Hybrid Semantic-Spatial Topic Modeling of Tourism Narratives: A Spatio-Temporal Analysis of TripAdvisor Data Using LDA, Embedding Alignment, and Swarm Intelligence

Abstract

This study introduces two novel schemas for spatio-temporal topic modeling of tourism narratives by extracting and aligning semantically rich, geographically grounded topics from TripAdvisor’s Morocco travel forum over a 15-year period (2008–2023). Schema I integrates Latent Dirichlet Allocation (LDA) with geotemporal conditioning, while Schema II employs a hybrid architecture that combines sentence embeddings, autoencoders, and a dual optimization strategy using Genetic Algorithms (GA) and Artificial Bee Colony (ABC) clustering. Spatial and temporal metadata were embedded to capture evolving and localized tourist discourse. Evaluation was conducted using coherence scores, Kullback–Leibler (KL) divergence, topic overlap matrices, and spatial clustering metrics. Schema II outperformed Schema I in coherence (0.62 vs. 0.49), topic separation, and spatial resolution. UMAP and t-SNE visualizations revealed distinct, well-formed semantic clusters that aligned with key tourist destinations such as Marrakech, Fes, and Casablanca. Heatmaps and temporal density plots showed seasonal and event-driven discourse spikes. The findings demonstrate the utility of combining neural embeddings with biologically inspired optimization algorithms for extracting interpretable, location-sensitive topics in tourism analytics. This hybrid approach offers a robust framework for destination planning, tourist experience analysis, and cultural trend monitoring.