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

Related works

This section will summarize the review and survey articles discussing data augmentation or text classification in Arabic or general (in other languages). Feng et al. (2021), Shorten, Khoshgoftaar & Furht (2021), Bayer, Kaufhold & Reuter (2022) introduced the concept of data augmentation to the research community. Li, Hou & Che (2022) discussed DA methods in three categories namely paraphrasing, noise, and sampling to enhance data training diversity. The study highlighted the applications of DA, tasks, and challenges. At the same time, Bayer, Kaufhold & Reuter (2022) provided a comprehensive taxonomy of data augmentation methods. Their review offers a concise overview by highlighting promising approaches and future directions. Feng et al. (2021) discussed data augmentation techniques and their application in NLP. The study summarized the trends and challenges associated with DA approaches. Shorten, Khoshgoftaar & Furht (2021) summarized the significant data augmentation motifs and frameworks for textual data. The authors addressed the challenges of overfitting and generalization in deep learning and compared the application of DA to computer vision.

Meanwhile, Abdeen et al. (2019) discussed Arabic text classification over two decades (2000–2019). They reported on the techniques used for text classification, the most popular datasets, feature selection methods, and the effect of stemming methods. Aljedani, Alotaibi & Taileb (2020) reviewed multi-label Arabic text classification techniques and related applications and challenges. Alammary (2022) introduced a survey of BERT models and their applications in ATC. They compared Arabic BERT to English BERT from the performance perspective, underscored the limitations, and suggested areas for future research. Next, Alruily (2021) comprehensively analyzed the classification of Arabic Tweets which was carried out based on lexicon-based approaches, machine learning algorithms, as well as supervised, unsupervised, and hybrid learning. The review of Wahdan, Al-Emran & Shaalan (2024) focused on ATC, its area, application, and future research direction. The authors suggested different areas for exploration, such as handling unbalanced dataset issues and improving preprocessing techniques like stemming or tokenization.

Motivation of the study

Despite the growing body of research on text classification and data augmentation within the NLP field, there remains a significant gap concerning ATC specifically enhanced through DA techniques, as shown in Table 1. This review addresses this critical gap, offering the first comprehensive review of data augmentation strategies tailored for ATC. Existing reviews and surveys have either focused on ATC without delving into data augmentation (e.g., Abdeen et al., 2019; Aljedani, Alotaibi & Taileb, 2020; Alruily, 2021; Alammary, 2022; Wahdan, Al-Emran & Shaalan, 2024) or have explored data augmentation methods without specific attention to Arabic text (e.g., Feng et al., 2021; Shorten, Khoshgoftaar & Furht, 2021; Li, Hou & Che, 2022; Bayer, Kaufhold & Reuter, 2022). Consequently, this review strives to contribute substantially to the research community by synthesizing these two vital areas, providing a foundational resource for future research and practical applications in Arabic NLP.

Therefore, this review article addresses several research questions regarding applying DA techniques in the context of ATC. It tries to answer the following questions:

RQ1: What are the DA techniques used for Arabic text classification?

RQ2: Does the augmented data get evaluated?

RQ3: What domains use the DA techniques in the context of the Arabic language?

RQ4: What is the importance of DA techniques in the context of the Arabic language?

RQ5: What are the challenges of using DA techniques in the Arabic language?

The intended audience

Two distinct yet interconnected groups are the target audience for this research article. The first group is composed of academic researchers specifically those in the Arabic natural language processing (ANLP) domain who want to explore innovative DA techniques tailored to ATC and need to start or proceed with their work by exploring the latest techniques, what datasets, in what domains, and the challenges of data augmentation in text classification for the Arabic language. The second group is the computer science developers who are familiar with technical implementations but wish to better understand the specific challenges that arise in Arabic NLP and provide potential tools or already embedded methods to speed up the research for Arabic NLP. In addition to identifying challenges and future directions for Arabic NLP, this article highlights how these augmentation techniques can enhance the performance of text classification models. Furthermore, researchers and developers will be inspired to adopt these techniques and further advance ANLP.

