Description

Context & overview

The European labour market requires tools capable of analysing skills at scale and aligning them with standard taxonomies, such as ESCO (European Skills, Competences, Qualifications and Occupations). Traditional methods for skill extraction from job postings or CVs demand extensive labelled data, which is costly and difficult to maintain in multilingual contexts. To address this challenge, researchers proposed an approach based on Large Language Models (LLMs) that can perform zero-shot skill extraction and matching. This means that the model does not need task-specific training data but can directly map skills mentioned in free text to ESCO entries, enabling more agile and scalable labour market intelligence.

 

Challenges addressed

Lack of annotated datasets for training supervised models to map job postings to ESCO.	Language variation in job descriptions, which complicates accurate skill recognition.
Need for scalability across multiple European countries and languages.	Alignment with ESCO as a standard European taxonomy to ensure comparability of results across contexts.

 

Solution implemented

The researchers developed a zero-shot methodology where LLMs are used to identify and match skills directly to ESCO without prior fine-tuning. The solution includes three components:

Synthetic data generation	The LLM generates AI training examples covering the full ESCO taxonomy, helping to recognise a wide range of skills.
Retriever	Given a job posting, the system retrieves potentially relevant ESCO skills based on semantic similarity.
Reranker	A second LLM is applied to refine the ranking and select the most accurate skills.

This approach leverages LLMs both to generate auxiliary data and to perform the actual matching, reducing the reliance on external labelled datasets.

 

Impact & results

The methodology shows how AI can reduce dependency on manual annotation, enabling real-time skill intelligence across countries and sectors. Because of the strengthen of the link between raw labour market signals and ESCO, the system supports evidence-based curriculum design and responsive VET programmes. In addition, the accuracy is improved and this demonstrates the value of multi-step AI tool.

Reference Link

https://www.alphaxiv.org/overview/2307.03539v1