Detecting Insults on Social Network Platforms Using a Deep Learning Transformer-Based Model

Abstract

Social network platforms ought to be a place where no user gets insulted via verbal abuse. Cyberbullying and online abusive language is becoming a major concern for all stakeholders. Research on insult detection has seen a surge in interest over the last few years to address this problem. The emergence of deep learning techniques and their promising achievements in various fields have promoted their use as a natural alternative to tackle the many challenges related to insult detection. The authors describe in this chapter a supervised deep learning model to capture the online comments' semantics and detect insults. The proposed approach relies on two main components: (i) Text pre-processing and representation, and (ii) A Transformer-based deep learning model trained to classify the comments submitted on social media platforms as an insult or insult free based on the BERT method. The obtained results indicate that the proposed model outperforms CNNs and RNNs in the detection of insults for the same OLID benchmark yielding a macro F1-score of 0.83 and an accuracy of 86%.