The structure of this article is as follows: The next section explains the methodology used to conduct this review. It describes the inclusion and exclusion criteria with the assessment of the selected articles. Followed by analysis and discussion of selected articles. The analysis highlights in depth the DA techniques, the importance, and the challenges. In the last section, we will conclude the article.

Search strategy

The search was conducted using academic databases, including Web of Science (WOS), ACM Digital Library (ACM), IEEE Xplore Digital Library, ACL Anthology, and ScienceDirect (SD). Furthermore, a query was conducted on a connected article website to get the most relevant articles on Arabic data augmentation (ADA). The search was limited to materials published from 2019 to 2024 to obtain the most recent ones. Additionally, only journal articles and conference articles were selected. All the documents that can be accessed with their content were selected; otherwise, they were discarded. Meanwhile, different keywords were used including “data augmentation”, “Arabic text”, “paraphrasing”, “data generation”, and “text classification”.

Inclusion criteria

• –

  Articles published between 2019 and 2024,

• –

  Articles in English,

• –

  Articles including journal and conference articles,

• –

  Articles in full-text access,

• –

  Articles address data augmentation in Arabic text classification.

Exclusion criteria

• –

  Articles address data augmentation in other languages,

• –

  Articles address data augmentation in Arabic, but they are in image or voice,

• –

  Articles cannot access their content.

To assess the collected articles, first, articles were filtered by screening their titles and abstract contents to ensure that only relevant articles were selected. The availability of data augmentation techniques was the primary factor for selection. Then, we focused on the data augmentation methods for the Arabic language in text domains including sentiment analysis, text detection, text identification, or wherever it contains textual datasets. All the information was recorded for further analysis. In the end, only 20 articles out of 287 articles were found to comply with the required criteria and passed the filtration process. Then these selected articles were analyzed from different angles including motivation, data augmentation technique, domain, dataset, and data source as follows:

• Motivation

Most researchers conducted a study due to a noticeable problem in previous work. The criterion for motivation refers to why these DA approaches were applied in the research.

• Data augmentation techniques

Several methods are used to perform the augmentation, ranging from the simplest ones like feeding the punctuation marks to the dataset, to the complicated ones using the generative models. This criterion refers to the DA techniques used for the Arabic language.

• Domain

Different applications and domains use data augmentation methods. This criterion refers to the most prominent fields employing DA methods in the Arabic context. These domains may concentrate on text, speech, or images. However, the scope of this article is only on textual data for augmentation.

• Datasets

Applying data augmentation in different domains leads to the usage of various datasets. This criterion clusters the datasets for DA approaches based on their characteristics or applications to highlight the benchmark dataset in Arabic.

• Data source

Conducting data augmentation techniques on textual data triggers another criterion that points to the source of this data. This criterion will explain more about the data sources used in the selected articles such as social media, news, and reviews. Besides that, this criterion considers different varieties of Arabic, including Modern Standard Arabic (MSA) or different varieties.

RQ1: Data augmentation techniques in Arabic text classification

The application of DA techniques in ATC has gained attention in recent years. Research on DA in other languages, such as English, has been more prevalent. Exploring DA specifically for Arabic text is a relatively newer area of study (Duwairi & Abushaqra, 2021). This attention to using DA is driven by the growing interest in NLP research and the need to address the challenges of working with limited labeled or imbalanced data in Arabic. Figure 1 shows the overall trend of growing research interest and activity in the field over the past six years, despite some fluctuations. Different DA approaches have been applied to handwritten recognition, like in Alwaqfi et al. (2023), Eltay et al. (2022), Lamtougui et al. (2023) or speech recognition in Al-Onazi et al. (2022), Lounnas, Lichouri & Abbas (2022) as is easier to implement state-of-the-art methods in these domains.

