Artificial Intelligence-Based Approach for Misogyny and Sarcasm Detection from Arabic Texts

Social media networking is a prominent topic in real life, particularly at the current moment. The impact of comments has been investigated in several studies. Twitter, Facebook, and Instagram are just a few of the social media networks that are used to broadcast different news worldwide. In this paper, a comprehensive AI-based study is presented to automatically detect the Arabic text misogyny and sarcasm in binary and multiclass scenarios. The key of the proposed AI approach is to distinguish various topics of misogyny and sarcasm from Arabic tweets in social media networks. A comprehensive study is achieved for detecting both misogyny and sarcasm via adopting seven state-of-the-art NLP classifiers: ARABERT, PAC, LRC, RFC, LSVC, DTC, and KNNC. To fine tune, validate, and evaluate all of these techniques, two Arabic tweets datasets (i.e., misogyny and Abu Farah datasets) are used. For the experimental study, two scenarios are proposed for each case study (misogyny or sarcasm): binary and multiclass problems. For misogyny detection, the best accuracy is achieved using the AraBERT classifier with 91.0% for binary classification scenario and 89.0% for the multiclass scenario. For sarcasm detection, the best accuracy is achieved using the AraBERT as well with 88% for binary classification scenario and 77.0% for the multiclass scenario. The proposed method appears to be effective in detecting misogyny and sarcasm in social media platforms with suggesting AraBERT as a superior state-of-the-art deep learning classifier.


Introduction
Posts and tweets on social media platforms allow people to share their thoughts, emotions, and sentiments. Regularly, Internet bullying is observed, which is considered a form of attacking, and it has escalated against people [1,2]. with a growth in the number of people using social media sites such as Twitter to express themselves [3]. e lack of tools and benchmark datasets are the main challenges in this field [4]. e goal here is to construct an accurate method for detecting misogyny and sarcasm from Arabic text. e limitation of studies in Arabic context serves as motivation to research and build practical smart solutions through the design and development of automatic identification systems [5].
ere are around 447 million native speakers who speak Arabic as their first language, with many more speaking it as a second or religious language. As a result, Arabic has become one of the world's most frequently spoken languages. At the same time, Arabic is the primary language of 22 Arab countries and many other Islamic countries. e writing of Arabic is written from right to left. It is the world's fifth most spoken language [6][7][8]. Arabic has a rich morphology and a complex orthography [9].
Text detection is one of the big challenges of NLP which still has many limitations to work with, especially for the Arabic language when compared to English. ere are many types of text detection such as prediction of human behavior [10], hate speech detection [11], exploring halal tourism [12], gender detection [13], misogyny detection [14], sarcasm detection [15], detection and classification of psychopathic personality [16], fake news detection [17], and detection of dialectal in the Arabic language [18].
is study aims to design a model for automatically detecting misogyny and sarcasm using machine learning (ML) and deep learning (DL) methods with different benchmark datasets [19,20]. e proposed study is evaluating state-of-the-art deep learning classifiers for multiple topics (i.e., misogyny and sarcasm) at the same time. Whereas most Arabic text detection latest studies were focusing only on a single topic. e requirements to develop a method for detecting any type of Arabic text, such as sarcasm and hate speech, represent the Arab society needs. It is critical since our system must be able to reliably detect any suspicious text. e contributions for this work are summarized as follows: (i) A comprehensive study is performed via seven ML/ DL classifiers with two scenarios: binary and multiclass classification problems (ii) Two different datasets are used to evaluate the proposed research concept for distinguishing Arabic misogyny and sarcasm from Arabic tweets (iii) Multiple topics of misogyny and sarcasm are simultaneously detected via the proposed scenarios (iv) e proposed study investigates the best deep learning classifier for Arabic texts in terms of increasing the surfing life on social media networks e remainder of this paper is organized as follows: In Section 2, we address the related works in misogyny and sarcasm identification in Arabic. Section 3 illustrates the suggested model and model architecture. Following that, in Section 4, we present the experimental analysis. After that, the discussion will be given in Section 5, and the conclusion will be presented with a brief summarization about the future work in Section 6.

