Abstract
With the current increase in the number of online users, there has been a concomitant increase in the amount of data shared online. Techniques for discovering knowledge from these data can provide us with valuable information when it comes to detecting different problems, including violence. Violence is one of the significant problems humanity has faced in recent years all over the world, and this is especially a problem in Arabic countries. To address this issue, this research focuses on detecting violence-related tweets to help in solving this problem. Text mining is an important technique that can be used to find and predict information from text. In this study, a text classification model is built for detecting violence in Arabic dialects on Twitter using different feature-reduction approaches. The experiment comprises bagging, K-nearest neighbors (KNN), and Bayesian boosting using different extraction features, namely, root-based stemming, light stemming, and n-grams. In addition, the study used the following feature-reduction techniques: support vector machine (SVM), Chi-squared (CHI), the Gini index, correlation, rules, information gain (IG), deviation, symmetrical uncertainty, and the IG ratio. The experiment showed that the bagging with tri-gram approach has the highest accuracy at 86.61%, and a combination of IG with SVM from reduction features registers an accuracy of 90.59%.