On the other hand, Arabic textual data has not been extensively studied, as we will clarify below. Arabic DA techniques are more commonly applied to image datasets than textual data in the context of the Arabic language. DA is widely used in Arabic literature primarily to categorize sentiment (Refai, Abu-Soud & Abdel-Rahman, 2023) in Arabic short texts. However, further research and exploration of DA techniques in ATC tasks is necessary, particularly in fields beyond sentiment analysis. Applying DA to textual data to handle class imbalance or data adequacy makes ATC models more accurate and perform better. DA techniques in Arabic are classified into data transformation, data generation, or expanded datasets using external data. Figure 2 shows the taxonomy of these methods which will be clarified in the following sub-sections.

RQ 2: Evaluation of data augmentation in Arabic text classification

Assessing the effectiveness and impact of DA techniques in ATC is essential. This subsection focuses on analyzing the second research question of this article, which concentrates on evaluating the efficacy of these techniques by using specific metrics. The evaluation involves measuring the impact on text quality (Refai, Abu-Soud & Abdel-Rahman, 2023), model accuracy (Hamza et al., 2022), and task-specific performance such as on the Arabic dialect (Carrasco, Elnagar & Lataifeh, 2021; Talafha et al., 2019). This subsection aims to provide insights and understand how DA techniques have been assessed for the reviewed articles.

❖ Evaluating data generated by generation models

Studies in this category first evaluate the quality of generated data using similarity scores before proceeding to classification. This approach ensures that the augmented data is relevant and maintains semantic integrity before it is used for training machine learning models.

Similarity scores

Refai, Abu-Soud & Abdel-Rahman (2023) used four similarity scores namely Cosine (Zobel & Moffat, 1998), Euclidean (Danielsson, 1980), Jaccard (Ivchenko & Honov, 1998), and BLEU (Papineni et al., 2002) methods to ensure the context, semantics, novelty, and diversity of the generated data. Likewise, Carrasco, Elnagar & Lataifeh (2021) used the discriminator of GAN, but the authors modified it to use the penalty function to ensure diverse instances using the Jaccard similarity score. These evaluations shed light on the ability of DA techniques to generate additional instances that effectively capture the nuances and complexities of Arabic text.

❖ Evaluating data created by transformation methods or external dataset

Studies employed transformation methods, such as synonym replacement or back translation, or augmented using external datasets directly used the augmented data for training machine or deep learning models. The evaluation focuses on the performance of these models in classifying the augmented data.

Machine/deep learning metrics

Alkadri, Elkorany & Ahmed (2022) applied three classifiers namely naïve Bayes (NB), linear regression, and linear SVM, before and after the augmentation of spam detection tasks to check the improvement in the classification process after using the DA methods. Meanwhile, Beseiso & Elmousalami (2020) focused on enhancing the sentiment analysis for short Arabic Twitter texts. The authors performed the augmentation to increase the size of the dataset, hence they evaluated the augmented data using sentiment analysis classifiers. Mohammed & Kora (2019) assessed the performance of long short-term memory (LSTM) using four measures namely average accuracy, average precision, average recall, and average F1-score. Likewise, Omran et al. (2023) used the LSTM classifier with augmented data to monitor the impact of augmentation on classifiers. Wafa’Q, Mustafa & Ali (2022) used the AraBERT and MARBERT transformer models to analyze the data, while AlAwawdeh & Abandah (2021) utilized a multinomial NB classifier to assess the augmented data. Hamza et al. (2022) evaluated the performance of DA on bidirectional LSTM. The proposed augmentation method by Duwairi & Abushaqra (2021) was evaluated for accuracy using three ML classifiers: SVM, NB, and k-nearest neighbors. Abuzayed & Al-Khalifa (2021) employed MARBERT and QCRI Arabic and Dialectal BERT (QARiB) to assess the augmented data using the accuracy, F1-score, precision and recall. Gaanoun & Benelallam (2022) used naïve Bayes and several Arabic BERT models to evaluate the efficacy of augmented data through accuracy and macro F1 score. Badri, Kboubi & Habacha Chaibi (2024) utilized DziriBERT and BERT-based-Arabic models and got robust results for accuracy, F1-score, precision and recall. The AraBERT model is mostly used by several studies Gaanoun & Benelallam (2022), Hussein et al. (2022), Laskar et al. (2022), Fadel, Abulnaja & Saleh (2023), Shah et al. (2024) to investigate the effectiveness of augmented data in classification. Table 6 explains the evaluation type, methods, and metrics used to assess the augmented data.