Related Works
Detecting sarcasm text is one of several tasks in NLP. ere exist some work in the literature, which has been done for Arabic text detection. As we mention above, there are many types of sarcasm detection of Arabic text, and where we plan to survey some of these works is as follows: Detection of Arabic text is still in its early stages [21]. Lichouri et al. in [22] presented a comparison between two LSVC and BiLSTM classifiers for Arabic text detection. ey presented an intuitive but straightforward detection system based on different stages and combinations. Mulki and Ghanem in [2] introduced an Arabic Levantine Twitter dataset for Misogynistic language (LeT-Mi) to be the first benchmark dataset for Arabic misogyny. Muaad et al. in [23] presented a model to detect misogyny from Arabic text and compare between machine and deep learning algorithms. Al-Yahya et al.in [24] made a comparative study of representation transformer-based language and classification neural network models to detect text in general. Gaanoun and Benelallam in [25] designed a new system using a hybrid ensemble of Gaussian Naive Bayes, MarBERT, and Mazajak embedding. is study has got an F1-sarcastic score of 51% on sarcasm and an F1-PN of 71% for sentiment detection. Husain and Uzuner in [26] designed the system to detect sarcasm and sentiment analysis for Arabic text. Naski et al. in [27] designed a system to classify sentiment for four languages and to detect sarcasm using (mBERT, AraBERT, ARBERT, and MARBERT). Talafha et al. in [28] tried to solve the problem of sarcasm detection from Arabic texts using different methods. e work solves the task as a regression problem; also, this paper finds sarcasm in a given tweet. Faraj et al. in [29] joined the shared task on sarcasm and sentiment detection for the Arabic language. eir goal was to detect whether a tweet was sarcastic or not. ey have used an AraBERT pretrained model with an ensemble technique. eir model achieved an accuracy score of 0.78, 30 in this competition. Abuzayed and Al-Khalifa in [30] implemented seven DL models by using an augmentation technique. eir models were able to detect whether a tweet is a sarcasm or not. Wadhawan in [31] used the ArSarcasm-v2 dataset. ey processed it by altering different parts of the text. en, experiments with AraELECTRA and Ara-BERT transformer-based models have been conducted. e tasks of sarcasm and sentiment detection have been conducted. Abu Farha et al. in [32] provided an overview of the new ArSarcasm-v2 dataset, which is used for the shared task. ey also provided a high-level description of the top participating teams in the shared task. In the end, the best 0.62 F1-score and 0.74 FPN results for both sarcasm detection and sentiment analysis tasks have been registered. El Mahdaouy et al. in [33] introduced an end-to-end deep MTL model, allowing knowledge interaction between the two tasks. According to our relevant research, there are limited studies in detecting Arabic text for more than one topic simultaneously, and detection text can contain many topics.

Architecture of the Proposed Deep Learning Framework for Arabic Texts.
e architecture stages of the misogyny and sarcasm detection approach are given in Figure 1. It consists of three main steps: dataset preparation and splitting, Arabic text representation, misogyny and sarcasm detection, and evaluation scenarios. e preparation text for further processing can be performed through the preprocessing step.
en, Arabic word representation and feature extraction steps could be performed. Finally, the classification part is performed for both binary and multiclass scenarios.

Preprocessing.
Text must be preprocessed in order to be more useful for representation and learning tasks. Preprocessing is a technique for converting data into a specified input data format that might be useful for ML and DL techniques. e primary goal of preprocessing is to remove punctuation, slang, stop words, and other unwanted words that exist in text data. e identification performance of the Arabic misogyny and sarcasm detection task may be harmed by this unpleasant word in the Arabic language. rough a preprocessing phase, we exclude any non-Arabic terms, stop words, and punctuations in this work. Most of the time, we need to work with different preprocessing or combinations of them as we have performed here by using tokenization and stemming. A good text representation can be learned using a variety of conventional and advanced preprocessing approaches [34,35]. Tokenization is the process of dividing a text into words, phrases, or characters called tokens. Whereas here, the model converts a text (sentence) into words. en, these words will be converted to weights or embedding vectors depending on which representation we will select. en, stop word removal has been performed with machine learning.

Arabic Text Representation (Feature Extraction).
Preprocessing data will be ready for representation purposes. Various feature extraction models are used in this section. In the experimental part, the performance was poor with BoW for a machine learning method due to the loss of semantic and syntactic information between words. As a result, we use another representation technique known as AraBERT to handle semantics and syntactic which increase performance by increasing accuracy.
e BoW is merely a one-hot encoding extension. It is utilized in a variety of fields, such as NLP. e semantic relationship between words, as well as the order of words and grammar, is neglected in the matrix of words created with BoW. We will explore the other type of representation in the next section.

Arabic Text Classification.
Text is recognized and classified to a right label called a classification task. Several algorithms have been implemented here, as we will discuss in Section 5. We will explore different ML and DL techniques called AraBERT. BERTrefers to bidirectional encoder representations from transformers which is a contextualized word representation employed in various domains, including natural language understanding, sentiment analysis, and natural language generation, such as text translation.
e BERT training is based on two tasks. e first task is a masked language model, and the second task is the next sentence prediction. We adopted the AraBERT model proposed in [36], which is an effective model for the Arabic language text representation and classification task. We implemented some classifiers such as passive-aggressive classifier as (PAC), logistic regression classifier (LRC), random forest classifier (RFC), linear SVC classifier (LSVC), decision tree classifier (DTC), and K nearest neighbor classifier (KNNC).  e misogyny dataset is prepared for binary and multiclass problems, as shown in Tables 1 and  2. Table 1 shows the misogyny data distribution in the case of binary classification scenarios (i.e., non-misogyny vs. misogyny).

