Automatic detection of cyberbullying behaviour on social media using Stacked Bi-Gru attention with BERT model

Cyberbullying behaviour has drawn more attention as social media usage has grown. Teen suicide has been related to cyberbullying, among other serious and harmful effects on a person’s life. Using the appropriate natural language processing and machine learning techniques, it is possible to proactively identify bullying content to reduce and eventually eradicate cyberbullying. Accordingly, the article proposed an automated deep-learning model for detecting aggressive activity in cyberbullying. Initially, the data was extracted from the social media platform using Formspring, Instagram and MySpace datasets for perceiving cyberbullying behaviour, then the collected data are input for preprocessing. To remove the raw data, several preprocessing processes have been introduced. They consist of removing stop words, white spaces for punctuation, and changing the comments to lowercase. Lexical Density (LD) has been one of the metrics used to gauge language complexity generally. As a result, the study made use of the Feature Density (FD) to calculate how complicated certain natural language datasets are using the linguistically backed preprocessing model. After preprocessing, the data are input to the feature selection process which selects the pertinent features or attributes to include in predictive modelling and which to leave out. Since, the article proposed a Binary Chimp Optimization (BCO)-based Feature Selection (BCO-FSS) technique, which selects the subset of features for classification performance improvement. The selected features are exploited for cyberbullying behaviour detection. To identify the exploit of social media for cyberbullying text content, the article suggested Stacked Bidirectional Gated Recurrent Unit (SBiGRU) Attention for learning spatial location information and sequential semantic representations using a Bi-GRU. Additionally, the BERT model is employed as a base classifier to recognize and categorise aggressive behaviour in the textual content. The Matlab software is employed for simulation. For accuracy, precision, recall, and F1-Score, this experiment yielded a practically perfect outcome with values of 99.12%, 94.73%, 97.45%, and 93.91% respectively.