A Comparative Study of Topic Modeling Algorithms for Online Reviews: The Case of Casino Integrated Resorts in South Korea (LDA, STM, and BERTopic)

When applying topic modeling to online reviews in tourism and hospitality research, scholars face a largely undocumented decision: which algorithm to use.This study addresses that choice empirically by benchmarking three representative algorithms-Latent Dirichlet Allocation (LDA), the Structural Topic Model (STM; implemented as Dirichlet-Multinomial Regression), and the embeddingbased BERTopic-on a single common corpus.The corpus comprises Google Maps user reviews of seven casino integrated resorts and hotel casinos in South Korea.Rather than comparing results reported under heterogeneous conditions, we hold the corpus fixed and evaluate all three algorithms on four unified criteria: topic coherence (c_v), topic diversity, stability across random seeds (Jaccard similarity), and covariate responsiveness (operator and regional effects).STM achieved the highest coherence (0.517) and diversity (0.920); BERTopic was the most stable (0.763, 95% CI 0.641-0.868)and produced the most fine-grained niche topics, and LDA outperformed BERTopic on coherence (0.499) yet showed the lowest diversity (0.820).Crucially, these rankings are conditional: coherence rankings reversed with the number of topics, and covariate responsiveness was structurally favourable to the model embedding covariates internally, whereas stability rankings held across topic-number and hyperparameter conditions.By translating these differences into an algorithm-selection guideline provisional to the present corpus and analytic conditions, this study provides an empirical reference for a methodological choice that prior tourism research has typically left unjustified.