Introduction

Cyberbullying has recently been recognized as a national health concern by social media users, and creating a detection model in the modern period has been shown to be scientifically beneficial [1]. On social media, users publish a variety of content, including documents, images, and videos, and engage in online communication. People connect with social media through cellphones or computers. The most popular social media platforms are Facebook, Twitter, Instagram, TikTok, and others. Social media is now utilised for a wide range of objectives, including education and business [2]. Social media is causing a lot of new jobs to be created, which is good for the world economy [1]. Social networking has a lot of benefits, but it also has some drawbacks. In order to hurt the reputations and sentiments of others, malicious users of this medium engage in dishonest and deceptive behavior. Recently, cyberbullying has become one of the most serious phenomena in internet media [3]. Cyberbullying and cyber-harassment are terms used to describe bullying or harassment that occurs online. Cyberbullying and cyberharassment are both forms of online bullying. Due to the widespread adoption of digital and technological advancements in the modern world, cyberbullying has become more common among adolescents [4].

Online social networks (OSNs) play a significant role in facilitating social connection, but they also foster antisocial behavior like trolling, cyberbullying, and hate speech. Cyberbullying is the practise of expressing hostile or hateful remarks through the use of short message services or media platforms found on the Internet. Natural language processing (NLP) for automatic detection is hence the first step towards stopping cyberbullying [5].

Cyberbullying in Social Media Platforms

Most people are aware that cyberbullying occurs when someone uses a variety of communication channels, including text messages, emails, social media, and other online platforms, to publish or send hurtful words or comments, including images. Young children and teenagers are commonly the targets of cyberbullying because they are more open to new technology, such as the Internet [5]. A sort of harassment in which one person insults or offends another is described as cyberbullying [6]. Cyberbullying is the willful, persistent, and hostile use of technology to harass or damage another person.

Unlike more conventional types of bullying, cyberbullying can happen every day of the week, at any hour of the day or night. It occurs in a variety of ways, including sending text messages, spreading rumors, and posting embarrassing videos and images on social media [7]. Cyberbullying is different from traditional bullying in that it occurs online, where the message will either be delivered to the victim directly or posted in open forums where anyone can view it. Although these cyberbullies can be anyone, they frequently know their victims [8]. They may be a friend or classmate at times. Modern communication channels offer a wide range of cyberbullying techniques. The following forms of cyberbullying, for instance, have been documented [9,10,11,12].

Cyberbullying Types

  1. 1.

    Flaming: starting a battle online.

  2. 2.

    Harassment: the victim of harassment receives nasty and insulting communications on a regular basis this is the type of cyberbullying that most publications are seeking to stop.

  3. 3.

    Cyberstalking: the victim receives offensive or intimidating messages that make them feel endangered.

  4. 4.

    Masquerade: the bully presents a false persona.

  5. 5.

    Trolling: making offensive comments on social media with the intent of offending other users.

  6. 6.

    Discredit: spreading unfavorable rumors about someone else and divulging private information about them in public.

  7. 7.

    Absence: the Phenomena of excluding someone from a group of individuals

  8. 8.

    Petty crime: the act of taking someone’s identity online and fabricating a false profile to deceive others.

  9. 9.

    Dissing: the phenomena in which information or opinions about another person are spread with the intention of harming that person’s reputation or popularity.

  10. 10.

    Trickery: trickery is the act of having someone trust you with their secrets and personal information, only for them to utilize that confidence to disclose that information online to the public.

  11. 11.

    Frapping: frapping is the practise of someone using your social media accounts to pretend to be you, publish messages under your name, and lead people astray.

The Impact of Cyberbullying

The pandemic’s impacts have led to a 40% spike in cyberbullying attacks.

