No abstract

Bullying is a national problem for families, courts, schools, and the economy. Social, educational, and professional lives of victims are affected. Early detection of bullies mitigates destructive effects of bullying. Our previous research found, given specific characteristics of an actor, actor logics can be developed utilizing input from natural language processing and graph analysis. Given similar characteristics of cyberbullies, in this paper, we create specific actor logics and apply these to a select social media dataset for the purpose of rapid identification of cyberbullying.

Utilization of traditional sentiment analysis for predicting the outcome of an event on a social network depends on: precise understanding of what topics relate to the event, selective elimination of trends that don't fit, and in most cases, expert knowledge of major players of the event. Sentiment analysis has traditionally taken one of two approaches to derive a quantitative value from qualitative text. These approaches include the "bag of words model", and the usage of "NLP" to attempt a real understanding of the text. Each of these methods yield very similar accuracy results with the exception of some special use cases. To do so, however, they both impose a large computational burden on the analytic system. Newer approaches have this same problem. No matter what approach is used, SA typically caps out around 80% in accuracy. However, accuracy is the result of both polarity and degree of polarity, nothing else. In this paper we present a method for hybridizing traditional SA methods to better determine shifts in opinion over time within social networks. This hybridization process involves augmenting traditional SA measurements with contextual understanding, and knowledge about writers' demographics. Our goal is to not only to improve accuracy, but to do so with minimal impact to computation requirements.