MFTCXplain : A Brazilian Election Hate Speech Morality Corpus with Human-Annotated Rationales

The Brazilian Election Hate Speech Morality Corpus with Rationale (MFTCXplain) is a multidisciplinary, cross-cultural initiative designed to explore the relationship between morality, hate speech, misinformation, and election propaganda on Twitter during the Brazilian elections. The project has created an annotated corpus of 5,550 tweets and comments from Instagram in Portuguese, focusing on offensive language and hate speech and using moral foundation theory (MFT) to capture moral values. Each annotation includes both the moral labels and the rationale behind the label, providing a rich dataset for investigating explainability in AI, as well as bias mitigation. This allows for testing various AI model explainability approaches, including self-explanation and post-hoc explanation, while also exploring how to mitigate bias in morally and culturally sensitive tasks.

The MFTCxplain project involves a diverse team of social psychologists, computer scientists, and linguists, from Brazil, Germany, Italy, Australia, Los Angeles, and Portland. Annotators come from three distinct cultural regions in Brazil: the north (with a black rural indigenous population), the south (notably European-influenced), and the southeast (including the highly diverse city of São Paulo). This cultural diversity allows the team to measure biases and subjectivity across regions, providing insights into both intra-national and cross-national differences. The dataset will also help compare models trained in different cultural contexts, uncovering how moral, political, and cultural values influence the perception and spread of hate speech, misinformation, and propaganda detection. The timing of this corpus is particularly relevant due to the media frenzy surrounding the Brazilian elections. A former ally of Brazilian President Jair Bolsonaro claimed that the leader used a government-funded ``digital militia'' to spread propaganda targeting political enemies, particularly focusing on religious and cultural issues to polarize the electorate (Rennó, 2020). This highlights the significance of misinformation and propaganda during this period, making the project's focus on moral tones in speech highly relevant.

Previous research has established the link between morality and shared content in digital environments (Abdurahman et al., 2023), as well as the connection between morality and hate speech (Kennedy et al., 2023), supporting the need for a detailed analysis of the moral framing of election propaganda. Proposed and initiated by Francielle Vargas, a computer scientist at the University of São Paulo, and project manager Jackson Trager, a social psychologist at the University of Southern California, this project is at the intersection of sociotechnical systems, AI and society, election studies, and culturally sensitive AI. It aims to contribute to these fields through representative data collection and explainability in AI models, with a specific focus on bias mitigation strategies in cross-cultural contexts.

Key Terms: Natural Language Processing; Sociotechnical Systems; AI and Society; Election Studies; Human-Centered AI; Representative Data Collection; Culturally Sensitive AI; Misinformation, Hate Speech, and Propaganda; Explainability in AI Models; Bias Mitigation.

References
  • Abdurahman, A., Trager, J., Atari, M., Mostafazadeh Davani, A., & Dehghani, M. (2023). Morality and shared content in the digital age. Journal of Social Media Studies, 15(3), 117-133.
  • Kennedy, B., Golazizian, P., Trager, J., Atari, M., Hoover, J., Mostafazadeh Davani, A., & Dehghani, M. (2023). The (moral) language of hate. PNAS Nexus, 2(7).
  • Samuels, David. (2018). Money, Elections, and Democracy in Brazil. Latin American Politics and Society, Volume 43 , Issue 2 , Summer 2001 , pp. 27 - 48.
  • Rennó, L. R. (2020). The Bolsonaro Voter: Issue Positions and Vote Choice in the 2018 Brazilian Presidential Elections. Latin American Politics and Society , Volume 62 , Issue 4 , November 2020 , pp. 1 - 23.

Leaders
Team
  • Flor Plaza. Department of Computer Science, Bocconi University, Italy
  • Diego Alves. Department of Language Science and Technology, Saarland University, Germany
  • Thales Bertaglia. Faculty of Law, Economics and Governance, Maastricht University, Nederlands
  • Matteo Guida. Faculty of Engineering and Information Systems, University of Melbourne, Australia
  • Morteza Dehghani. Department of Psychology and Computer Science, University of Southern California, USA

Publications
    • To appear


Resources
Dataset
  • MFTCXplain: A Benchmark Explainable Hate Speech Morality Dataset with Human-Annotated Rationales.

Sponsorship