A Comparative Study of NLP Transformer-Based and CNN-GloVe Models with Flask Application for Research Article Classification

This paper compares five deep learning approaches for classifying research articles: SciBERT, RoBERTa, ELECTRA, DistilbERT, and CNN-GloVe. A dataset of 2,136 research papers was used, which were categorized into six categories: Artificial Intelligence, Decision Support, Cybersecurity, IoT, Networked Systems, and Mobile Applications. The dataset was partitioned into a training with 1,495 samples, a validation set of 320 samples, and a test set of 321 samples. The results of the experiments reveal that SciBERT achieves the highest accuracy of 97.2 % and F1 score of 97.18%. However, CNN-GloVe provides 11.64 seconds for training time while achieving 95.02 % accuracy, representing an exceptional performance-to-computational-cost ratio. ELECTRA has 96.26%, DistilBERT has 95.64%, and RoBERTa achieves 93.77%, offering varying balances of accuracy and efficiency. Category-specific analysis shows that certain models excel in particular domains, with consistently high performance across all models for Cybersecurity classification, while IoT classification proves more challenging. Further, a Flask-based web interface implementing the SciBERT model for real-time research paper classification has been developed. This system enables the automated categorization of scientific documents, with potential applications in conference management, reviewer assignments, and research organization.