Arabic text classification: the need for multi-labeling systems.

Area of Science:

• Natural Language Processing
• Machine Learning
• Information Retrieval

Background:

• Text categorization, or assigning labels to documents, is crucial for organizing information.
• Arabic news articles present unique challenges for automated classification due to linguistic nuances.
• Existing methods often struggle with the complexity of multi-label classification tasks.

Purpose of the Study:

• To develop and evaluate automatic text categorization models for Arabic news articles.
• To compare the performance of shallow learning and deep learning approaches for single and multi-label classification.
• To identify the most effective classification models for diverse Arabic news content.

Main Methods:

• Construction of two large datasets: one with 90k single-labeled articles and another with over 290k multi-tagged articles from Arabic news portals.
• Evaluation of ten shallow learning classifiers and an ensemble model for single-label classification.
• Implementation and testing of both shallow (Logistic Regression, XGBoost with OneVsRest) and deep learning (CNN, LSTM, GRU variants) multi-label classification approaches.

Main Results:

• For single-label classification, SVM achieved the highest accuracy at 97.9%, while AdaBoost scored 87.7%.
• For multi-label classification, shallow learning models achieved accuracies up to 84.7% (XGBoost).
• Deep learning models significantly outperformed shallow learning, with CGRU achieving the highest accuracy of 94.85%.

Conclusions:

• Multi-label text categorization is more effective than single-label for complex news articles.
• Deep learning models, specifically CGRU, demonstrate superior performance for multi-label Arabic news classification.
• The developed datasets and models offer a valuable resource for advancing Arabic NLP research.