Text Mining Techniques for Cyberbullying Detection: State of the Art

Abstract

The dramatic growth of social media during the last years has been associated with the emergence of a new bullying types. Platforms such as Facebook, Twitter, YouTube, and others are now privileged ways to disseminate all kinds of information. Indeed, communicating through social media without revealing the real identity has emerged an ideal atmosphere for cyberbullying, where people can pour out their hatred. Therefore, become very urgent to find automated methods to detect cyberbullying through text mining techniques. So, many researchers have recently investigated various approaches, and the number of scientific studies about this topic is growing very rapidly. Nonetheless, the methods are used to classify the phenomenon and evaluation methods are still under discussion. Subsequently, comparing the results between the studies and identifying their performance is still difficult. Therefore, the current systematic review has been conducted with the aim of survey the researches and studies that have been conducted so far by the research community in the topic of cyberbullying classification based on text language. In order to direct future studies on the topic to a more consistent and compatible perspective on recent works, we undertook a deep review of evaluation methods, features, dataset size, language, and dataset source of the latest research in this field. We made a choice to focus more on techniques that adopted neural networks and machine learning algorithms. After conducting systematic searches and applying the inclusion criteria, 16 different studies were included. It was found that the best accuracy was achieved when a deep learning approach is used particularly CNN approach. It was found also that, SVM is the most common classifier in both Arabic and Latin languages and outperformed the other classifiers. Also, the most widely used feature is N-Gram especially bigram and trigram. Furthermore, results show that Twitter is the main source for the collected datasets, and there are no unified datasets. There is also a shortage of studies in Arabic texts for cyberbullying identification in contrast with English texts.

Instagram-Based Benchmark Dataset for Cyberbullying Detection in Arabic Text

“…Thus, the need for an existing Arabic dataset to evaluate automatic algorithms for detecting and classifying offensive speech is critical. Indeed, the majority of the available cyberbullying datasets are in English [19][20][21], while there are only a few available in Arabic. Hence, in this section, we provide an overview of the most recent Arabic datasets that addressed cyberbullying, abusive, and offensive language.…”

“…Therefore, in our work, we introduce a corpus with (sub-class categorization (multi-class)) that could help in studying cognitive processes, such as investigating whether cyberbullying is linked with increased levels of anxiety in online audiences. .However, the majority of the available cyberbullying datasets are in English [19,20], with only a few available in Arabic, none of which are data collected from the Instagram platform [18,21]; the number of Arabic datasets available for offensive/hate speech/cyberbullying auto detection is limited compared to in the English language. Furthermore, none of them are collected from Instagram.…”

Sentiment Analysis of Emirati Dialect

“…Social networking has become one of the most popular communication platforms nowadays. People of different ages, cultures, and socioeconomic classes utilize social network environments to communicate a variety of messages to a worldwide audience [1,2]. Twitter, Instagram, and Facebook are examples of social networking platforms that allow users to communicate and freely discuss their thoughts and opinions in a non-constrictive atmosphere.…”

“…It is worth noting that SPSS software restricted the following rules for correctly implementing Cohen's kappa coefficient: (1) The raters' responses were graded on a nominal scale, and the categories must be mutually exclusive (in our experiment: positive, negative, or neutral). (2) The response data were made up of paired observations of the same phenomena, which meant that both raters evaluated the identical observations (in our experiment, annotators evaluated the same comments, i.e., text). (3) The two raters were independent, which meant that one rater's decision did not influence the decision of the other (in our experiment, the annotators were independent).…”

A Systematic Review for Sentiment Analysis of Arabic Dialect Texts Researches

“…Social Media has become a part of daily life; It is increasingly hard to survive in this modern era of digital media without a digital footprint [1]. People are increasingly using social media platforms for various purposes and reasons, resulting in a vast amount of online data being created on a daily basis [2].…”