Detection of cyberhate speech towards female sport in the Arabic Xsphere

Methods

Ethical approval was obtained from the Institutional Review Board of PNU University before commencing the study (23-0734). The data was collected through semi-structured interviews with The data was collected through semi-structured interviews with 33 Saudi females, who are either athletes, players for professional sports clubs, players for university clubs, or sports fans. Most of them mentioned having a presence on social media, either public or private. The majority of participants, totaling 22, are aged between 18 and 25. There are fewer participants, only 3, aged between 26 and 30, and just one person aged 31 to 36. Among the fans, 5 are aged between 18 and 25, while only 2 are 37 or older. This suggests that the majority of individuals involved are young, with only a few older participants and fans. Appendix S1A presents each participant’s demographics.

The interviews were conducted either face-to-face or online, based on the participant’s preference or request, demonstrating flexibility in the data collection process. The engagement with the participants commenced by presenting the study’s objectives and rationale, followed by obtaining their informed consent, ensuring ethical research practices. After the participants signed the consent forms, the interviews began.

The interview questions in Appendix S1B were crafted drawing from prior research on cyberhate and bullying, exemplified by studies like (Dredge, Gleeson & De La Piedad Garcia, 2014; Nyman & Provozin, 2019), with a focus of exploring the experiences of female athletes and sports fans with hate speech in online communities.

The average interview time was approximately 20 min. All interviews were audio-recorded. The researchers have transcribed all interviews and subsequently translated them into English.

Data analysis

The researchers analyzed the transcripts with directed content analysis, which “is the intellectual process of categorizing qualitative textual data into clusters of similar entities, or conceptual categories, to identify consistent patterns and relationships between variables or themes” (Julien, 2008).

Results

Based on the analysis, three themes were discovered that answer

RQ1: What are the specific forms of hate speech experienced by female athletes and female sports fans on social media?

Theme 1: Gender-based derogatory comments

Gender-based derogatory comments include derogatory comments and slurs based on gender. Female athletes and fans may be targeted with sexist insults and language meant to demean or belittle them based on their gender. Twenty-one out of 33 participants talked about unfair treatment on their social media platforms. If a mistake they made became public, they were more likely to receive abuse for it than men.

Participants described receiving comments suggesting that women do not belong in certain sports or that they should confine themselves to traditional gender roles.

“They tell me go to your house, your kitchen is your home, you have nothing to do with football, they always say it like that.” (interview 17)

All the comments women receive say that they do not belong, whereas when a mistake of a male sportsman becomes public, none of the comments he receives say that he does not belong in that sport.

One participant stated:

“If a female coach were to state a sports fact, many would not believe her. On the other hand, if a male coach were to state the exact same fact, everyone would believe him.” (interview 9)

Participants also stated that when men find a woman on a social media platform who seems to be better at a sport than they are, these men make comparisons and say things like “they are only better because they’re on steroids,” implying that no woman can be better than a man in any sport without additional help.

Theme 2: General misogyny

General misogyny in hate speech reflects underlying biases and prejudices against women in sports, specifically female athletes. There is a discrepancy in how female athletes and their performance are treated on social media: When they win, they do not receive the support and recognition they deserve. In contrast, when they lose, they are bombarded by negative comments. This suggests that the mistreatment is based on their gender rather than their actual performance.

Sixteen out of 33 participants mentioned that female athletes do not get any support from men if they win, but if they lose, they get many negative comments on their social media platforms.

“Yes, many tweets would be about the victory of the girls or about positive things if they had a victory or achievement, but I would not find any males in the comment section. But if there was a mistake or anything negative, males would be present in the comment section and they would criticize and say the reason for this is because she is a girl, if she was a boy, she would not have made that mistake.” (interview 14)

Some participants stated that when a women’s sports team loses, all the comments have comments such as that women do not belong in sports. However, when a men’s sports team loses, the comments may also be negative, but they never mention that men do not belong there.

“I mean okay, all of it is criticism, but they differ, I mean, for the girls, their criticisms are ridiculous, as if they say that you have no place in the field of sports and the men’s criticism, it has to do with football, criticism of the player, he is not playing well.” (10).

