Contextual-Lexicon Approach for Abusive Language Detection

Proceedings of Recent Advances in Natural Language Processing, pages 1438–1447 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_161 1438 Contextual-Lexicon Approach for Abusive Language Detection Francielle Vargas*†, Fabiana G´oes*, Isabelle Carvalho‡ Fabr´ıcio Benevenuto†, Thiago A. S. Pardo* *Institute of Mathematical and Computer Sciences, University of S˜ao Paulo, Brazil ‡Ribeir˜ao Preto Medical School, University of S˜ao Paulo, Brazil †Computer Science Department, Federal University of Minas Gerais, Brazil {francielleavargas,fabianagoes,isabelle.carvalho}@usp.br fabricio@dcc.ufmg.br, taspardo@icmc.usp.br Abstract Since a lexicon-based approach is more el- egant scientifically, explaining the solution components and being easier to generalize to other applications, this paper provides a new approach for offensive language and hate speech detection on social media, which em- bodies a lexicon of implicit and explicit offen- sive and swearing expressions annotated with contextual information. Due to the severity of the social media abusive comments in Brazil, and the lack of research in Portuguese, Brazil- ian Portuguese is the language used to validate the models. Nevertheless, our method may be applied to any other language. The con- ducted experiments show the effectiveness of the proposed approach, outperforming the cur- rent baseline methods for the Portuguese lan- guage. 1 Introduction In Brazil, hate speech is prohibited. Nevertheless, in government and civil society, the regulation of hate speech is not effective due to the difficulty to identify, quantify and classify abusive comments. Indeed, this is rather a difficult requirement to sat- isfy. According to Mesquita (2018), the Safer- net non-governmental organization, which oper- ates in cooperation with public organizations in Brazil, as well as companies, such as Google, Facebook, and Twitter, proposed a collection of data on actions that violate human rights. The data is very worrisome: during the 2018 year’s elec- tion period, denunciations with xenophobia con- tent had an increase of 2,369.5%; apology and public incitement to violence and crimes against life, 630.52%; neo-nazism, 548.4%; homopho- bia, 350.2%; racism, 218.2%; and religious in- tolerance, 145.13% 1. Figure 1 shows the hate 1https://www.bbc.com/portuguese/ brasil-46146756 crimes evolution that occurred in the most pop- ulous Brazilian state 2. The data was collected from S˜ao Paulo public security government. The pink line provides data on religious intolerance crimes, red on homophobia/transphobia, blue on race/ethnicity/color, green on region/origin, yel- low on political intolerance, and light green on other crimes. Figure 1: Hate crimes occurrence in S˜ao Paulo from 2016 to begining of 2020. Indeed, it is generally agreed that the high in- cidence of hate crimes is boosted by the popular- ization of online social networks. In social media, people and organizations may use the language to defamation, oppression, and terrorism. The lan- guage used intentionally in order to disrespect, in- sult or attack the reader is denominated in litera- ture by abusive language, unless otherwise stated such as offensive language (C¸ ¨oltekin, 2020; Pite- nis et al., 2020; Razavi et al., 2010), hate speech (Schmidt and Wiegand, 2017; Waseem and Hovy, 2016) and cyberbullying (Rosa et al., 2019). According to Warner and Hirschberg (2012), hate speech is a particular form of abusive lan- 2https://www.ssp.sp.gov.br/ 1439 guage considering stereotypes …

Content truncated to 500 words