Scholarlink

<p>Imbalanced datasets are prone to bias due to overrepresentation of certain classes, which can affect model performance. To address this, data balancing techniques like oversampling and undersampling are used, though they may lead to overfitting or poor data representation. This paper focuses on classifying data into 8 to 26 distinct classes using multi-domain NLP datasets containing over 54,000 annotated comments. We use transformer models, enhanced with ensemble learning, to consolidate results and select the best-performing model based on the F1-measure macro-average. Our sub-models approach, which reduces the number of classes iteratively, helps mitigate imbalance and improves results. We achieve an F-measure of 0.84 for binary and 7-class tasks, and 0.65 for the 26-class task. Our method improves multi-class classification performance and can be applied to other languages and tasks.</p>

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

Submodels approach for (im)balanced Arabic data classification

  • 期刊论文

  • 发表日期：

    2026年5月28日

查看全文
  • Ibtissam Touahri

摘要

Imbalanced datasets are prone to bias due to overrepresentation of certain classes, which can affect model performance. To address this, data balancing techniques like oversampling and undersampling are used, though they may lead to overfitting or poor data representation. This paper focuses on classifying data into 8 to 26 distinct classes using multi-domain NLP datasets containing over 54,000 annotated comments. We use transformer models, enhanced with ensemble learning, to consolidate results and select the best-performing model based on the F1-measure macro-average. Our sub-models approach, which reduces the number of classes iteratively, helps mitigate imbalance and improves results. We achieve an F-measure of 0.84 for binary and 7-class tasks, and 0.65 for the 26-class task. Our method improves multi-class classification performance and can be applied to other languages and tasks.