CRIMES DE ÓDIO NA INTERNET: UM COMPARATIVO DE MODELOS DE APRENDIZADO DE MÁQUINA CLÁSSICOS PARA IDENTIFICAÇÃO DE DISCURSO NOCIVO

Abstract

In the contemporary world, technology plays a fundamental role in life in society, one of its main functions is to facilitate communication between individuals. Social networks, as the name itself says, function as a large virtual community, people in general have never been so exposed to the comments of third parties, nor has there ever been such ease in making comments about others, many even do so anonymously, which intensifies a sense of impunity regarding what is said, a fact that contributes when the intention of the comment is only to express hate. Hate crimes can be destructive in the lives of victims and must be treated with seriousness by governmental and private organizations that manage social networks. Aggressors benefit from the variation in national legislation on hate speech, from the difficulty of establishing a boundary for the constantly evolving cyberspace, and from algorithms that provide greater engagement to controversial comments. The spread of online hate speech challenges policymakers and the research community worldwide. To assist in combating these crimes, several techniques of natural language processing, data analysis, and threat detection have already been tested and presented through scientific texts which served as inspiration for the creation of this work. In this monograph, data mining techniques and supervised machine learning models were implemented to train models capable of identifying these crimes in diverse texts. To test them, a dataset generated dynamically by the Dynabench project from the context of common hate crimes that occur on the internet was used. Each technique tested is compared in order to document the limitations, best models, and strategies for detecting this type of crime.