OLID-BR: offensive language identification dataset for Brazilian Portuguese

Social media has revolutionized the manner in which our society is interconnected. While this extensive connectivity offers numerous benefits, it also accompanied by significant drawbacks, particularly terms of proliferation fake news and vast dissemination hate speech. Identifying offensive comments a critical task for ensuring safety users, why industry academia have been working on developing solutions to problem. Prior research speech detection predominantly focused English language, with few studies devoted other languages such as Portuguese. This paper introduces Offensive Language Identification Dataset Brazilian Portuguese (OLID-BR), high-quality NLP dataset language detection, we make publicly available. The contains 6,354 (extendable 13,538) labeled using fine-grained three-layer annotation schema compatible datasets languages, allows training multilingual/cross-lingual models. five tasks available OLID-BR allow comments, classification types offenses racism, LGBTQphobia, sexism, xenophobia, so on, identification type target extraction toxic spans comments. All those can enhance capabilities content moderation systems providing deep contextual analysis or highlighting that text toxic. We further experiment evaluate state-of-the-art BERT-based NER models, demonstrates usefulness development toxicity texts.