“In football matches, when a male makes mistakes, it’s normal, it’s a normal thing in the stadium, people make mistakes, but when they see it from a female, they see it as a lack of awareness of the thing they are doing, as if they are just playing and they don’t know anything in the first place, this is a societal view of women.” (interview 13)

Another way players can be affected is when their family names are out in public; most women are encouraged to participate under their first names only, so as not to focus attention on their families, which corresponds with the cultural values of the Middle East. This is closely tied to misogyny because it diminishes their recognition, reinforces traditional gender roles, and imposes conformity, reflecting sexist and misogynistic norms. Consequently, this practice may limit the personal recognition and visibility of these athletes, making it challenging for them to build individual brands and careers in sports.

“I mean I faced lack of support, but in terms of name, I mean, it’s OK to be public on social media, but the name shouldn’t come out, as in use a nickname.” (interview 10)

Theme 3: Appearance-related discrimination

Appearance-related remarks involve hate speech that centers around an individual’s physical appearance, including their body, clothing, or overall look. Eight participants said that they receive many negative comments if their physical appearance did not represent that of a typical woman. For instance, a participant stated: “I noticed they bully girls; they comment on their shapes and athletic bodies . Unfortunately, there is still a certain stereotypical thinking that girls should look a certain way, and this has nothing to do with it. Every person is free to appear in the way they like” (7).

Another participant talked about comments she has received: “Getting comments on the body in a very unbearable way;” “it can be sexual comments” (interview 32).

“ A strong girl who lifts heavy weights , for example, for sure she’s taking something, it’s impossible to be normal.” (27)

“I am going to be honest; boys see a girl who is muscled, sometimes they would come and say why is she more muscular than me and whatever else. I see it in the comments, they don’t like to see anyone better than them, I mean, they have the madness of greatness.” (interview 16)

“I see that with females, I mean they emphasize her outward appearance more than her skills…and if she makes a mistake they say, this is not your place; you should be studying or tending a home, or something like that, this is not for you.” (interview 6)

Although some excerpts touch on gender-related issues, they each highlight distinct forms of discrimination encountered by female athletes. For example, while it may seem that excerpt (interview 9) and (interview 27) are similar, excerpt (interview 9) sheds light on the credibility challenges female coaches encounter, such as receiving less recognition and support. In contrast, excerpt (interview 27) underscores the damaging stereotypes surrounding the physical appearance of female athletes. Thus, the excerpts are appropriately categorized under their respective themes based on the nature of the discrimination they illustrate.

Methodology

As this research aims to develop an automated classifier to classify Arabic tweets using natural language processing (NLP) techniques for the purpose of detecting hate speech targeted at women in sports, the methodology covers data collation, annotation, pre-processing, classification, and evaluation. This approach is crucial for obtaining reliable and valid results in any research study. In this section, we outline the specific steps taken to achieve our research objectives, including the model architecture for Arabic hate-speech detection (Fig. 1).

Data annotation.

Annotation is crucial in developing a supervised machine learning model. In this case, the dataset was divided among six annotators (authors of this work), each responsible for labeling 1,247 tweets. All of annotators are Arabic-speaking Saudi female bachelor students majoring in computer science. The annotators were trained to identify tweets containing hate speech against women in sports. The qualitative data from interviews aided in the development of annotation criteria. Researchers used participant insights to identify what constitutes a hateful comment. The team encountered challenging cases during annotation, where the tweet’s content was ambiguous. To address these issues, meetings and discussions were held among the annotators to reach a consensus on label definitions, ensuring that all annotators were satisfied with the labelling criteria. This collaborative approach helps maintain the quality of the labeled data, which is essential for developing an effective machine learning model for detecting hate speech. To guarantee the precision and uniformity of the annotations, 1,247 tweets annotated by different annotators were randomly selected and reviewed by other annotators. The annotators agreed on most of the labels. The annotation process involved three distinct labels, as described in Table 1. This table provides the definitions, examples, and counts for each label used during annotation. To validate the team’s annotations, we hired an external annotator—an Arabic-speaking Egyptian male postgraduate in computer science—and paid $100 to review all tweet labeled as 0 and 1. The inter-annotator reliability between the team and the external reviewer was calculated using Cohen’s kappa (Cohen, 1960) (resulting in a strong agreement level of 91%.

