Architecture of Cyberbullying Recognizer in Video Game Chat Using Deep Learning Model with BiLSTM

Abstract

Cyberbullying is a social problem that has been growing in recent years through social media and other contexts such as videogames which is the most visited context by children and adolescents in these times. The solutions proposed in recent years to address the prediction of texts with cyberbullying used machine learning algorithms of classification. In view of this, we propose an architecture as a solution, using Natural Language Processing algorithms and a Deep Learning model with BiLSTM for the classification of texts in video game chats. The architecture is composed of 3 modules: Pre-processing, NLP and Clustering. The proposed architecture is sequential, the first Pre-processing module is in charge of cleaning the dataset of numbers, punctuation marks, Stop Words; the second NLP Module is in charge of generating a feature vector and a vocabulary with the Word2Vec algorithm; the third Module is in charge of classifying our dataset if it is cyberbullying thanks to the unsupervised Kmeans and TF-IDF algorithm; finally the Training Module is in charge of training our BiLSTM model which is composed of an Embedding layer, a BiLSTM layer, we show as a result the architecture test obtaining 97.91% accuracy.