Human Behavior Assessment using Ensemble Models

Automated Detection of Cyberbullying Against Women and Immigrants and Cross-domain Adaptability Thushari Atapattu1, Mahen Herath2, Georgia Zhang1 and Katrina Falkner1 1School of Computer Science, The University of Adelaide, Adelaide, Australia 2Department of Computer Science & Engineering, University of Moratuwa, Katubedda, Sri Lanka email: thushari.atapattu@adelaide.edu.au Abstract Cyberbullying is a prevalent and growing so- cial problem due to the surge of social me- dia technology usage. Minorities, women, and adolescents are among the common victims of cyberbullying. Despite the advancement of NLP technologies, the automated cyberbully- ing detection remains challenging. This pa- per focuses on advancing the technology us- ing state-of-the-art NLP techniques. We use a Twitter dataset from SemEval 2019 - Task 5 (HatEval) on hate speech against women and immigrants. Our best performing ensemble model based on DistilBERT has achieved 0.73 and 0.74 of F1 score in the task of classifying hate speech (Task A) and aggressiveness and target (Task B) respectively. We adapt the en- semble model developed for Task A to classify offensive language in external datasets and achieved ∼0.7 of F1 score using three bench- mark datasets, enabling promising results for cross-domain adaptability. We conduct a qual- itative analysis of misclassified tweets to pro- vide insightful recommendations for future cy- berbullying research. 1 Introduction and Motivation Cyberbullying is ”the repetitive use of aggressive language amongst peers with the intention to harm others through digital media” (Rosa et al., 2019). Due to the surge of social media technology use, cyberbullying has become a prevalent and growing social problem. Unlike in the physical environment, cyberspace, in particular, online social platforms are not yet evolved sufficiently to prevent people from communicating without disclosing identities, spreading rumours, and harassing others. The risk of and potential consequences caused by cyberbul- lying are critical including both physical and men- tal health risk to victims. The impact and conse- quences are common to all generations (e.g. young, elderly) including emotional and psychological dis- tress, decline in personal/academic development, anti-social behaviour, and, potentially, suicide. The criticality of this societal problem is demon- strated from a study by Yale University, comment- ing “cyberbullying victims are 2 to 9 times more likely to consider committing suicide” across the globe.1 Within Australia, the eSafety Commis- sioner comments “one in every five Australian chil- dren aged eight to seventeen are victims of cyber- bullying (2018)”.2 Adolescents, minorities (e.g. refugees, LGBTQI) and women are among com- mon targets of cyberbullying. According to Bully- ing Statistics3, over half of adolescents are victims of cyberbullying and about the same percentage are involved in bullying. Despite recent research advancement in hate speech detection (Fortuna and Nunes, 2018), au- tomated identification of cyberbullying attempts (i.e. repetitive hate speech against an individual or a group) remains as a challenging subtask of NLP. Due to diverse variants of language (e.g. hate, intimidation, sarcasm, metaphors) used by bullies and the evolution of language (e.g. slang), particu- larly among adolescents, the automated detection of cyberbullying is extremely challenging. The example below appears to be misogynistic as it in- …

Content truncated to 500 words