Prototype-Guided SDG-Aware Topic Discovery from Multilingual Tourism Reviews

Abstract

Tourism reviews provide continuous visitorgenerated evidence for destination monitoring, but sustainabilityrelated information is often implicitly embedded in ordinary travel narratives. This study proposes a prototype-guided, label-free NLP framework for discovering candidate Sustainable Development Goal (SDG)-related themes from multilingual tourism reviews. Using Google Maps reviews from tourism points of interest in Mandalika and the broader Lombok destination context, we compare five language-representation and embedding scenarios based on original multilingual text, English translation, Indonesian translation, and languagespecific embedding models. The framework combines sentence embeddings, BERTopic-based topic induction, prototype-based SDG alignment, adaptive thresholding, and dual-anchor semantic filtering. The results reveal a trade-off between topic coherence and clustering density: the original multilingual BGE-M3 scenario achieved the highest topic coherence $(c_{v}=0.5147)$ but also produced a high outlier ratio, whereas the Indonesian E5-based scenario produced a more compact clustering structure with a lower outlier proportion and relatively strong coherence. In the original multilingual scenario, approximately one-fifth of retained documents were categorized as SDG-related, mainly reflecting themes associated with culture and communities, waste management, infrastructure access, informal vendors, and governance-related visitor issues. The findings show that multilingual tourism reviews can complement formal sustainability monitoring by revealing recurring visitor-reported issues.