Brazilian Portuguese Hate Speech Classification using BERTimbau

Abstract

Hate speech is a language that attacks or denigrates a specific group based on their characteristics, such as their race, ethnicity, or sexual orientation. Hate speech became widespread and spread through social networks, blogs, videos, and other communication channels. With anonymity and a sense of impunity, people feel encouraged to spread their hatred on the internet. In this work, we used the BERT model for the Portuguese language called BERTimbau to classify hate speech in three datasets in Portuguese, available in the literature: OFFCOMBR-2, OFFCOMBR-3, and Fortuna et. al. (2019) dataset. Still, we performed some preprocessing and an oversampling technique on the datasets. Finally, we compared the results obtained with results obtained by works available in the literature. Experiments with BERTimbau, using preprocessing and oversampling obtained better results than other classification techniques.