Hardware-effective Approaches for Skill Extraction in Job Offers and Resumes
Talk, NLP Seminars, University of Sheffield, Online
This talk will present hardware-effective methods for skill extraction. In this task, resource optimization is often not a priority, and methods evaluation is limited to job offers due to the challenges of data privacy. We will understand the trade-off between task performance and computing resources, from rule-based systems to BERT-based large language models in both scenarios, resumes, and job offers. Our results show that competitive skill matching can be achieved using minimal hardware, where rule-based and semantic systems run solely in CPU and neural models use minimal GPUs, with less than 25 minutes of training time (1 node, 24GB GPU RAM).