Experts and Machines against Bullies: A Hybrid Approach to Detect Cyberbullies

Abstract

Abstract. Cyberbullying is becoming a major concern in online environments with troubling consequences. However, most of the technical studies have focused on the detection of cyberbullying through identifying harassing comments rather than preventing the incidents by detecting the bullies. In this work we study the automatic detection of bully users on YouTube. We compare three types of automatic detection: an expert system, supervised machine learning models, and a hybrid type combining the two. All these systems assign a score indicating the level of "bulliness" of online bullies. We demonstrate that the expert system outperforms the machine learning models. The hybrid classifier shows an even better performance. IntroductionWith the growth of the use of Internet as a social medium, a new form of bullying has emerged, called cyberbullying. Cyberbullying is defined as an aggressive, intentional act carried out by a group or individual, using electronic forms of contact repeatedly and over time against a victim who cannot easily defend him or herself [1]. One of the most common forms is the posting of hateful comments about someone in social networks. Many social studies have been conducted to provide support and training for adults and teenagers [2,3]. The majority of the existing technical studies on cyberbullying have concentrated on the detection of bullying or harassing comments [4-6], while there is hardly work on the more challenging task of detecting cyberbullies and studies for this area of research are largely missing. There are few exceptions however, that point out an interesting direction for the incorporation of user information in detecting offensive contents, but more advanced user information or personal characteristics such as writing style or possible network activities has not been included in these studies [7,8]. Cyberbullying prevention based on user profiles was addressed for the first time in our latest study in which an expert system was developed that assigns scores to social network users to indicate their level of 'bulliness' and their potential for future misbehaviour based on the history of their activities [9]. In the previous work we did not investigate machine learning models. In this study we focus again on the detection of bully users in online social networks but now we look into the efficiency of both expert systems and machine learning models for identifying the potential bully users. We compare the performance of both systems for the task of assigning a score to social network users that indicates their level of bulliness. We demonstrate that the expert system outperforms the machine learner and can be effectively combined in a hybrid classifier. The approach we propose can be used for building monitoring tools to stop potential bullies from conducting further harm. Data Collection and Feature SelectionIn this section we will explain the characteristics of the corpus used in this study. We also describe the feature space and the three feature categories that have been used...

Harnessing the Power of Text Mining for the Detection of Abusive Content in Social Media

“…A barrier for the use of text mining techniques for abusive content detection is the lack of labelled datasets in the field. At present, researchers collect data, and annotate by one of two approaches -their own labelling effort [30,9,1,19] which is timeconsuming or through the use of crowdsourcing services [25,2] such as Amazon's Mechanical Turk which can be costly.…”

“…Other research takes extra information into account to enhance classification accuracy such as user profile features [9,8] including age and gender; semantic features of the user comment [30,9,1,32] such as parts of speech, number of pronouns; and features such as profanity word occurrences [1,32,4]. Text context features have also been analysed in recent years.…”

“…Wellknown classifiers have been used in this domain including Support Vector Machines (SVM) [9,16], Näive Bayes [4,13], logistic regression [29], and decision trees [25,16]. Classifier ensemble solutions have also proven successful.…”

“…Dadvar et al [10] detected cyberbullies instead of detecting text content; Nahar et al [21] used social networks to present a graph model, identifying the most active cyberbullying predators and victims; Xu et al [31] explored the detection of roles within cyberbullying, and identified those of bully, victim, accuser and reporter. In addition, research [30,13,29] used Latent Dirichlet Allocation (LDA) to extract the main topics for each text content.…”

“…D2-YouTube Dadvar et al [9] created a corpus of comments from the video upload site, YouTube, by scraping comments from sensitive cyberbullying topics within the site. Each labelled instance consists of a single user's comment.…”