Hardware-effective Approaches for Skill Extraction in Job Offers and Resumes

This technical report presents experiments focused on the automatic extraction of skills from job offers and resumes, emphasizing hardware-effective methods that require minimal computational resources and annotated data. The study evaluates various extraction methods, including rule-based, semantic, and neural models, assessing their performance on both public and commercial datasets. The findings indicate that while standalone rule-based and semantic models exhibit limited and variable performance, neural models demonstrate competitive stability and effectiveness, even with smaller datasets, achieving approximately 30% improvement. The report details the hardware requirements for these methods, highlighting the use of CPU-based systems with less than 8 GB of RAM for simpler models and GPUs for neural models, with a maximum memory usage of 24 GB and training times under 25 minutes. The research aims to develop incremental and hybrid approaches tailored to organizational needs, addressing the challenges posed by limited resources and annotated data in the skill extraction process.