SocialBully: A Social Information-Driven Cyberbullying Detector with Similarity-Based Word Embedding

Abstract

Cyberbullying depicts the form of bullying that is spread through information and communication technologies, most common of which is through the Internet. When compared to traditional bullying, cyberbullying can spread quicker and reach a wider audience and, because of this, can have severe effects on people’s mental health. In this paper, we propose a new cyberbullying detection method – SocialBully. SocialBully systematically integrates both users’ social information and their posted textual information to achieve a high detection accuracy. It uses a new word embedding method “SimWord” to represent words based on the similarity of their co-occurrence vectors and exploits a graph embedding method to obtain the social representation of users. Based on these representations, SocialBully uses a bidirectional Long Short-Term Memory to detect bullying texts. Extensive experiments are conducted on four real-world datasets to validate the effectiveness of the proposed method.