Skill2Vec: A Novel Hierarchical Embedding Algorithm for Technical Skill Representation in Automated Resume Parsing Systems

Abstract

Job portals these days, they receive so many applications that manual screening has become next to impossible. And here lies a tricky problem. When HR teams try to match candidates with jobs, they struggle because everyone writes their skills differently on resumes. We looked at existing tools like Word2Vec and FastText. They work okay for regular text processing, but technical skills? Not so much. See, the thing is, skills have relationships that these tools simply cannot understand. Django and Python, for instance - one depends on the other. This is not just about two words showing up together in documents.

So, we built Skill2Vec. What does it do exactly? Four things mainly. First part is HCA - it keeps skills grouped with their parent categories. Second is SCW - skills appearing together get more weightage. Third, DCW - context window changes based on how skills relate. Fourth, GR - maintains structure during the whole training process. We ran tests on 2,407 actual resumes. Total words were 646,138 and skill mentions came to 50,059 across 226 skills. Our skill database had 1,033 entries in 21 categories. Results? Skill2Vec beat Word2Vec by good margins - 23.7% on similarity tasks, 31.2% on clustering, 18.9% on analogies. We also made a Flask app with REST APIs for real deployment.