Francielle Vargas

Hate speech detection models rely on surface-level lexical features, increasing vulnerability to spurious correlations and limiting robustness, cultural contextualization, and interpretability. We propose Supervised Moral Rationale Attention (SMRA), the first self-explaining hate speech detection framework to incorporate moral rationales as direct supervision for attention alignment. Based on Moral Foundations Theory, SMRA aligns token-level attention with expert-annotated moral rationales, guiding models to attend to morally salient spans rather than spurious lexical patterns. Unlike prior rationale-supervised or post-hoc approaches, SMRA integrates moral rationale supervision directly into the training objective, producing inherently interpretable and contextualized explanations. To support our framework, we also introduce HateBRMoralXplain, a Brazilian Portuguese benchmark dataset annotated with hate labels, moral categories, token-level moral rationales, and socio-political metadata. Across binary hate speech detection and multi-label moral sentiment classification, SMRA consistently improves performance (e.g., +0.9 and +1.5 F1, respectively) while substantially enhancing explanation faithfulness, increasing IoU F1 (+7.4 pp) and Token F1 (+5.0 pp). Although explanations become more concise, sufficiency improves (+2.3 pp) and fairness remains stable, indicating more faithful rationales without performance or bias trade-offs

Although Natural Language Processing (NLP) has traditionally relied on inherently interpretable “white-box” techniques, such as rule-based algorithms, decision trees, hidden Markov models, and logistic regression, the adoption of Large Language Models (LLMs) and language embeddings (often considered “black-box”) has significantly reduced interpretability. This lack of transparency introduces considerable risks, including biases, which have become a major concern in the field of Artificial Intelligence (AI). This Ph.D. thesis addresses these critical gaps by proposing new resources that ensure interpretability and fairness in NLP models for automated fact-checking and hate speech detection tasks. Specifically, we introduce five benchmark datasets (HateBR, HateBRXplain, HausaHate, MOL, and FactNews), four novel post-hoc and self-explaining methods (SELFAR, SSA, B+M and SRA), and one web platform (NoHateBrazil) designed to improve the interpretability and fairness of hate speech detection. The proposed models outperform the existing baselines for Portuguese and Hausa, both underrepresented languages. This research contributes to ongoing discussions on responsible and explainable AI, bridging the gap between model performance and interpretability to achieve positive real-world social impact. Finally, this thesis has had a significant impact both nationally and internationally, receiving citations from prestigious universities and research institutes abroad, inspiring new M.Sc. and Ph.D. in Brazil, and being recognized with multiple awards, including the Google LARA, the Maria Carolina Monard AI Award, and as a finalist at the Brazilian Computer Society Thesis and Dissertation Award.

Misinformation and hate speech form a socially harmful cycle. Research shows that misinformation can amplify hate speech targeting social identity groups and reinforce harmful stereotypes. To combat this cycle, a wide range of Natural Language Processing (NLP) methods have been proposed. Nevertheless, while NLP has historically relied on inherently explainable “white-box” techniques, such as rule-based algorithms, decision trees, hidden markov models, and logistic regression, the adoption of Large Language Models (LLMs) and language embeddings (often considered “black-box”) has significantly reduced interpretability. This lack of transparency introduces considerable risks, including biases, which have become a major concern in AI. This Ph.D. thesis addresses these critical gaps by proposing new resources that ensure explainability and bias mitigation in NLP models for these tasks. Specifically, it introduces five benchmark datasets (HateBR, HateBRXplain, HausaHate, MOL, and FactNews), three novel methods (SELFAR, SSA, and B+M), and one web system (NoHateBrazil) designed to improve the explainability and fairness of automated fact-checking and hate speech detection. The proposed models outperform existing baselines for Portuguese and Hausa, both underrepresented languages. This research contributes to ongoing discussions on responsible and explainable AI, bridging the gap between model performance and interpretability for realworld applications. Finally, it has had a significant impact both nationally and internationally, receiving citations from prestigious universities and research institutes abroad, and inspiring new M.Sc. and Ph.D. projects in Brazil.

Most existing discourse annotation guidelines have focused on the English language. As a result, there is a significant lack of research and resources concerning computational discourse-level language understanding and generation for other languages. To fill this relevant gap, we introduce the first discourse annotation guideline using the rhetorical structure theory (RST) for low-resource languages. Specifically, this guideline provides accurate examples of discourse coherence relations in three romance languages: Italian, Portuguese, and Spanish. We further discuss theoretical definitions of RST and compare different artificial intelligence discourse frameworks, hence offering a reliable and accessible survey to new researchers and annotators.

This paper provides data resources for low-resource hate speech detection. Specifically, we introduce two different data resources: (i) the HateBR 2.0 corpus, which is composed of 7,000 comments extracted from Brazilian politicians’ accounts on Instagram and manually annotated a binary class (offensive versus non-offensive) and hate speech targets. It consists of an updated version of the HateBR corpus, in which highly similar and one-word comments were replaced; and (ii) the multilingual offensive lexicon (MOL), which consists of 1,000 explicit and implicit terms and expressions annotated with context information. The lexicon also comprises native-speaker translations and its cultural adaptations in English, Spanish, French, German, and Turkish. Both corpus and lexicon were annotated by three different experts and achieved high inter-annotator agreement. Lastly, we implemented baseline experiments on the proposed data resources. Results demonstrate the reliability of data outperforming baseline dataset results in Portuguese, besides presenting promising results for hate speech detection in different languages.

Surendrabikram Thapa, Kritesh Rauniyar, Farhan Jafri, Hariram Veeramani, Raghav Jain, Sandesh Jain, Francielle Vargas, Ali Hürriyetoğlu, Usman Naseem. Proceedings of the 7th Workshop on Challenges and Applications of Automated Extraction of Socio-political Events from Text (CASE 2024). 2024.

Francielle Vargas, Isadora Salles, Diego Alves, Ameeta Agrawal, Thiago A. S. Pardo, Fabrício Benevenuto. Proceedings of the Seventh Fact Extraction and VERification Workshop (FEVER). 2024.

Francielle Vargas, Samuel Guimarães, Shamsuddeen Hassan Muhammad, Diego Alves, Ibrahim Said Ahmad, Idris Abdulmumin, Diallo Mohamed, Thiago Pardo, Fabrício Benevenuto. Proceedings of the 8th Workshop on Online Abuse and Harms (WOAH 2024). 2024.

This paper provides data resources for low-resource hate speech detection. Specifically, we introduce a large-scale expert annotated corpus of Brazilian Instagram comments and a context-aware offensive lexicon, which was manually extracted by a linguist from the proposed corpus and annotated with contextual information. We further provide native-speaker translations and adaptations from the specialized lexicon for other low-resource languages. The corpus consists of 7,000 document-level multi-layer annotations: (i) a binary offensive class, (ii) offensiveness-level classes, and (iii) nine hate speech targets. The context-aware offensive lexicon holds 1,000 explicit and implicit terms and expressions with pejorative connotations annotated with context-dependent offensiveness and context-independent offensiveness labels. Both corpus and lexicon were annotated by three different experts and achieved high inter-annotator agreement. Finally, we implemented baseline experiments on both data resources (corpus and lexicon). Results show the reliability of the proposed data, outperforming baseline dataset results in Portuguese, as well as presenting promising results for hate speech detection in different languages.