Table 1:

Overview of dataset labels.

Label	Definition	Count	Example	Translated example
0(Support)	A tweet containing expressions that show support or encouragement for women participating in sports activities or events.	1,787		The best thing is that sports are beneficial for girls and their health. Congratulations to all Saudi girls.
1(Hateful)	A tweet containing expressions that show negativity or hate towards women participating in sports activities or events.	2,025		Sports were ruined when women became its fans your place is in the kitchen.
				The stupidity of girls continues and does not end cumulative genetic stupidity What did your mother know about football?
				Women getting involved in sports is disgusting
				God does not help who teaches girls football
				Women’s bodies are built for KABSA and LUQAIMAT, what brought them to sports
2(Irrelevant)	A tweet that does not contain any expression or opinion towards women participating in sports activities or events.	3,675		The account name should be Women’s Sports News.

DOI: 10.7717/peerjcs.2138/table-1

Feature extraction.

After performing essential data preprocessing tasks such as text cleaning and stop word removal, two distinct feature extraction methods were applied to the preprocessed tweets TF-IDF (Qaiser & Ali, 2018) and POS tagging. TF-IDF assigns weights to words based on their frequency within a document while considering their inverse frequency across the entire dataset. This approach helps identify significant words within each document, facilitating subsequent classification. The TF-IDF technique is commonly used in hate speech classification, because it effectively highlights important words in a text. Recent research, like that by Pariyani et al. (2021) and Mugambi (2017), has demonstrated the utility of TF-IDF for identifying hate speech on platforms like X and in other modern social media contexts. This technique is valued for its ability to uncover significant linguistic patterns in hate speech (Mugambi, 2017; Qaiser & Ali, 2018).

Table 2:

Examples of tweets before and after applying pre-processing steps.

Before pre-processing	After pre-processing

DOI: 10.7717/peerjcs.2138/table-2

POS involves assigning a grammatical category (e.g., noun, verb, adjective) to each word in the text. This enriches the feature extraction process by providing information about the grammatical structure and contextual usage of words in the documents. The use of POS tagging in hate-speech classification has gained prominence due to its effectiveness in capturing linguistic nuances. Recent studies by (Watanabe, Bouazizi & Ohtsuki, 2018; Bauwelinck & Lefever, 2019) and Lefever emphasize the importance of linguistic analysis and emotional context, making POS tagging a relevant and valuable technique for hate-speech detection.

The tool we used to generate the features is the Arabic Linguistic Pipeline (ALP) developed by (Abed Alhakim et al., 2022). The choice for ALP was made based on a comparative study of the five most common Arabic POS taggers, the Stanford Arabic tagger (https://nlp.stanford.edu/software/tagger.shtml), CAMeL Tools (https://github.com/CAMeL-Lab/camel_tools), MADAMIRA (https://camel.abudhabi.nyu.edu/madamira), Farasa (https://farasa.qcri.org/POS), and ALP (http://arabicnlp.pro/). The study used text samples from Saudi novels and found that the ALP outperformed the other taggers (Alluhaibi et al., 2021).

ALP helped in the extraction of POS features and generated 245 different Arabic POS tags for class 0 and 225 tags for class 1, adding a valuable linguistic dimension to the feature set for classification. Table 3 shows a sample of tagged tweets classified as 1.

These two methods, TF-IDF and POS, were used independently, each contributing to the extraction of valuable features for classification.

Model training.

Following feature extraction, we trained several traditional supervised machine learning models to classify tweets as either hate speech or non-hate speech. While deep learning algorithms have shown promise in various NLP tasks, including hate speech detection (Badjatiya et al., 2017), they typically require very large amounts of data to perform well (Dhola & Mann, 2021; Mujahid et al., 2023) and our dataset size may not provide sufficient examples for effective training. Therefore, we chose traditional machine learning algorithms which are known to perform well with relatively small dataset size and can provide interpretable results.

