NoHateBrazil: A Brazilian Portuguese Text Offensiveness Analysis System

Proceedings of Recent Advances in Natural Language Processing, pages 1180–1186 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_125 1180 NoHateBrazil: A Brazilian Portuguese Text Offensiveness Analysis System Francielle Vargas12, Isabelle Carvalho1, Wolfgang S. Schmeisser-Nieto3 Fabrício Benevenuto2, Thiago A. S. Pardo1 1Institute of Mathematical and Computer Sciences, University of São Paulo, Brazil 2Computer Science Department, Federal University of Minas Gerais, Brazil 3 Department of Linguistics, University of Barcelona, Spain francielleavargas@usp.br, isabelle.carvalho@alumni.usp.br wolfgang.schmeisser@ub.edu, fabricio@dcc.ufmg.br, taspardo@icmc.usp.br Abstract Hate speech is a surely relevant problem in Brazil. Nevertheless, its regulation is not effec- tive due to the difficulty to identify, quantify and classify offensive comments. Here, we in- troduce a novel system for offensive comment analysis in Brazilian Portuguese. The system ti- tled NoHateBrazil1 recognizes explicit and im- plicit offensiveness in context at a fine-grained level. Specifically, we propose a framework for data collection, human annotation and machine learning models that were used to build the sys- tem. In addition, we assess the potential of our system to reflect stereotypical beliefs against marginalized groups by contrasting them with counter-stereotypes. As a result, a friendly web application was implemented, which be- sides presenting relevant performance, showed promising results towards mitigation of the risk of reinforcing social stereotypes. Lastly, new measures were proposed to improve the ex- plainability of offensiveness classification and reliability of the model’s predictions. 1 Introduction The scenario of hateful comments in Brazil is se- vere and entails the creation of safety and fairness technologies. During the elections in 2018 and 2022, the denunciations against xenophobia con- tent had an increase of 2,369.5%; apology and pub- lic incitement to violence and crimes against life, 630.52%, and misogyny and race-ethical, increased by 1,639% and 595%2, respectively. Hate speech is a particular form of offensive lan- guage that considers stereotypes to express an ide- ology of hate (Warner and Hirschberg, 2012; Sahoo et al., 2022; AlKhamissi et al., 2022). While sys- tems that classify hateful content are undoubtedly relevant, these technologies are being developed with scarce consideration of their potential biases (Nadeem et al., 2021; Sap et al., 2019; Chang et al., 1Demo: http://143.107.183.175:14581/ 2https://new.safernet.org.br/ 2019; Bordia and Bowman, 2019; Blodgett et al., 2020). These systems may discriminate against the groups they are designed to protect (Davidson et al., 2019), reflecting social stereotypes and being able to perpetuate social inequalities when propagated at scale (Davani et al., 2023). To the best of our knowledge, no systems have attempted to analyze text offensiveness in Brazilian Portuguese. Therefore, the main contribution of this paper3 is providing the first web system titled NoHateBrazil for Brazilian Portuguese offensive comments classification. The NoHateBrazil sys- tem receives two different inputs. The first input consists of a single comment written directly into the initial screen. The second input consists of a file in CSV format containing a set of comments. In the following outputs, three pieces of informa- tion are exhibited: (i) offensiveness categories; (ii) offensiveness overall score; and (iii) prediction re- liability score, which we describe in Section 2.1. Towards providing a reliable text …

Content truncated to 500 words