A Vietnamese Job–Candidate Matching System Based on Embedding Models and Gradient Boosting

Abstract

The rapidly increasing demand for information technology (IT) professionals in Vietnam has significantly driven online recruitment activities. Therefore, employers and job seekers continuously create a large volume of recruitment posts and candidate profiles. However, most existing recruitment platforms in Vietnam are still based on traditional, rule-based filtering methods or keyword matching methods. Therefore, in this study, we propose integrating modern natural language processing (NLP) techniques by extracting semantic features using ViBERT (Refined Bidirectional Encoded Representations from Transducers) and SBERT (Sentence-BERT) methods, while supplementing lexical features with TF-IDF (Term Frequency-Inverse Document Frequency). These features are combined with statistical attributes such as skill overlap and position matching to train and evaluate several gradient boosting models,. Experimental results show that CatBoost is the best efficiency, obtaining an F1 score of 0.7823, an AUC of 0.8571, and an nDCG@10 of 0.8739.