OLID-BR: Offensive Language Identification Dataset for Brazilian Portuguese

Social media has transformed the way society is connect. Although there are many advantages to such large scale connection, there are unquestionable drawbacks, for example in the way fake news is easily spread and in particular the vast dissemination of hate speech. Identifying offensive comments is a critical task for ensuring the safety of users, which is why industry and academia have been working on developing solutions to this problem. Previous work on the automatic detection of offensive comments focuses mainly on English, with very little work for languages such as Portuguese. In this paper, we present a new dataset for toxicity detection in Portuguese containing 6,354 (extendable to 13,538) comments annotated using a hierarchical scheme with multiple levels of granularity, which gives us the possibility of training models for multiple NLP tasks. The dataset is used to train machine learning models which can be useful to identify several aspects of hate speech. The results of the experiments show that the dataset is useful in the development of intelligent systems for the detection of offensive comments in Portuguese.