Motivation

We observed that the main goal of using DA methods is either to tackle data scarcity (Mohammed & Kora, 2019; Talafha et al., 2019; Beseiso & Elmousalami, 2020; AlAwawdeh & Abandah, 2021; Carrasco, Elnagar & Lataifeh, 2021; Duwairi & Abushaqra, 2021; Hamza et al., 2022; Laskar et al., 2022; Gaanoun & Benelallam, 2022; Omran et al., 2023; Fadel, Abulnaja & Saleh, 2023; Badri, Kboubi & Habacha Chaibi, 2024) or class imbalance (Gaanoun & Benelallam, 2020; Abuzayed & Al-Khalifa, 2021; Wafa’Q, Mustafa & Ali, 2022; Alkadri, Elkorany & Ahmed, 2022; Hussein et al., 2022; Bensoltane & Zaki, 2023; Refai, Abu-Soud & Abdel-Rahman, 2023; Badri, Kboubi & Habacha Chaibi, 2024; Shah et al., 2024) to improve the performance of the models. Eleven studies, as clarified in Table 7, were used to handle the limited availability of the labeled training datasets. In comparison, nine studies were used to tackle the class imbalance problem.

Data augmentation methods

Different DA approaches created data to enhance the performance model. Some studies used more than one method for augmentation in case one is insufficient or for diversity. Hence, depending on the augmentation method used for this comparison criteria, each study was counted more than once. Two studies, Bensoltane & Zaki (2023), Beseiso & Elmousalami (2020) used language model predictions to augment the dataset. In addition, one study Bensoltane & Zaki (2023) used the back translation method. Two studies Alkadri, Elkorany & Ahmed (2022), Fadel, Abulnaja & Saleh (2023) used the word embedding to generate new data. Meanwhile, three studies Carrasco, Elnagar & Lataifeh (2021), Refai, Abu-Soud & Abdel-Rahman (2023), Shah et al. (2024) used generative models like GAN or GPTs models to perform the augmentation. Some studies focused on the linguistic characteristics of the language; methods like synonym replacement using Arabic WordNet or grammatical or syntactic rules and negation rules were applied (Duwairi & Abushaqra, 2021). The noise techniques such as random word swapping or shuffling (Mohammed & Kora, 2019; Omran et al., 2023), word deletion (Wafa’Q, Mustafa & Ali, 2022), and word insertion (Badri, Kboubi & Habacha Chaibi, 2024) were applied for the simplest augmentation. Furthermore, rule-based methods were employed by AlAwawdeh & Abandah (2021), Hamza et al. (2022) using transitive, symmetry, and reflexive relations to augment the dataset. Laskar et al. (2022) used the root/stem rules using an Arabic light-stemmer to produce new augmented data. Moreover, some studies Gaanoun & Benelallam (2020), Abuzayed & Al-Khalifa (2021), Hussein et al. (2022) augmented their original data by adding a new external dataset. Table 8 provides a summary of data augmentation techniques in Arabic text classification.

As we observed from Table 8, methods focusing on semantic preservation, such as paraphrasing via word embeddings or back translation, illustrate significant improvements in classification performance, particularly F1 scores. For instance, replacing similar words using word embeddings increased the F1 score from 65% to 89%. However, the dependency on pre-trained models and their inherent limitations, such as vocabulary constraints, points to a need for more adaptable and comprehensive embedding resources tailored for the Arabic language.

Conversely, noise-based techniques, such as random word swap or deletion, yielded inconsistent results. While these methods can introduce helpful variability, their unpredictable impact on semantic integrity suggests potential barriers. Random swapping achieved a moderate accuracy increase of 14.06% in some models, but often compromised the meaning of text, particularly for longer sentences. This suggests that noise-based augmentations need to be applied cautiously, potentially in combination with more sophisticated semantic-preserving methods to balance variability with coherence. The rule-based augmentation methods, mainly those employing syntactic and semantic rules, demonstrated considerable promise, with the highest reported accuracy of 93.86% using syntactic transformations like transitive or symmetric rules. This points out that rule-based methods can effectively capture linguistic transformations, offering more control compared to stochastic approaches. However, these approaches are labor-intensive and challenging to scale, as developing exhaustive linguistic rules is time-consuming.

