Social networks such as Facebook or Twitter promote the communication between people but they also lead to some excessive uses on the Internet such as cyberbullying for malicious users. In addition, the accessibility of the social network also allows cyberbullying to occur at anytime and evoke more harm from other users’ dissemination. This study collects cyberbullying cases in Twitter and attempts to establish an auto-detection model of cyberbullying tweets base on the text, readability, sentiment score, and other user information to predict the tweets with harassment and ridicule cyberbullying tweets. The novelty of this study is using the readability analysis that has not been considered in past studies to reflect the author's education level, age, and social status. Three data mining techniques, k-nearest neighbors, support vector machine, and decision tree are used in this study to detect the cyberbullying tweets and select the best performance model for cyberbullying prediction.