ENSEMBLE OF MODELS FOR AUTOMATIC CLASSIFICATION OF RESUME TEXTS

The aim of this study is to enhance the accuracy of automatic classification of resume texts under limited computational resources. The objective is to design and experimentally validate an ensemble of models, fine-tuned on the open-source ResumeAtlas dataset, augmented with taxonomic regularization and probability calibration, to boost classification accuracy and robustness. The methods of investigations include the proposed pipeline containing (I) a TfidfSentenceSelector module for extracting seven most informative sentences; (II) domain-adaptive pretraining (DAPT) of DeBERTa-v3-large; (III) R-Drop regularization; (IV) Longformer-large-4096 for full document processing; (V) logit averaging and temperature scaling. The author performs evaluation on the Résumé Atlas open corpus using Top-k Accuracy and F1 metrics. The novelty of the work lies in showing that TF-IDF resume compression retains almost all essential facts while reducing input size fivefold, and that the ensemble of “resume summary + long context” yields additional gains in accuracy. Also proven is the efficacy of single-epoch DAPT and R-Drop under restricted GPU time. The study results in achieving Top-1 = 0.93, Top-3 = 0.97, Top-5 = 0.98, Top-10 = 0.99, and F1-macro = 0.93, surpassing the baseline Résumé Atlas model by 2 percentage points and classical TF-IDF+SVM by 11 percentage points in Top-1 accuracy. The entire training cycle completes in approximately 4 hours on a single NVID-IA A100-80 GB GPU. The findings state that the combination of domain-adaptive pretraining, text compression, and heterogeneous attention provides a novel open-source solution for resume classification, potentially transferrable to other professional taxonomies.