Identifying Tourist Preferences from Online Reviews: An Integrated BERTopic-LLM Approach with Bidirectional Validation

Understanding tourist preferences is crucial for destination competitiveness, yet existing approaches struggle to achieve semantic coherence and identify low-frequency preferences but rich information (long-tail preferences) for tourism service development. Furthermore, few studies have employed bidirectional systematic validation in the analysis of large-scale user-generated content (UGC). This study addresses this gap by proposing an integrated BERTopic–LLM framework to identify and structure tourist preferences. Consequently, a bidirectional validation procedure was used to increase the accuracy of preference classification, enhance robustness, and reduce manual adjustments. Using 265,518 TripAdvisor reviews of Tokyo across restaurants, hotels, and attractions, BERTopic extracted 1,071 distinct topics, and an LLM then structured them into 212 distinct preferences. Frequency scarcity and topic specificity were used to identify 89 long-tail preferences (42% of all identified preferences) from 6.1% of the total reviews. The novel framework provides researchers and practitioners with a robust approach for converting unstructured UGC into structured preferences. The findings carry practical implications for destination management, including the redesign and development of customized tourism products and services.