Identificação e análise de discurso sexista em comentários curtos provenientes de portais de notícia

Abstract

Identifying the spread of hate and prejudice speech is an ongoing challenge for online social media environments. Accurate characterization and identification are key to treating and mitigating violence, as well as educating users effectively and assertively. The online dissemination of hate can be targeted at different groups of people, which generates various classes of hate speech, such as racism, homophobia or sexism. The latter is a topic whose social interest objective has grown as the female figure overcomes the barriers of gender inequality. In particular, sexist speech propagates and encourages derogatory and abusive behavior against women. In this work, we build a corpus of sexist speech in Portuguese collected from popularly accepted news portals, such as G1 and UOL, for example. The work presents the three main contributions: (1) the process of creating the corpus and rotating comments (sexist/non-sexist); (2) a characterization and analysis of the corpus and behavior of anonymous labelers; (3) an assessment of machine learning techniques for detecting sexist/non-sexist comments. Preliminary results show that, by using automatic classifiers, it is possible to identify sexist comments with an F1 measure and precision above 0.9 and recall close to 0.8.