Sentiment Classification Using BERT-CNN and SMOTE: A Case Study on Hotel Reviews Dataset

The increasing importance of user-generated content in the hospitality industry necessitates advanced sentiment analysis tools to derive actionable insights from customer reviews. Traditional methods often struggle with the complexities of natural language, such as contextual dependencies and nuanced emotional expressions. This research leverages the BERT-CNN hybrid model, which combines BERT’s contextual language understanding with CNN’s feature extraction capabilities, to address these challenges and improve sentiment classification accuracy. Using a dataset of 1,828 hotel reviews from Eastparc Hotel Yogyakarta, the model achieved an impressive accuracy of 99.59%, with precision and recall exceeding 0.99. The application of SMOTE effectively resolved class imbalance, enhancing the model’s ability to generalize across diverse sentiment classes. Training and validation loss curves exhibited steady convergence, indicating robust learning and minimal overfitting. These results provided actionable insights into customer satisfaction, offering targeted recommendations for enhancing service quality and operational strategies. This study demonstrates the practicality of integrating advanced machine learning architectures in sentiment analysis, enabling the hospitality sector to transform unstructured feedback into meaningful insights. The findings contribute to academic advancements in natural language processing and practical innovations in customer experience management. Future research may expand this framework to other domains, further underscoring its adaptability and impact.