Detect Chinese Cyber Bullying by Analyzing User Behaviors and Language Patterns

Abstract

With the rapid growth of social media, people are more aware of cyber bullying on the internet. The most important aspect for preventing cyber bullying is to detect the abusive content. In this paper, we build a Long Short-Term MemoryNeural Network-Deterministic Finite Automaton (LND) model which considers not only the language content, but also the user’s characteristics and historical speech on social network. Due to the lack of labeled content, we utilize Douban’s reviewers data by analyzing speech patterns with polarized emotions. Then the learned model is applied to analyze Chinese cyber bully behaviors on Weibo. As a result, the accuracy of detecting cyber bullying increases from 89% (sensitive lexicon filtering method) to 95% by considering user’s behavior features and language emotional polarity scores. Our model is capable of analyze real celebrities’ Weibo webpages and assists prevention of cyber bullying on social media.