Socially responsible and explainable automated fact-checking and hate speech detection

• Ano de defesa: 2024
• Autor(a) principal: Vargas, Francielle Alves
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Automated fact-checking and hate speech detection; Explainability and interpretability; Explicabilidade e interpretabilidade; Fairness; Inteligência artificial responsável; Justiça; Natural language processing; Processamento de linguagem natural; Responsible artificial intelligence; Verificação automatizada de fatos e detecção de discurso de ódio
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-07012025-155212/

Misinformation and hate speech have a negative impact on society, particularly in conflict-affected areas and politically polarized countries. These issues are fueled by longstanding and ingrained social, cultural, political, ethnic, religious, and other divisions and rivalries, often exacerbated by misinformation through sophisticated belief systems, including propaganda and conspiracy theories. In this context, a wide range of models for fact-checking and hate speech detection have been proposed. However, while Natural Language Processing (NLP) has traditionally relied on techniques that are inherently explainable (often referred to as white box techniques, such as rule-based algorithms, decision trees, hidden markov models, and logistic regressions), the advent and popularity of Large-Scale Language Models (LLMs)-often considered black box techniques has led to a decrease in interpretability. The use of language embeddings as features has further contributed to this trend. Consequently, most existing fact-checking and hate speech detection models struggle to provide relevant rationales (explanations) for their predictions, highlighting a lack of transparency that poses significant risks, such as the prevalence of biases. This has recently been identified as a major concern in the field. For instance, biases in hate speech technologies may reinforce discrimination against groups based on their social identity when propagated at scale. Similarly, biases in fact-checking may increase political polarization due to limitations in impartiality or deliberate attempts to favor or disfavor particular individuals or viewpoints. To address these critical gaps, this thesis introduces a study of fact-checking and hate speech detection technologies and their potential ethical implications. Specifically, it provides five benchmark data resources (e.g. HateBR, HateBRXplain, HausaHate, MOL, and FactNews) and three new methods (e.g., SELFAR, SSA, and B+M) for automated fact-checking and hate speech detection, ensuring that the data and models are explainable and socially responsible. Notably, the HateBR and the B+M outperformed current baselines in Portuguese. Ultimately, we hope that our study, data resources, and methods will advance research on misinformation and hate speech, significantly contributing to the ongoing discussions on responsible AI, explainability and interpretability, and fairness in Natural Language Processing and Machine Learning.