Youth suicides are on the rise, and 20% of that increase is due to cyberbullying. Attacks involving cyberbullying are expected to reach an all-time high in 2025, according to 60% of experts (Source: Statista.com). Cyberbullying affects more than 38% of persons who use online media platforms every day. In response to the cyberbullying attack, about 25% of kids self-harm as a coping mechanism for the humiliation. If they experience cyberbullying, adolescents between the ages of 12 and 18 are more likely to experience social and health problems in the future in the Journal of Adolescence. Cyberbullying is spreading alarmingly despite the fact that many individuals are aware of cyberattacks. 64 percent of victims who get hostile instant messages claim to have personally interacted with the sender. Online teenagers reported receiving inappropriate forwarding of private messages in nearly one in six cases (15%) (Source: Pew Research Centre). The cyberbullying incidences in India are depicted in Fig. 1. In the year 2019, there were 542 reported cases of cyberstalking, and in the following year, there was a sharp rise to 739 cases, for a total of 1000 cases. The percentage of parents whose children have experienced bullying is shown in Fig. 2 by age. In India, children between the ages of 14 and 18 encounter the largest percentage of cyberbullying incidents—59.9%. The suicide and homicide rates are shown in Fig. 3 as a result of cyberbullying.

Fig. 1
figure 1

Cyberbullying incidents in India

Full size image
Fig. 2
figure 2

Cyberbullying percentage by age

Full size image
Fig. 3
figure 3

Suicide and homicide death rates in the USA

Full size image

Background and Related Work

An effective solution for spotting online abusive and bullying texts is created by combining natural language processing with machine learning techniques. Using distinct properties including phrase frequency, inverse text frequency, and bag-of-words, the accuracy of distinct is tested and used for categorising comments as bullying or non-bullying [1].The study employed Twitter datasets to attempt to suggest an ensemble model by analysing the classification methods used to pinpoint instances of cyberbullying. The suggested ensemble model combines the models of various machine learning classifiers and each of the individual component classifiers on which the datasets are trained to create the predictions in an effort to outperform the individual models. By adding these preliminary forecasts, a final conclusion is reached. Numerous strategies, such as bagging, boosting, alternative stacking, and voting, can be used to make this final forecast. The assessment methods include Adaptive Boosting, Naive Bayes, K nearest neighbours, decision tree, logistic regression, random forest, Bagging classifiers, linear support vector classifier, and stochastic gradient descent [13].

Image-based cyberbullying is recognised using optical character recognition (OCR) technology, and its component effects are assessed on a dummy system. Cyberbullying is automatically detected using natural language processing and machine learning algorithms, and textual data is matched for the corresponding features of the transaction. We developed a semi-supervised method for detecting cyberbullying using the BERT model, which uses the five features that can be utilised to characterise a cyberbullying post or message. The BERT model surpassed traditional machine learning models with an accuracy of 91.90 percent after being trained over two cycles, accounting exclusively for emotive characteristics. The BERT model can generate more accurate findings when a large amount of data is employed [14].

The methods of cyberbullying are investigated using a consolidated deep learning model to automatically detect aggressive behaviour. This method uses three multichannel deep learning models, including convolutional neural networks, bidirectional gated recurrent units, and transformer blocks, to categorise Twitter comments into two groups. Combining data from three well-known datasets on hate speech allowed researchers to assess the effectiveness of the suggested approach. The suggested strategy delivered positive results. The accuracy of the suggested strategy is about 88 percent [5].

Each bully tweet on Twitter has a value of 1, indicating that they have all been positively recognised, according to the results of the machine learning techniques used to detect cyberbullying. Different machine learning models receive an equal amount of tweets from the Twitter dataset that are bullies and non-bullies. With a precision of 91%, recall of 94%, and F1-score of 93%, the logistic regression classifier correctly distinguishes between bullying and non-bullying in tweets. Cyberbullying will not hurt users, thanks to users’ ability to prevent it [6].

Different machine learning algorithms for cyberbullying detection are covered, as well as the many cyberbullying categories, data sources, and sources of cyberbullying data for research. The dearth of publicly accessible statistics and the absence of multimedia content-based detection were cited as barriers to cyberbullying detection [8].

Very few people have tested methods that do not entail supervision, thus supervised learning techniques were used to detect cyberbullying. This gives researchers more space to work and gives harassers more room to control their targets. Unattended methods are getting more attention, but supervised methods have historically dominated the detection of cyberbullying. It is also feasible to claim that supervised methods fail to address class disparity, whereas unsupervised ones can. The paper focuses on recent investigations of non-supervised text-based cyberbullying detection and makes recommendations for further research. It calls for more research on unsupervised approaches and emphasises the seriousness and intensity of harassment in the online environment [15].

