This paper presents a method for semantic mediation based on Natural Language Processing (NLP), designed to ensure interoperability between heterogeneous software systems. The method integrates modern transformer-based language models to extract context-aware semantic representations from natural language descriptions and metadata. A multi-level architecture is proposed, involving data preprocessing, semantic embedding generation, and algorithms for semantic similarity calculation and data transformation. Experimental validation within the medical domain demonstrates high accuracy and significant time savings compared to manual integration methods. The proposed approach offers scalability, domain adaptability, and a reduction in the need for expert-driven interface development. Future directions include active learning, domain adaptation, and the development of interactive mediation management tools.