سامانه همایش های دانشگاه آزاد اسلامی واحد نجف آباد

صفحه اصلی / سومین کنفرانس بین المللی مهندسی و علوم کامپیوتر و اولین کنفرانس ملی سامانه های هوشمند، کلان داده و رایانش پیشرفته

An Explainable SBERT-Based Framework for Resume–Job Matching in Applicant Tracking Systems

نویسندگان :

Parisa Rezaei Aliabadi¹ Hamid Rastegari² Bita Yazdani³ Masoud Barati⁴

1- Nahafabad Branch, Islamic Azad University 2- Nahafabad Branch, Islamic Azad University 3- Nahafabad Branch, Islamic Azad University 4- Nahafabad Branch, Islamic Azad University

کلمات کلیدی :

job matching،sentence-bert(sbert)،explainable artificial intelligence،transformers،cosine similarity،resume screening،applicant tracking system(ats)

چکیده :

Resume–job matching has become a fundamental functionality of contemporary Applicant Tracking Systems (ATS). Despite its importance, a large portion of existing solutions still depend on keyword driven rules or highly complex deep learning architectures whose decision making processes remain opaque. In parallel, progress in transformer-based language models, including large-scale pre-trained architectures, has substantially enhanced the ability to capture semantic relationships in recruitment-related textual data. However, most prior studies focus primarily on similarity computation or ranking accuracy, without providing interpretable and auditable hiring decisions. This paper proposes an end-to-end and explainable framework for resume–job matching that combines Sentence-BERT (SBERT) embeddings with a lightweight logistic regression classifier. The proposed pipeline employs minimal semantics-preserving preprocessing, contextual sentence embeddings, similarity-based feature construction, and SHAP-based explanation of model predictions. The framework is evaluated using both classification and ranking metrics and benchmarked against encoder-based transformers and decoder-based LLM approaches reported in recent recruitment literature. Experimental results demonstrate that the proposed method achieves competitive performance while offering improved transparency, reproducibility, and suitability for real-world ATS deployment. Additionally, the proposed method has increased the Precision@10 accuracy by 4.55% and the F1-score by 2.70%.