Experimental Analysis
Whereas, Table 2 shows the misogyny data distribution in case of multiclassification problem. In this case, eight different classes are considered which are discredit, stereo typing and objectification, damning, threat of violence, derailing, dominance, sexual harassment, and non-misogyny.

Sarcasm Dataset.
In the same way, the sarcasm dataset is prepared for binary and multiclass problems. Tables 3 and  4 show the detail of data distribution per class for the binary and multiclass scenarios, respectively.

Evaluation Metrics.
For this study, the evaluation metrics of recall, precision, F1-score, and accuracy are used. e mathematical definition of these indices are expressed as follows: where TP refers to true positives, TN refers to true negatives, FP refers to false positives, and FN refers to false negatives. ese parameters are derived using the confusion matrices for both classification scenarios: binary and multiclass problems [36][37][38][39][40].

Discussion and Results
is part highlights that the results and discussions of the different methods have been used describing the several trials with two different datasets. ese experimentations are focused on two tasks: binary classification and multiclassification for both misogyny and sarcasm. Table 5 and Figure 2 show the results of the misogyny detection for binary classification. e linear SVC model was better than other ML in terms of accuracy, precision, recall, and F-1 score. At the same time, we executed AraBERTv2, as transfer learning algorithm that provides high accuracy but it costs in terms of execution time more than other ML techniques.

Misogyny Multiclass Classification.
Multiclassification is more complicated than binary classification, so learning the pattern here is less accurate as we have seen in Table 6. e results of multiclassification have been presented in Table 6 and Figure 3. e linear SVC model performs better than the others in terms of accuracy and the random forest model performs poorly, according to the findings.
Simultaneously, we were using ARABERTv2 that provided outstanding accuracy but required more time than others. Lastly, we do want to point out that an imbalanced dataset will have an impact on categorization accuracy. As we have seen in Table 2, the class dominance contains only 38 comments document after dividing data to train and test the number of documents will be very less and learning about this class will be very less too. So, augmentation of data is recommended for future work. Table 7 and Figure 4 illustrate the results of the sarcasm classification for the sarcasm dataset. e linear SVC model exceeds all ML model in terms of accuracy, precision, recall, and F1-score. At the    Computational Intelligence and Neuroscience 5 same time, we employed ARABERTv2, that provided good accuracy but took a long time compared to the ML model. Table 8 and Figure 5 show the outcomes of the multidetection and categorization process. e linear SVC model in terms of F1-score outperforms the others, but the logistic regression model in terms of precision was better, and both of the models were equal in terms of accuracy and recall. According to the findings, the performance of the random forest classifier model is weak. At the same time, we used ARABERTv2, that provided great accuracy but consumed more time than the ML approach.

Sarcasm Multiclass Classification.
Finally, we would like to draw attention to the fact that the dataset was skewed. For example, as seen in Table 4, the positive class contains only two thousand five hundred, but other classes have more than six thousand, so the learning for these classes will not be same, and working with augmentation technique will be better for getting good accuracy as future work.
From the results in Tables 5 and 6 for the binary classification as well as Tables 5 and 6 for the multiclassification, we see that the accuracy 90% and 89% for binary classification were better than multiclassification which were 82% and 77%, respectively. Finally working with deep learning gives better results but a lack of data and unbalance of the dataset is one of the challenges for researchers in Arabic natural language processing.

Conclusion
e problem of the sarcasm detection of Arabic text has become a major problem for Arab people. In this work, we introduce a model to detect misogyny and sarcasm from Arabic text. We have carried out our work utilizing two datasets called (misogyny and Abu Farah datasets). Our results provide excellent accuracy equal to 91% and 88% for binary classification, respectively and 82% and 77% for multiclassification, respectively, using the

Computational Intelligence and Neuroscience
AraBERT deep learning classifier.
is work indicates that there are various outstanding issues, starting with finding benchmark datasets and lexicons for Arabic text detection topics. At the same time, data augmentation techniques can be used to address dataset imbalances. Finally, future studies should examine the model to find the relation between various topics and the mixed language problem.

Data Availability
e data used to support the findings of this study are available from the corresponding author upon request.

Conflicts of Interest
e authors declare no conflicts of interest.