Sentiment analysis in hotel guest reviews has become essential for evaluating customer satisfaction and service quality. This study improves sentiment classification accuracy by utilizing the Multilingual BERT model with an improved performance evaluation framework. Using the Knowledge Discovery in Databases (KDD) methodology, this research involves data selection, preprocessing, transformation, sentiment classification, and performance evaluation. A dataset of 715 hotel reviews from Qubika Boutique Hotel, sourced from Agoda, was used to assess the model's effectiveness. The classification results showed high accuracy in identifying positive sentiment, with 98% precision, 97% memory, and 98% F1 score, as observed in 432 correctly classified reviews. However, challenges were identified in the classification of neutral sentiment, which achieved a precision of 87% with 127 correctly classified cases, and negative sentiment, where the accuracy was 92%, with 104 correctly identified reviews. The overlap in confidence scores, especially in the range of 0.4-0.6 between neutral and negative sentiment, highlights the need for improved contextual embedding and hybrid modeling techniques. The sentiment distribution analysis revealed that 60-70% of reviews were positive, 20-30% neutral, and 10-15% indicated dissatisfaction, underscoring the need for targeted service improvement. These findings provide valuable insights for data-driven decision-making in hospitality management, enabling businesses to strengthen service power and address critical areas of concern. Future research should focus on refining model interpretability, expanding multilingual datasets, and integrating real-time sentiment analysis to improve classification performance. Strengthening these aspects will contribute to a more robust and scalable sentiment analysis framework, ensuring greater precision in capturing the guest experience and optimizing service strategies in the hospitality industry.