Recognition of Toxicity of Reviews in Online Discussions

Abstract

The paper solves some problems belonging to the field of recognition of asocial behaviour in online space. Nowadays, it seems to be an important issue when we must question our way of dealing with the pandemic crisis. Social network users must deal with such unhealthy phenomena in online space as toxic comments and toxic troll authors that prevent constructive communication and knowledge sharing through the web space. We have proposed a new multimodal approach to social network analysis, which combines two methods, the first one to recognize toxic posts using machine learning and the second one to identify toxic authors in online space using sentiment analysis. The recurrent neural network was trained with different numbers of neurons in the hidden layers using three different types of hidden layers and optimizers along with various learning rates. Finally, the paper provides detailed results of extended experiments with deep learning models for recognition of toxic reviews, where a model generated by a combined deep learning architecture achieved accuracy over 0.9, and results of our novel approach to the detection of toxic troll reviewers achieving accuracy of 0.95. Our approach to troll recognition is based on a comparison of the sentiment related to the authors’ posts to sentiment related to all comments of an online discussion.