Data augmentation for Arabic text classification: a review of current methods, challenges and prospective directions.

Abstract

The effectiveness of data augmentation techniques, i.e., methods for artificially creating new data, has been demonstrated in many domains, from images to textual data. Data augmentation methods were established to manage different issues regarding the scarcity of training datasets or the class imbalance to enhance the performance of classifiers. This review article investigates data augmentation techniques for Arabic texts, specifically in the text classification field. A thorough review was conducted to give a concise and comprehensive understanding of these approaches in the context of Arabic classification. The focus of this article is on Arabic studies published from 2019 to 2024 about data augmentation in Arabic text classification. Inclusion and exclusion criteria were applied to ensure a comprehensive vision of these techniques in Arabic natural language processing (ANLP). It was found that data augmentation research for Arabic text classification dominates sentiment analysis and propaganda detection, with initial studies emerging in 2019; very few studies have investigated other domains like sarcasm detection or text categorization. We also observed the lack of benchmark datasets for performing the tasks. Most studies have focused on short texts, such as Twitter data or reviews, while research on long texts still needs to be explored. Additionally, various data augmentation methods still need to be examined for long texts to determine if techniques effective for short texts are also applicable to longer texts. A rigorous investigation and comparison of the most effective strategies is required due to the unique characteristics of the Arabic language. By doing so, we can better understand the processes involved in Arabic text classification and hence be able to select the most suitable data augmentation methods for specific tasks. This review contributes valuable insights into Arabic NLP and enriches the existing body of knowledge.