The automatic normalization of clinical entities is a critical step for enabling structured analysis of free-text medical records. This study proposes and evaluates four distinct retrieval strategies for normalizing entities extracted via Named Entity Recognition (NER) from Spanish psychiatric emergency notes, using Unified Medical Language System (UMLS) as the target terminology. A total of 768 annotated entities were mapped using MiniLM, Multilingual BERT, lexical matching, and the UMLS API. Results show that the API-based approach yields the best performance (Hit@3 = 56.8% and Hit@5 = 57.9%), effectively balancing accuracy and computational efficiency. Although embedding-based methods such as MiniLM and Multilingual BERT often outperform traditional techniques in other domains, they showed only marginal improvements over simple lexical matching in this context, while incurring significantly higher computational costs. These findings suggest that in the psychiatric domain, where language is often ambiguous, embedding-based approaches may fall short. Hybrid systems that combine fast retrieval with semantic reasoning are needed to capture the richness of psychiatric narratives.