COMPARATIVE STUDY OF TRANSFORMER-BASED MODELS FOR AUTOMATED RESUME CLASSIFICATION

Abstract

This study presents a comparative evaluation of transformer-based models and traditional machine learning approaches for automated resume classification—a key task in optimizing recruitment workflows. While traditional approaches like Support Vector Machines (SVM) with TF-IDF demonstrated the highest performance (93.26% accuracy and 95% F1-score), transformer models such as DistilBERT and RoBERTa showed competitive results with 93.27% and 91.34% accuracy, respectively, and fine-tuned BERT achieved 84.35% accuracy and an F1-score of 81.54%, indicating strong semantic understanding. In contrast, Word2Vec + LSTM performed poorly across all metrics, highlighting limitations in sequential modelling for resume data. The models were evaluated on a curated resume dataset available in both text and PDF formats using accuracy, precision, recall, and F1-score, with preprocessing steps including tokenization, stop-word removal, and lemmatization. To address class imbalance, we applied stratified sampling, macro-averaged evaluation metrics, early stopping, and simple data augmentation for underrepresented categories. Model training was conducted in a PyTorch environment using Hugging Face’s Transformers library. These findings highlight the continued relevance of traditional models in specific NLP tasks and underscore the importance of model selection based on task complexity and data characteristics