The use of machine learning techniques has been made to lessen or stop cyberbullying. These attempts, however, work because they rely on the interactions between the victims. Therefore, it’s essential to identify cyberbullying without the victims’ participation. In this work, we attempted to analyse this problem using a global dataset of 37,373 unique tweets from Twitter. The seven machine learning classifiers used machine learning techniques such, logistic regression, support vector machine, Ada boost, light gradient boosting machine, stochastic gradient descent, random forest, and naive Bayes. performance of all the aforementioned algorithms was assessed based on various metrics, including F1score, precision, accuracy, and recall, to estimate the global dataset’s recognition rates. With an accuracy of about 90.57 percent, the experiment’s findings show that Logistic Regression is preferable. The greatest F1 score for logistic was 0.928, the highest recall was 1.0, while the highest precision for stochastic gradient descent was 0.968 [16].

The sentiment analysis of the Twitter data is carried out using ordinal regression. Twitter samples from the Corpus of NLTK, including both good and negative tweets, are used in our research. This improvement was achieved by using Text Blob to determine the polarity of the tweets and lemmatization in place of stemming. To summarise our work, we performed sentiment analysis on Twitter data. The sensation is one of the five categories. Positive and negative tweets from the NLTK Corpus Twitter samples were mined for information. Only the pertinent information was preprocessed for the data extraction using lemmatization. We identified the polarity of the tweets using the Text Blob library. There were five groups created for them. On top of that, three machine learning (ML) techniques were used on the information. SVM classifier outperformed the Multinomial Logistic Regression classifier and Random Forest classifier in Twitter Sentiment Analysis with an accuracy of 86.60%. The accuracy of our Support Vector Classifier was enhanced by using GridSearchCV to choose the appropriate hyperparameters while applying SVM. We discovered that at “C”:1000 and “gamma,” 0.001 was the optimal value [17].

Utilising both user and textual features, the Twitter data was utilised to study abusive posting behaviour from a number of angles. The number of highly interrelated abusive behaviour standards was identified using a deep learning architecture; the suggested work makes use of easily available metadata to combine it with hidden patterns that are automatically extracted from tweeted text. With no need for model architecture modification for each activity, we can detect different types of abusive behaviour using this unified architecture in a seamless and transparent manner. We assess the suggested approach using various datasets that include a range of abusive Twitter behaviours. The findings indicate that it greatly outperforms the other techniques based on the dataset (an improvement in AUC of between 21 and 45%). Training a multi-input network is challenging. A technique that uses two input routes and alternates training between them was developed to further enhance performance across all evaluated datasets. We showed that the suggested training paradigm can perform noticeably better than a variety of other possible training techniques, such as ensemble, feature transfer, and concurrent training [18].

Using a congruent attention fusion capsule network, cyberbullying on social media was discovered. Due to the homogeneity of bullying types, dynamic routing between capsules can more adaptively identify the fine-grained categories in the cyberbullying text to further improve detection performance. An extensible congruent attention mechanism that strikes a balance between the fusions of detailed correlations between various feature subspaces was developed using a unified spatial representation of the composition. To automatically improve bullying characteristics and optimise the word embedding matrix, a similarity weighting approach based on word2vec that assesses the similarity between context words and external bullying vocabulary has also been investigated [19].

The goal of the study was to identify cyberbullying actors using text analysis and user credibility to educate users about the harm that cyberbullying causes. Information was taken from Twitter. Because the data were unlabeled, we created a web-based labelling tool to separate tweets that contained cyberbullying from tweets that did not. The technology provided us with 2053 derogatory words, 129 tweets containing swear words, 301 tweets about cyberbullying, 399 tweets with non-cyberbullying. Then, using SVM and KNN, we discovered and identified cyberbullying texts. The results show that SVM generates the greatest F1-score, which is 67% [20]. People occasionally use sarcasm to convey their feelings and thoughts, indicating the complete opposite of what they say. Sarcastic content can be shared by users in a number of different ways, including audio, photos, podcasts, text, and videos. The study uses text data taken from Twitter to examine COVID-19 ironic material with negative attitudes. We extracted the data using specific terms such hash tag-related, irony, sarcastic information, and sarcasm, and then we did an aggressive and offensive character study of people. Decision Tree, Naive Bayes, and the linear support vector classifier were used to conduct this research. The decision tree demonstrated the highest level of accuracy compared to libSVM and Naive Bayes with 90% accuracy (Table 1).

