http://hdl.handle.net/10071/44 Wed, 01 Apr 2026 05:57:06 GMT 2026-04-01T05:57:06Z http://hdl.handle.net/10071/36168 Título próprio: Aprendizado por transferência para correção automática de redação Autoria: Silveira, I. C.; Ribeiro, E.; Mamede, N.; Baptista, J. Resumo: A tarefa de Correção Automática de Redação tem despertado crescente interesse na área de processamento de texto em português. Entre os conjuntos de dados disponíveis, destaca-se um corpus de redações narrativas produzidas por alunos do 5º ao 9º ano do ensino fundamental no Brasil. Essas redações são avaliadas segundo quatro competências: registro formal, coerência temática, estrutura retórica narrativa e coesão textual. Este trabalho explora a criação de um sistema baseado em conhecimentos derivados de outro dataset (desenvolvido com base em textos produzidos para o ENEM) e de outras tarefas (cálculo de complexidade textual e análise de legibilidade). O sistema desenvolvido combina modelos neurais, características (features) curadas calculadas por programas de análise textual e seleção de features em um modelo de Aprendizado em Dois Estágios. Com isso, foi possível elevar a performance em relação ao estado-da-arte, nomeadamente, em 9% para a primeira competência, 5,5% para a terceira e 8,9% para a quarta. Wed, 01 Jan 2025 00:00:00 GMT http://hdl.handle.net/10071/36168 2025-01-01T00:00:00Z http://hdl.handle.net/10071/35734 Título próprio: Enhanced multiple instance learning for breast cancer detection in mammography: Adaptive patching, advanced pooling, and deep supervision Autoria: Sarwar, Fareeha; Garrido, Nuno Miguel de Figueiredo; Sebastiao, Pedro; Silveira, Margarida Resumo: This paper addresses the challenge of weakly supervised learning for breast cancer detection in mammography by introducing an Enhanced Embedded Space MI-Net model with deep supervision. The framework integrated adaptive patch creation, convolution feature extraction, and pooling methods -max, mean, log-sum-expo, attention, and gated attention pooling - evaluated in three MIL models, Instance Space mi-Net, Embedded Space MI-Net and Enhanced Embedded Space MI-Net. A key contribution is the incorporation of deep supervision, improving feature learning across network layers and enhancing bag-level classification performance. Experimental results on the CBIS / DDSM dataset demonstrate that the Enhanced MI-Net model achieves the highest AUC of 86% with attention pooling. This work addresses the gap in leveraging MIL techniques for high-resolution medical imaging without requiring detailed annotations, offering a robust and scalable solution for breast cancer detection.Clinical Relevance-This study highlights the potential of MIL-based models with attention pooling to accurately detect breast cancer in mammographic images without requiring detailed ROI annotations, offering a scalable and efficient diagnostic tool for clinical practice. Tue, 01 Jul 2025 00:00:00 GMT http://hdl.handle.net/10071/35734 2025-07-01T00:00:00Z http://hdl.handle.net/10071/35633 Título próprio: A transformer-based deep learning approach for detecting online hate speech in Spanish Autoria: Sanchez-Gomez, J. M.; Batista, F.; Vega-Rodríguez, M. A.; Pérez, C. J. Resumo: The amount of content published on the Internet has grown exponentially in recent times. Social networks have enabled this content to reach an even wider audience. However, the freedom of communication provided by these networks can consequently facilitate the spread of offensive language and hate speech. Although social media platforms have attempted to implement mechanisms for detecting and addressing such content, it remains an ongoing challenge, particularly for languages other than English, such as Spanish. One promising approach to tackle this problem is the application of Natural Language Processing (NLP) tools, which rely on the use of language models and deep learning for text classification. In this work, an approach for detecting Spanish Hate Speech with ALBETO (SHS-ALBETO) is proposed. Experimentation is conducted with HatEval dataset. The performance of SHS-ALBETO is compared with other competing models, such as BERT, BETO, and DistilBETO, along with other proposals from the state-of-the-art. SHS-ALBETO has improved the existing results in the scientific literature, simultaneously providing reduced computing times. Additionally, analyses of the results have revealed its advantages together with challenging aspects that must be addressed to further improve the performance of this kind of approach. Wed, 01 Jan 2025 00:00:00 GMT http://hdl.handle.net/10071/35633 2025-01-01T00:00:00Z http://hdl.handle.net/10071/35515 Título próprio: Unpacking online hate speech in Portuguese social media: A social-psychological and linguistic-discursive approach Autoria: Guerra, R.; Carvalho, P.; Marques, C.; Carmona, M.; Sarroeira, R.; Batista, F.; Ribeiro, R.; Fonseca, A.; Moro, S.; Silva, C. Resumo: Building on social psychology and language sciences, this research identified core social psychological, and linguistic-discursive features of online hate speech targeting racialized, migrant and LGBTI+ communities in two social media platforms in Portugal: YouTube, and Twitter/X. The research was based on the analysis of two annotated corpora comprising 24,739 YouTube comments and associated replies, and 29,758 contextualized tweets retrieved from 2775 conversations. Overall, the results, based on the detailed annotation framework developed in this study, revealed that i) online hate speech was mainly expressed in subtle ways (i.e., indirect hate speech); ii) the main underlying process of discrimination in both direct and indirect hate speech was outgroup derogation; iii) stereotypes, threats, and dehumanization were frequently used as discursive strategies to express online hate speech; iv) specific features, like emotions, often overlooked in hate speech annotated corpora, varied in their expression depending on the specific target community; v) the use of some discursive strategies, such as realistic and symbolic threats, seem to be dependent not only on the target community but also the social media platform; vi) discursive strategies and emotions mobilized in hate speech were correlated with specific rhetorical devices and fallacies. These findings provide valuable insights into the complex landscape of online hate speech and highlight the importance of interdisciplinary, context and culturally sensitive approaches in understanding this phenomenon. Wed, 01 Jan 2025 00:00:00 GMT http://hdl.handle.net/10071/35515 2025-01-01T00:00:00Z