In this study, four machine learning models from various classes have been applied, including linear, non-linear, tree-based, and non-tree-based methods. Most of these algorithms—SVM, Random Forest, Decision Tree, and XGBoost—are widely recognized for their effectiveness in hate speech text classification tasks across different languages. All models were implemented in Python version 3.11 using Anaconda 7. A brief description of each model is provided below:

Table 3:

Sample of tagged tweets using the ALP tool.

Pre-Processed Tweets	Tagged Tweets using ALP

DOI: 10.7717/peerjcs.2138/table-3

• Support vector machines (SVM) (Cortes & Vapink, 1995): SVM is a widely used machine learning algorithm for classification and regression tasks. It works by maximizing the margin between the data points of different classes, using a decision boundary called a hyperplane.

• Random Forest (Breiman, 2001): Random Forest is an ensemble learning method that consists of multiple decision trees. It leverages a combination of bootstrap aggregating (bagging) and feature randomization to improve predictive accuracy. The individual decision trees are trained on random subsets of the data, and their outputs are combined to make predictions. This ensemble approach enhances model performance and reduces the risk of overfitting.

• Decision Tree (Quinlan, 1986): Decision Trees are hierarchical structures used for decision-making in classification and regression tasks. They recursively split the input data into subsets based on the values of input features, with the goal of maximizing the homogeneity of the resulting subsets.

• Extreme Gradient Boosting (XGBoost) (Chen & Guestrin, 2016): XGBoost is an optimized implementation of gradient boosted trees, which combines multiple weak classifiers (decision trees) to form a strong classifier. It uses a technique called boosting to iteratively improve the model by refining the weights of the weak classifiers.

These machine learning algorithms were selected due to their ability to handle high-dimensional data and their effectiveness in various text classification tasks. By training and comparing the performance of these models, we aimed to identify the best classifier for detecting hate speech against women in sports in the Arabic tweets dataset.

Model evaluation

In our model evaluation, we conducted a thorough assessment of two distinct natural language processing tasks: TF-IDF and POS tagging. To validate our findings, we split the dataset into 70% for training and 30% for testing, and we employed a 10-fold cross-validation approach as suggested by Kohavi (1995). We measured the effectiveness of our trained models using standard machine learning performance metrics, including accuracy, precision, recall, and F1-score, which are widely recognized for evaluating model performance (Sokolova & Lapalme, 2009).

• True positives (TP): True positives represent the instances where the model correctly predicts positive cases.

• False positives (FP): False positives occur when the model incorrectly predicts positive cases that are actually negative.

• Accuracy: Accuracy is a measure of the proportion of correct predictions made by the model over the total number of predictions. It provides an overall assessment of how well the model is performing. A c c u r a c y = T r u e P o s i t i v e s + T r u e N e g a t i v e s T r u e P o s i t i v e s + F a l s e P o s i t i v e s + T r u e N e g a t i v e s + F a l s e N e g a t i v e s

• Precision: Precision is a metric that quantifies the accuracy of positive predictions made by the model. It measures the proportion of true positive predictions (correctly predicted positive cases) out of all positive predictions. P r e c i s i o n = T r u e P o s i t i v e s T r u e P o s i t i v e s + F a l s e P o s i t i v e s

• Recall: Recall, also known as sensitivity or true positive rate, assesses the model’s ability to identify all relevant instances in the dataset. It measures the proportion of true positive predictions out of all actual positive cases. R e c a l l = T r u e P o s i t i v e s T r u e P o s i t i v e s + F a l s e N e g a t i v e s

F1-score: The F1-score is a harmonic mean of precision and recall. It provides a balance between these two metrics and is particularly useful when you want to find a single metric that combines both precision and recall. F 1 s c o r e = 2 ∗ P r e c i s i o n ∗ R e c a l l P r e c i s i o n + R e c a l l x.

