Hybrid Pipelines for Intelligent Human Resources Text Classification: LLMs, RAG, and Generative AI

Abstract

The digital transformation of Human Resources Information Systems (HRIS) requires advanced approaches to process unstructured textual data originating from Curriculum Vitae (CVs), cover letters, and job postings. Traditional text classification methods exhibit limitations when faced with current needs for contextual understanding and fine-grained skill detection. This paper proposes a hybrid pipeline combining advanced text classification, contrastive learning (SimCSE, Contriever), Retrieval-Augmented Generation (RAG), and generative AI (LLMs) to enhance candidate pre-screening, CV–job matching, and profile generation. Experimental results obtained on a corpus of 50,000 CVs show that the hybrid pipeline with RAG achieves an accuracy of 94.2% with a macro-F1 score reaching 92.3%, outperforming standard Transformer-based approaches and improving performance by +2.5% compared to the hybrid pipeline without RAG. When integrated into an HRIS, the proposed system accelerates recruitment processes while improving accuracy and efficiency, all while maintaining inference times compatible with operational deployment.