The generation-based techniques, including those leveraging GANs, AraGPT2, and ChatGPT, showed significant improvements, with reported increases in F1 scores by up to 13% in certain datasets compared to producing synthetic data based on Mix method (split and merge) as illustrated in Table 8. Such methods create highly diverse and contextually relevant text, addressing the scarcity of labeled data. However, the computational cost and complexity of tuning these models are notable challenges. Despite these constraints, generation methods are powerful because they can synthesize realistic examples that enrich the training set far beyond what simpler strategies can achieve. Expanding data through external datasets also appears strategic to address the limitations of small native datasets in Arabic NLP. The integration of translated datasets, such as SemEval-2020, yielded a moderate boost to the performance of models like AraBERT, with an F1 score of 0.649. This highlights the benefits of cross-linguistic resource expansion, though translation quality and cultural context must be rigorously managed to avoid biases and inaccuracies. Consequently, a hybrid DA strategy for ATC is recommended to maximize effectiveness while balancing resource constraints and linguistic integrity. Despite the observations presented in this table, it is not possible to draw a definitive conclusion. It is important to note that the studies included in this article often do not use the same datasets or apply the same methods. Although in some cases the dataset is the same, the DA approaches vary. As a result, we cannot say with certainty whether this approach is suitable in all contexts.

Augmentation level

Most paraphrasing methods are applied to word-level augmentation since they are effective for altering a specific word to maintain the context and semantics of the text. Meanwhile, phrase or sentence-level augmentation is used mostly with rules to change more than one word. This augmentation shows significant improvement and is adequate for logical consistency and structure reasoning tasks. The Document-level augmentation using generation models or external datasets is highly effective in maintaining semantic coherence and augmenting diverse and contextual texts. However, it is essential to ensure the quality of the generated or added text and its alignment with the labels of the original dataset. Table 9 outlines the suitable level of augmentation for each DA technique.

RQ3: Domain

Figure 3 categorizes the DA methods on various application areas such as sentiment analysis (Mohammed & Kora, 2019; Beseiso & Elmousalami, 2020; Duwairi & Abushaqra, 2021; Abuzayed & Al-Khalifa, 2021; Refai, Abu-Soud & Abdel-Rahman, 2023; Omran et al., 2023), dialect identification (Carrasco, Elnagar & Lataifeh, 2021; Talafha et al., 2019; Gaanoun & Benelallam, 2020), sarcasm detection (Abuzayed & Al-Khalifa, 2021; Wafa’Q, Mustafa & Ali, 2022), question-answering tasks (AlAwawdeh & Abandah, 2021; Hamza et al., 2022), hate speech detection (Badri, Kboubi & Habacha Chaibi, 2024), spam detection (Alkadri, Elkorany & Ahmed, 2022) and propaganda detection (Gaanoun & Benelallam, 2022; Hussein et al., 2022; Laskar et al., 2022; Shah et al., 2024) and finally for aspect-based sentiment analysis (Bensoltane & Zaki, 2023; Fadel, Abulnaja & Saleh, 2023). Each domain’s prominence indicates where DA techniques are most actively applied and explored, underscoring the interdisciplinary nature of NLP research in ATC.

Data augmentation techniques have been employed across various domains to enhance the performance of the classification models. In sentiment analysis, multiple methods such as AraGPT2, pre-trained language model predictions, and syntactic rules with WordNet have enriched the dataset and improved model accuracy and robustness. For instance, AraGPT2 has been applied to generate new texts, significantly enhancing F1 scores across various datasets. Similarly, pre-trained language model predictions and the use of WordNet for synonym replacement, combined with syntactic rules, have shown substantial improvements in sentiment classification tasks. Methods like shuffling word order and using generative adversarial networks (GANs) have been proven effective in dialect identification. GANs introduce variability and increase the robustness of models to handle the intricacies of different dialects. For question prediction and detection, rule-based approaches have achieved high accuracy and F1 scores, demonstrating the effectiveness of structured augmentation in this domain.