The performance of the four machine learning models in classifying tweets as hateful or non-hateful was evaluated using two different feature sets: TF-IDF and POS features in Tables 4 and 5, respectively. In Table 4, utilizing TF-IDF features, all models demonstrated relatively strong performance. XGBoost stood out with the highest accuracy of 0.86 and precision and recall, each of 0.87 and 0.86, respectively. Random Forest also performed well, with an accuracy of 0.85 and precision and recall, each of 0.87 and 0.84, respectively. In Table 5, employing POS features, the models showed slightly lower but still reasonable performance. Random Forest had the highest accuracy of 0.84, with precision, recall, and F1-score values of 0.85, 0.83, and 0.83 respectively. SVM followed closely with an accuracy of 0.82 and similar precision, recall, and F1-score values of 0.84, 0.82, and 0.82 respectively. XGBoost showed also performance with an accuracy, precision, recall, and F1-score of 0.82. These results suggest that both TF-IDF and POS features are valuable in hate speech identification, and Random Forest consistently delivered good performance (0.85, 0.84) using TF-IDF and POS features respectively. Moreover, the results indicate that decision trees have the lowest performance, primarily attributed to their poor performance in text classification, as noted in Sun et al. (2011). This observation is supported by Pranckevičius & Marcinkevičius (2017), which consistently reports decision trees’ lower accuracy compared to models like Random Forest, logistic regression, and SVM in text reviews classification tasks.

Figures 2 and 3 illustrate the confusion matrix for the Random Forest model using TF-IDF and POS features respectively. The matrix is a visual summary of the classification accuracy shows the number of correct predications and number of errors and error types.

In conclusion, our evaluation of these machine learning models using various feature sets has indeed yielded valuable insights. For instance, we found that both the XGBoost and Random Forest models, when using TF-IDF features, excelled in detecting different forms of hateful text (achieving an accuracy of 86% and 85%, respectively), making them highly suitable for hateful speech classification tasks. Conversely, the Random Forest model demonstrated the best accuracy (84%) compared to others when utilizing POS features, establishing it as a strong choice for hateful speech classification tasks. These findings provide crucial guidance for selecting the appropriate machine learning models for specific NLP tasks, ensuring the optimal model is chosen for the job.

Important features

Finding and showing the most important features in a dataset used for binary classification is crucial. It helps pick the best features and provide insights into which feature(s) affect the model’s predictions the most. Figure 4 displays the top five POS features from a pool of 263 features, utilizing the Random Forest classifier. Below, a brief overview of each color-coded feature depicted in Fig. 4:

“Definite article + PIN: Plural irregular noun”: Refers to a plural noun with the definite article and inflectional endings.

“D: Definite article + SFAJ: Singular feminine adjective”: Describes a singular feminine adjective preceded by the definite article.

“D: Definite article + SFN: Singular feminine noun”: Indicates a singular feminine noun with the definite article.

“SFN: Singular feminine noun”: Simply a singular feminine noun without a definite article.

“SMN: Singular masculine noun”: Denotes a singular masculine noun without the definite article.

These linguistic features provide essential information about the grammatical attributes of Arabic words, aiding in the analysis and comprehension of Arabic texts by revealing details like gender, number, and definiteness. For instance, the label “D: + SFAJ ” indicates a singular feminine adjective preceded by the definite article, enabling readers to identify the precise grammatical function of the word in context.

These linguistic features shed light on how hate speech towards women in sports is linguistically constructed. Plural irregular nouns with the definite article may imply collective derogatory references to female athletes. Meanwhile, the presence of definite articles with singular feminine adjectives and nouns suggests individualized attacks on specific female figures. Conversely, the absence of the definite article in singular feminine and masculine nouns may indicate a shift towards generalized or dehumanizing language targeting women or men in sports. These features offer valuable insights into the linguistic dynamics of hate speech in the sports domain.

A potential integration of qualitative findings with linguistic analysis may allow exploration of how specific linguistic features, such as the presence or absence of definite articles in plural and singular nouns, correlate with various hate speech themes directed at women in sports. For instance, the presence of plural irregular nouns with the definite article may align with gender-based derogatory comments, while the absence of the definite article in singular masculine nouns might relate to instances of general misogyny.

It’s crucial to recognize the complexity of Arabic part-of-speech (POS) tagging, arising from its rich morphology. The intricate structure, with numerous inflections, poses challenges in transferring POS features to other languages. Arabic’s effectiveness in POS tagging relies on understanding its unique morphological and syntactic features, which may not easily translate to languages with different structures. Despite these challenges, exploring the applicability of these features in other languages is important.