Online Aggression Identification Using Ensembled Transformer-based Language Models

The ubiquitous online aggressive content is rapidly increasing with the increase in the popularity of Social media platforms. Social media has been used by a few to vent out and demean an individual or community and can disturb society’s social harmony. Therefore, restraining aggressive online content has become an urgent need. Past studies on aggressive behaviour identification have done an excellent job, but an absolute combat method is yet to be designed. In this paper, an XGBoost ensembling of large fine-tuned Transformer-based Language Models viz., XLNet, RoBERTa, ALBERT, and BERT are proposed to detect Online Aggression. The performance of the proposed approach is evaluated on various metrics such as Accuracy, Micro, Macro, and Weighted aggregation of Precision, Recall, and F1-Score. The proposed approach gave better results (like 13.18% and 7.47% increase in performance concerning Weighted F1-Score and Accuracy, respectively) compared to varied individual and ensemble Transformer-based and Deep Learning models.