Brute Force Works Best Against Bullying

The problem of harmful and offending messages on the Internet has existed for many years. One of the reasons such activities evolved was the anonymity of communication on the Internet, giving users the feeling that anything can go unpunished. Recently the problem has been officially defined and labeled as cyberbullying (CB)1. In Japan the problem has become serious enough to be noticed by the Ministry of Education [MEXT 2008]. In 2007 Japanese school personnel and members of Parent-Teacher Association (PTA) have started monitoring activities under the general name Internet Patrol (later: net-patrol) to spot Web sites containing such inappropriate contents. However, the net-patrol is performed manually as a volunteer work. Countless amounts of data on the Internet make this an uphill task. This situation motivated us to help and ease the burden of the net-patrol members and create a net-patrol crawler automatically spotting cyberbullying entries on the Web and reporting them to appropriate organs. In the following sections we firstly present some of the previous research related to ours. Next, we describe our method and the dataset used in this research. In the method we apply a combinatorial approach to language modeling, resembling brute force search algorithms, to extract sophisticated patterns from sentences. Next, we use them in text classification task. Finally, we explain the evaluation settings, analyze and discuss the results. Evaluation on actual cyberbullying data showed our method outperformed previous ones while minimizing human effort.