Unsupervised Skill Identification from Job Ads

Abstract

Identifying skills in the huge number of job ads posted every day can give multiple entities the big picture of job market requirements. Such requirements will not only help universities adapt their curricula to produce more employable graduates but will also give meaningful insights for job seekers. Unfortunately, analyzing these ads is challenging because they are submitted in raw text and need to be structured in order to identify the required skills. Most prior work, however, focuses on detecting and exact matching skills in job ads through keyword matching from built-up skill bases. In this paper, we develop a new methodology for identifying skills in job ads. This methodology is two-fold. First, the methodology identifies skills from job ads expressed in sentences where we use sentence embeddings to detect skills in job ads. Secondly, technical words for software names and certifications are identified using Wikipedia. Experiments on a sample of real-world job ads show that the precision and recall yield promising results in mapping and extracting hard skills from job ads. Our findings can offer insights to universities and graduates about the top skills and competencies requested in the job market.