Abstract
In the context of rapidly evolving digital tourism and the exponential growth of user generated textual data, there is a growing demand for personalized recommender systems capable of effectively interpreting the semantic nuances of user queries. This study presents a comparative evaluation based on models RuBERT, TinyBERT, and XLM-RoBERTa for the task of recommending tourist destinations from textual user preferences. We propose a system architecture that includes data preprocessing, generation of query embeddings, retrieval of relevant tourist locations using the Faiss library, and subsequent ranking of results. For experimentation we compiled a corpus of 28,104 user reviews and queries collected from tourism platforms. The results indicate that XLM-RoBERTa delivers the strongest performance in multilingual settings, particularly for cross-lingual queries. RuBERT attains higher accuracy on Russian language texts and TinyBERT provides a favorable balance between effectiveness and computational cost, which makes it suitable for resource constrained applications. The findings can inform the development of intelligent tourism services, conversational agents, and generation personalized recommendation systems.