Summary:
This thesis concerns the application and comparison of machine learning algorithms for sentiment analysis in order to detect cyberbullying. The algorithms are applied to two different datasets: SOSNet Twitter Dataset and Suspicious Tweets Dataset.
The purpose of the work was, in addition to simple detection of online bullying, to further find the type of bullying according to specific criteria such as age, gender, nationality, etc. Furthermore, the linguistic elements of the texts of each category are presented, as well as the results of other researches regarding the incidence of cyberbullying according to the personal characteristics of the person under study. As extensions of the present study, the creation of a system which will take into account personal axes/criteria such as gender, nationality, sexual orientation etc. for the detection of online bullying is proposed. In addition, the linguistic features are collected so that the research can be adapted to Greek data as well. The first effort for this extension takes place in the present study.
Finally, all the methodologies that have been implemented in similar research are described in detail, as well as the one preferred in the current one. The results of all algorithms on each data set are listed and commented extensively. The detection of cyberbullying and its correct categorization are done with high accuracy. However, some minor failures are pointed out and a future goal is to create a neural network to potentially improve these failures.
