SkillSense: A Context-Aware Machine Learning Framework for Automated Skill Scoring in Resumes

Keyword-based applicant tracking systems treat every mention of a skill as equal evidence of competence and therefore overlook qualified candidates whose resumes do not reproduce the exact phrasing of a job description. This paper presents SkillSense, a context-aware machine learning framework that extracts skills from resumes and scores their relevance to a target job role. The framework combines transformer-based named entity recognition for skill extraction, normalisation against the O*NET taxonomy, feature engineering that fuses TF-IDF vectors, 768-dimensional BERT embeddings and candidate metadata, and an attention-weighted scoring function that accounts for the priority of each requirement in the job description, the industry domain and the experience level. Six supervised classifiers (Naive Bayes, Logistic Regression, Support Vector Machine, Random Forest, XGBoost and a fine-tuned BERT classifier) were compared with five-fold stratified cross-validation on a curated dataset of 1,200 anonymised resumes and 350 job descriptions from five domains, of which 500 resume–job pairs were labelled by expert annotators. The BERT classifier performed best, reaching 95.1% accuracy, 0.942 precision, 0.929 recall and a 0.935 F1-score, while XGBoost was the strongest classical model at 93.4% accuracy. Context-aware weighting improved performance by 12–15% over context-agnostic baselines. SkillSense has been deployed as a three-tier web application that returns ranked candidates with a 0–100 score and a Recommended / Not-Recommended label, giving recruiters consistent, comparable and role-specific evidence for shortlisting.