Guest-origin and temporal heterogeneity in online hotel reviews: A structural topic modeling analysis

This study examines whether the relative salience of hotel-review themes varies across guest-origin groups and review years. Structural Topic Modeling was applied to 29,304 Booking.com reviews from 144 hotels in Vietnam collected between 2022 and 2025. A seven-topic solution captured staff/location/breakfast satisfaction, room cleanliness and comfort, amenities and food, walkability, personalized hospitality, leisure settings, and booking/travel logistics. Probability-scale contrasts were standardized over the complementary metadata distribution, with hotel-clustered CR2 inference accounting for within-hotel dependence. Guest-origin heterogeneity was markedly more stable than temporal heterogeneity. Under hotel-clustered inference, 15 of 21 guest-origin contrasts met the within-family false-discovery-rate threshold, compared with 1 of 21 2022-referenced temporal contrasts; 11 guest-origin contrasts showed strong multi-specification evidence. The largest difference was a 27.6-percentage-point higher expected prevalence of the Room cleanliness and comfort topic in Domestic than Western reviews; Domestic reviews also showed an 18.4-percentage-point lower expected prevalence of Staff, location, and breakfast satisfaction . Asian reviews exhibited a distinct profile spanning amenities, walkability, personalized hospitality, travel logistics, and leisure-related content. Temporal evidence was less stable: the 2025–2022 increase in Personalized hospitality and attentive service attenuated after hotel fixed effects, although a separate 2023-referenced analysis identified six supported temporal contrasts. The findings show that guest-origin context is associated with persistent differences in what consumers discuss, whereas annual differences are more dependent on hotel composition and temporal reference. The study advances hotel-review analytics by integrating metadata-conditioned topic prevalence with dependence-aware inference and provides context-specific diagnostic signals beyond aggregate ratings.