Mapping ESCO Skills Taxonomy to Educational Offers Using Natural Language Processing and Large Language Models

The European Skills, Competences, Qualifications, and Occupations (ESCO) taxonomy serves as a standardized multilingual framework designed to categorize and describe skills, competences, qualifications, and occupations. Its primary purpose is to support professional mobility and foster a more integrated labor market across Europe by providing a common language for stakeholders in employment, education, and training. Despite its comprehensiveness, ESCO faces challenges in accurately mapping unstructured text, such as course descriptions or job postings, to its skill taxonomy due to the limitations of existing APIs and the complexity of natural language processing (NLP).To address these challenges, this research proposes a computational system capable of inferring ESCO skills from textual data by leveraging advanced NLP techniques, sentence embeddings, HDBSCAN-based clustering, and Large Language Models (LLMs). The system evaluates similarity between input queries and ESCO skills using three approaches: (1) NLP only, (2) LLM-only recommendations via the official Deepseek API, and (3) a hybrid approach combining NLP and LLM outputs.HDBSCAN-based clustering is applied to group similar skills into coherent clusters, improving interpretability while addressing the limitations of spherical clustering methods like K-Means. Deepseek-V3, a state-of-the-art LLM developed by DeepSeek AI, is integrated via the official API to refine and validate skill mappings. Preliminary results demonstrate the system’s potential to automate skill inference with high accuracy, benefiting universities, students, and employers. User testing indicates that the hybrid approach consistently outperforms standalone methods, with participants rating its results as much more relevant. Notably, while the ESCO API performs comparably to the LLM-only approach for English inputs, it underperforms significantly for Portuguese inputs, highlighting the need for improved multilingual support. Challenges remain, particularly with ambiguous input texts and mappings involving non-English languages, such as Portuguese. Future work will focus on refining the system for multilingual support and addressing other edge cases.