The digital transformation of the hospitality sector has been driven by recent advances in the capabilities of conversational agents to effectively deliver services to users. Conversational performance of chatbots is nevertheless challenged by the high cardinality of request intents. This study aims at exploring the rich structure behind request intents to support their detection, therefore guiding conversations. To this end, we propose a novel approach that learns a hierarchical structure of intents based on their semantic similarities and predictive dependencies (shared errors), and subsequently incorporates these structures into the- learning of Large Language Models for intent detection. Gathered results from Hijiffy and Clinc150 case studies highlight statistically significant improvements across performance metrics against state-of-the-art approaches, highlighting the relevance of modeling structural relationships between intents. The particular ability of the proposed models to handle the high imbalance and cardinality of intents opens new opportunities for user adherence towards agent-mediated services in the hospitality domain.