The influence of data augmentation methods varies by domain, reflecting their potential suitability and impact. For instance, in sentiment analysis, techniques such as text generation have been suggested to help models generalize better by exposing them to a broader range of linguistic variations and expressions of sentiment. Similarly, GANs have been explored in dialect identification as they can simulate the structural patterns of words and phrases in different dialects. On the other hand, rules-based methods have been utilized in question prediction and detection to ensure that augmented data maintains logical consistency and relevance. However, it is essential to note that the limited number of studies addressing each of these applications makes it difficult to draw definitive conclusions about the most suitable data augmentation techniques for each domain. A more comprehensive examination of DA techniques across a broad range of studies is needed to validate their effectiveness for specific applications.

Dataset

The availability of benchmark datasets is rare when it comes to the Arabic language. As shown in Table 10, two studies Carrasco, Elnagar & Lataifeh (2021), Talafha et al. (2019) evaluated their methods on the same dataset, namely the Multi-Arabic Dialect Applications and Resources (MADAR) (Bouamor et al., 2018) dataset. This dataset contains 52,000 parallel sentences written in MSA and the local dialect of 25 cities. Besides that, two studies AlAwawdeh & Abandah (2021), Hamza et al. (2022) evaluated their rules on the same dataset, i.e., Mawdoo3. Sentiment analysis datasets mainly concentrate on tweets, product reviews, and hotel review data. Some studies used the Arabic Sentiment Tweets Dataset (ASTD) (Nabil, Aly & Atiya, 2015) for sentiment analysis (Beseiso & Elmousalami, 2020; Refai, Abu-Soud & Abdel-Rahman, 2023) as a benchmark dataset to conduct their experiments. Details of the datasets are explained clearly in Table 10.

Based on the dataset sizes presented in the table, there is a significant variation across Arabic NLP tasks and domains. A dataset such as Multi Arabic Dialect Applications and Resources (MADAR), used for dialect identification, contains more than 52,000 samples, while others have substantially fewer. In the case of product reviews, for example, only 300 reviews are included in the sentiment analysis dataset. As a result of this disparity in dataset sizes, DA techniques have become imperative, particularly in domains with limited resources or a lack of high-quality annotations. There is a complex relationship between dataset size and domain, and it varies depending on the task at hand. There are many types of sentiment analysis datasets, such as reviews and tweets, which range in size from just 300 to 63,000 tweets (LABRv2) (Aly & Atiya, 2013). The variability in data availability suggests that certain domains, like social media, are more easily accessed, whereas others, like product reviews, face more challenges during data collection and annotation. The necessity of DA becomes especially evident in domains with limited data, since models developed from small datasets may not be generalized in real-world situations.

Data source

Table 10 exhibits a range of data sources, including curated datasets like Quora and iSarcasmEval (Farha et al., 2022), social media data like Facebook and Twitter, and product reviews from e-commerce sites like Amazon and souq.com. This range offers a good foundation for using data augmentation methods such as noise injection, paraphrasing, and back translation. For example, augmentation techniques like back translation might be helpful when product reviews were translated from English into Modern Standard Arabic (MSA) and then Bahraini. By using this technique, any potential translation-induced biases can be identified and mitigated by translating the content back into its native language. Additionally, it increases the dataset’s linguistic diversity, which raises the accuracy of the models that were trained on it.

Furthermore, augmentation methods can help handle informal language, and dialects in social media data, which is a significant element in the datasets (e.g., LABRv2, AraSarcasm (Farha, Zaghouani & Magdy, 2021)). Using strategies like paraphrasing or word substitution can create diverse data making the models more robust to different variations of the same sentiment. Augmentation confirms that data covers several linguistic expressions which are crucial for domains like sentiment analysis and sarcasm detection. Eventually, applying data augmentation techniques to these different data sources helps improve model performance by reducing biases and increasing the generalizability of the training data across various domains and dialects.

