Zero-Shot Skills Extraction Using Large Language Models

This technical report presents a novel approach for skills extraction from job descriptions using large language models (LLMs). The objective is to enhance the identification of skills required in the labor market, leveraging the ESCO framework, which catalogs over 13,000 skills. The authors propose an end-to-end zero-shot system that generates synthetic training data for skills extraction, employing a classifier to identify skill mentions in job postings. The report details a two-step skills matching process, where the first step involves generating potential skill matches, followed by a re-ranking phase utilizing LLMs. The findings indicate that the proposed method significantly improves performance metrics, achieving a notable increase in the RP@10 score compared to previous techniques. Additionally, the report discusses the challenges of skills extraction, such as the complexity of the ESCO taxonomy and the need for extensive training data. The integration of LLMs is shown to enhance the efficiency of skills matching pipelines without requiring human annotations.