Sustainable cyberbullying detection with category-maximized relevance of harmful phrases and double-filtered automatic optimization

We developed a supporting solution for “cyberbullying” prevention based on recent discoveries in Artificial Intelligence and Natural Language Processing. Cyberbullying, defined as using the Internet to humiliate and slander other people has become a serious problem. In Japan members of the Parent-Teacher Association manually perform Web monitoring to stop cyberbullying activities. Unfortunately, reading through the whole Web manually is an impossible task. Although the complexity of cyberbullying makes it a problem unsolvable solely with the help of technology, we found that technology could make cyberbullying prevention more efficient. We developed a novel method of automatic detection of cyberbullying entries on the Internet. In the method we use seed words from three categories to calculate a semantic orientation score and then maximize the relevance of categories. The proposed method outperformed baseline settings in both laboratory and real world conditions. The developed system was deployed and tested in practice. After a year of testing we noticed a greater than 30 percent-point-drop in its performance. We hypothesize on the reasons for the drop. To regain the lost performance and retain it in the future we propose additional improvements including automatic acquisition and filtering of seed words. Experimentally selected optimal improvements regained much of the lost performance.