Table 1 Comparative summary of different approaches in cyberbullying detection
Full size table

Cyberbullying Detection Algorithms

The two main parts of the cyberbullying detection techniques are discussed.

  • Part 1: using natural language processing to detect cyberbullying.

  • Part 2: machine learning.

Cyberbullying Detection in Natural Language Processing

One area of research in this area is the detection of offensive information using natural language processing (NLP). The most effective method for describing how Natural Language Processing works is “LANGUAGE LEVELS” [23]. These levels are used by people to interpret spoken or written languages. Processing applications make use of linguistic abilities [23, 24].

Machine Learning in Cyberbullying Detection

The datasets are used with any of the following machine learning algorithms to find bullying-related messages on social media.

Decision Tree

Decision tree classifiers can be used for both classification and regression [1]. It can help in decision-making and representation. It is a framework that resembles a tree, with each leaf node standing in for a choice and each inside node for a condition. A classification tree returns the class of the target. A regression tree will show the anticipated value for an addressed input.

Naive Bayes

Naive Bayes is a powerful machine learning method that is based on the Bayes theorem [25]. The probability of an object is used by the algorithm to create predictions. Multi-class classification and binary problems can be swiftly fixed with this technique.

Support Vector Machine

Similar to a decision tree, the support vector machine (SVM) is a supervised machine learning technique that may be applied to both classification and regression. It can distinguish the classes in n-dimensional space in a special way. In practice, SVM creates a series of hyperplanes in an infinite-dimensional space using a kernel that turns an input data space into the desired form. For instance, Linear Kernel employs instances dot products in the following manner: K(x, xi) = sum (x, xi).

Evaluation Metrics

Researchers assess how well a suggested model separates cyberbullying from non-cyberbullying using a range of evaluation metrics. Examining typical evaluation criteria employed by academics is crucial to understanding the efficacy of models. The most popular measures for evaluating cyberbullying classifiers on social media websites are as follows:

Accuracy

Enhanced accuracy in identifying cyberbullying content and determining the victim’s emotional state following a cyberbullying incident via real-time streaming data

$$\mathrm{Accuracy}=\frac{(\mathrm{TP}+\mathrm{TN})}{(\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN})}$$

The prediction models for cyberbullying could be assessed using.

Precision: determines the percentage of successfully or accurately identified samples

$$\mathrm{Precision}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$$

Recall: provides the percentage of real positives.

$$\mathrm{Recall}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$

F1 Measure: provides the choral group with memory and accuracy

$$F1-\mathrm{Measure}=\frac{-2*\mathrm{Precision}*\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}$$

TP stands for true positive, TN for true negative, FP for false positive, and FN for false negative, respectively.

Data for the performance measure evaluation metrics for social media websites of cyberbullying classifiers are shown in Table 2. The data demonstrates unequivocally that the stochastic gradient descent (SGD) classifier has the highest accuracy of 90.60% [16].

Table 2 Algorithms performance summary based on evaluation metrics
Full size table

Conclusions and Future Directions

Social networks have mostly taken over our everyday routines because using them makes it so simple to engage with others. However, the development of antisocial behaviour like trolling, hate speech, and cyberbullying on social networks like Twitter and the negative effects that social media users experience make this a crucial topic to research [5]. An effective strategy for addressing the cyberbullying issue is the use of deep learning techniques to identify the social media content that encourages it. The examination of research conducted revealed that there has been very little research on the identification of cyberbullying, and that detection using popular multimedia content, such as videos, music, etc., has been overdone using text-based research. The use of ML approaches is restricted by the availability of data because a testing and training dataset is necessary. In addition, the majority of the available datasets only contain text data, therefore the researchers had to create their own data. The concept of implementing a real-time cyberbullying detection system, which will be useful for identifying and stopping cyberbullying instantly, can be another area of research, and working with various different languages can open up research avenues [26]. Deep Learning techniques can be used to classify cyberbullying in a more precise manner, and they will perform better than the machine learning algorithms currently in use.