The ability to extract skill information from unstructured text documents, such as resumes and job advertisements, can deliver efficiencies for recruitment and training and insights for policy and decision makers. The task requires the ability to differentiate between the thousands of broad and narrow skills required across the contemporary labour market. We describe the development of a Deep BERT-based Semantic Skill Matching (DBSSM) classifier that responds to this challenge by incorporating a combination of complementary classification algorithms, domain-expert fine tuning of weights and thresholds and labour market information contained in the European Skills, Competencies, Qualifications and Occupations (ESCO) taxonomy. The performance of the DBSSM classifier was tested on a dataset of anonymized resumes and found to achieve a Mean Average Precision at 1 (MAP@1) of 0.413 and an Average F1 (AF1) score of 0.206, exceeding the performance of alternative approaches. As well as supporting efficient candidate shortlisting (for recruitment) the classifier has applications for job seekers, education and training providers, employers, policy makers and researchers.