<![CDATA[The increasing number of scientific articles makes the process of searching, filtering, and organizing relevant literature more challenging. This problem becomes more complex because one scientific article can be associated with more than one research topic. This study aims to implement and evaluate a SciBERT-TextCNN model for multilabel classification of scientific articles in the Computer Science field based on title and abstract metadata. The dataset used was arXiv article metadata obtained from Kaggle and limited to Computer Science categories. The research stages included data preprocessing, Multi-Label Random Under-Sampling, tokenization using SciBERT tokenizer, model training, threshold tuning, and evaluation. The model was tested using five scenarios: SciBERT frozen and fine-tuning the last 1, 2, 3, and 4 layers. The best performance was obtained by fine-tuning the last 2 SciBERT layers, achieving precision of 0.708, recall of 0.726, F1-score of 0.715, macro F1-score of 0.649, and Hamming Loss of 0.02333. These results show that SciBERT-TextCNN can be used to support automatic multilabel classification of scientific articles.]]>
Copyrights © 2026
