Scholarlink

In online medical services, the underrepresentation of most specialties in patient questions complicates the model’s learning task of categorizing patients’ needs. Resampling techniques were one of the solutions proposed to solve the problem of unbalanced class distribution, which is achieved through two main techniques, undersampling and oversampling. To this purpose, this paper offers a comparative study of various resampling techniques, which were applied to achieve a balanced class distribution in a dataset of Arabic medical questions classification. The resampling techniques were employed in conjunction with two DL architectures (LSTM and CNN) in the classification task. The building experiments were evaluated using a dataset of 59,244 Arabic medical questions. The conducted experiments were evaluated using evaluation metrics, such as \(P_{m}\) , \(R_{m}\) , \(F1_{m}\) score, and \(MCC\) . The outcome illustrates the superiority of TomekLinks as an undersampling technique coupled with LSTM with an \(MCC\) of 68.04% and ADASYN as an oversampling technique in conjunction with LSTM with an \(MCC\) of 69.30%.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

The Effect of Resampling Techniques on Arabic Medical Question Categorization

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mohammed Bahbib,
  • Majid Ben Yakhlef

摘要

In online medical services, the underrepresentation of most specialties in patient questions complicates the model’s learning task of categorizing patients’ needs. Resampling techniques were one of the solutions proposed to solve the problem of unbalanced class distribution, which is achieved through two main techniques, undersampling and oversampling. To this purpose, this paper offers a comparative study of various resampling techniques, which were applied to achieve a balanced class distribution in a dataset of Arabic medical questions classification. The resampling techniques were employed in conjunction with two DL architectures (LSTM and CNN) in the classification task. The building experiments were evaluated using a dataset of 59,244 Arabic medical questions. The conducted experiments were evaluated using evaluation metrics, such as \(P_{m}\) , \(R_{m}\) , \(F1_{m}\) score, and \(MCC\) . The outcome illustrates the superiority of TomekLinks as an undersampling technique coupled with LSTM with an \(MCC\) of 68.04% and ADASYN as an oversampling technique in conjunction with LSTM with an \(MCC\) of 69.30%.