.80%. However, further analysis is needed to determine whether these improvements are statistically significant and whether the hybrid model offers additional benefits beyond accuracy, such as remembering underrepresented sentiment categories.