Knowledge-Driven Unsupervised Skills Extraction for Graph-Based Talent Matching

In human resource management of large organisations, finding the best candidate for a job description requires an extensive examination of a large number of resume profiles. Even with the advent of Deep Information Retrieval and the supported semantic similarity search, identification of relevant skills within profiles requires thorough investigation over several aspects, including educational background, professional experience, achievements, etc. However, these techniques are based on the existence of domain-specific, human-annotated datasets, a laborious task that portrays high cost and a slow labeling progress. In this paper, we propose Resume2Skill-SE, an end-to-end architecture for interpretable skill-based talent matching. The solution consists of two components. The first module uses an unsupervised approach for skills extraction based on state-of-the-art text embeddings and efficient semantic similarity search. The second module creates a profile-skills bipartite graph and uses a proposed ranking formula for similar resume profiles, minimising the effect of potential errors from the skills extraction module. The optimal ranking formula was identified through an intuitive and automated evaluation method for getting relevance scores. The proposed technique delivers promising results while also including an interpretability layer by showing the common skills of a pair of resume profiles.