Dynamic topic modeling for aspect categorization in aspect-based sentiment analysis using BERTopic on hotel reviews

This study introduces a hybrid and explainable framework combining dynamic topic modeling (BERTopic), aspect-based sentiment analysis (ABSA), and LSTM-based sentiment classification with SHAP interpretability. A dataset of 31,876 Indonesian hotel reviews (2023-mid-2025) from Traveloka, TripAdvisor, and Google Reviews was used. BERTopic identifies evolving topics, semantically mapped to five service aspects: cleanliness, service, food, comfort, and location. Aspect-level sentiments are classified using LSTM, while SHAP provides word-level attribution. The framework outperforms traditional LDA in coherence, reveals temporal sentiment trends, and offers actionable insights for hotel managers. This approach contributes to scalable and interpretable sentiment analytics in the hospitality domain.