Ontology-Based Semantic Normalization of Resumes for Classification
TL;DR
Rather than scaling performance uniformly across the entire evaluation suite, the ontology layer acts as a targeted traceability and semantic refinement filter that contributes information beyond filtered-profile selection alone and produces a metric-dependent change in classifier behaviour at the validation-selected threshold.
Abstract
During the recruitment process, it is possible for CVs to appear well-organized. However, it is not always straightforward to compare them. The same competence may be denoted by different designations, and the levels of competence are not universally employed in the same manner. Natural Language Processing (NLP) methodologies can extract these data points; however, ensuring the consistency of this data across multiple CVs remains a challenge. In a multitude of cases, the comparability of two profiles remains ambiguous. In the present study, an ontological approach is adopted to solve this issue. The concept under discussion is that of the extraction of entities from CVs and their subsequent representation in a more structured form, utilizing RDF and an ontology aligned with ESCO—the multilingual classification of European Skills, Competences, and Occupations. Subsequently, the rules of SHACL are applied to verify the semantic coherence of the data; the validated data are transmitted to a model for classification. At this stage, the dataset becomes smaller, but semantically cleaner, more traceable, and enriched with validation indicators that can be used by the classification model. The proposed system is implemented as a set of microservices. A Spring Boot component coordinates the flow, whilst the Python services, implemented using Python 3.10.12 are responsible for the primary processing stages including extraction, validation and classification. A same-corpus ablation was conducted to separate ontology-guided profile selection from the contribution of the validation-derived quality features. On the same 35,770 filtered CV–job pairs, adding these features increased external benchmark accuracy from 0.794 to 0.809, recall from 0.760 to 0.865, F1-score from 0.749 to 0.786, and ROC-AUC from 0.881 to 0.887. A p-value of 0.00540 paired with a 1.54 effect ratio from McNemar’s test showed a statistically significant paired difference between the two configurations. However, precision decreased from 0.739 to 0.720 while Average Precision compressed from 0.851 down to 0.844. Rather than scaling performance uniformly across the entire evaluation suite, the ontology layer acts as a targeted traceability and semantic refinement filter that contributes information beyond filtered-profile selection alone and produces a metric-dependent change in classifier behaviour at the validation-selected threshold.