Using language models to extract and map emerging skills to the ESCO taxonomy

As the online labour market evolves, the ability to identify not only established skill requirements but also emerging skills directly from online job ads (OJAs) has become crucial for anticipating workforce needs and designing effective training programmes. To address this challenge, we introduce the SkiLLink pipeline, which is articulated into two main components: (i) an identification-and-filtering component that extracts new skills as they appear in OJAs using word embeddings and language models; and (ii) a mapping component that employs embedding-based recommendations and large language models to align the extracted skills with an established labour market taxonomy. This distinction is also reflected in our evaluation strategy. The extraction-and-filtering component is assessed qualitatively through validation by labour market experts across multiple EU countries, who reviewed the relevance and correctness of the emerging skills. The mapping component is evaluated quantitatively using two benchmark datasets and a baseline derived from the European Skills taxonomy (ESCO). We further perform a robustness analysis employing cosine similarity and twelve additional distance metrics to examine the stability of embedding-based alignment.