RQ4: Importance of data augmentation in Arabic text classification

The scarcity of labeled datasets or imbalanced classes for ATC poses significant challenges. DA techniques offer a solution by artificially expanding and diversifying the dataset. In this context, data augmentation plays a crucial role in improving the performance of classification models. It enables the modeling of a more prominent and representative dataset, ultimately enhancing the effectiveness of ATC. In the following points, we will further explore data augmentation’s specific benefits and advantages to ATC.

• Addressing data scarcity: This review article provides a comprehensive overview of different data augmentation (DA) techniques used to manage data scarcity in low-resource languages like Arabic. It examines the most effective DA techniques to overcome the limitations of training data required for classification models by considering methods such as paraphrasing, noise addition, and data generation. For example, as mentioned earlier in the RQ1 section, Carrasco, Elnagar & Lataifeh (2021) utilized GANs to generate robust text, thereby overcoming data scarcity and improving classification models.

• Increasing training data: This article also explores how DA can enrich training datasets. It extends this line of research by examining the implementation of various DA techniques to augment training data. For instance, Beseiso & Elmousalami (2020) utilized a language model (LM) trained on 70% of their dataset to predict the last word and expand the dataset. This approach ensures augmentation with contextually appropriate data. They introduced diverse features into the dataset by selecting the top three predictions of the language model as a paraphrasing approach. For example, with the input sentence “I have to go to”, the LM might predict “school, market, cinema” thereby enriching the dataset with new features. Similarly, using masked BERT, Bensoltane & Zaki (2023) created contextual semantic data which is crucial for making diverse training datasets and enhancing model performance.

• Mitigating imbalanced class problems: This article contributes to understanding class imbalance in classification datasets by discussing various DA methods to address this issue. For instance, Refai, Abu-Soud & Abdel-Rahman (2023), Shah et al. (2024) examined the use of AraGPT2 or ChatGPT to generate new data for classes with minority samples. By expanding these classes, the authors balanced the overall dataset. The GPT model leverages powerful pre-trained models on extensive data to produce effective results, thereby addressing class imbalance.

RQ5: Challenges and future work

Data augmentation is instrumental in enhancing the data pool for ATC. However, as transformative as these practices are, they face specific challenges due to the distinctive characteristics of the Arabic language. Linguistic complexities, multiple dialects, limited resources, and data imbalance weigh heavily on the efficacy of these techniques. The following points will delve into these challenges to advance the understanding and effectiveness of data augmentation in ATC.

• Lack of research: Data augmentation research for the Arabic context is limited compared to English. Future research should focus on comprehensive studies covering various Arabic text types. It may also be possible to identify the most promising DA techniques for Arabic text in future research by comparing their effectiveness and whether there is a difference between applying them to short text or long text, or MSA or dialect, as well as what DA methods are effective for each type of text. Therefore, researchers can provide more tailored solutions for Arabic NLP tasks. For example, generative models like GANs or GPT models can be employed widely for long texts to examine these techniques in such texts.

• Generalization: Due to the scarcity of research in this area, we cannot definitively generalize a single method of DA across all contexts. Future research should aim to develop and test DA methods tailored explicitly for short and long texts, given the unique features of MSA and dialects. When it comes to longer texts, it is essential to ensure that procedures such as synonym replacement, word shuffling, and rule-based approaches are effective without resulting in redundancy or incoherence.

• Inability to ensure the effectiveness of the DA method: There is a need for more rigorous evaluations of data generated by DA techniques, specifically for those involving generative models like GPT. Future research should examine the applications of the performance model and the novelties, diversity, and semantics of the generated data rather than just enhancing the performance model. As a result, different data augmentation techniques should be investigated from a more complete perspective.

• Focus on short texts: Most current studies on Arabic DA for text classification focus on short texts. Future research should include longer texts, developing and validating DA methods that maintain coherence and relevance over extended passages. Addressing the challenges and advantages of DA techniques will enhance the robustness and applicability of these methods in diverse Arabic NLP tasks.