Supervised Vector Weighting and BiLSTM-Attention: A Novel Approach for Arabic Medical Question Classification on Imbalanced Datasets
摘要
Online medical services, including telehealth and telemedicine platforms, play a pivotal role in enhancing healthcare delivery. Among these services, medical question-answering systems are crucial for providing patients with timely guidance and support. However, the large volume of questions spanning various medical specialties poses significant challenges for automated classification, particularly when compounded by the imbalanced distribution of medical classes and the inherent complexity of the Arabic language. Although existing approaches have addressed the problem of imbalanced datasets, prior research has focused on resampling techniques, which can impact the quality of the original dataset, especially in the medical domain. To address these challenges, this paper proposes a novel Supervised Vector Weighting (SVW) scheme, integrated with a deep learning architecture combining Bidirectional Long Short-Term Memory and a self-attention mechanism (BiLSTM-Attention). The proposed model aims to handle the imbalanced class distribution while simultaneously improving the extraction and understanding of contextual features in Arabic medical questions. Extensive experiments were conducted on an imbalanced Arabic medical question dataset, leveraging multiple resampling techniques and transformer-based approaches such as DistillBERT. A comprehensive comparative analysis demonstrates the superior performance of the proposed SVW and BiLSTM-Attention model across various evaluation metrics, including