While Part-of-speech tagging is considered a well-understood task in the literature, most work has focused on indo-european languages with fusional morphology. For low-resource agglutinative languages, such as brazillian indigenous languages, the task is commonly constrained by small annotated corpora, high lexical sparsity, and morphological patterns that are poorly represented by word-level models. This paper investigates the impact of unsupervised subword segmentation techniques on POS tagging for brazillian indigenous languages. We compare a word-level baseline with Byte Pair Encoding, Morfessor, and FlatCat on Bororo, Nheengatu, and Tupinamba corpora, including a controlled experiment on the size of the training corpus. Our findings suggest that unsupervised segmentation can reduce sparsity in low-resource POS tagging, although its benefit depends on the language, corpus size, and segmentation method.

The use of textbooks as primary sources of information has increasingly given way to tools based on Large Language Models (LLMs), raising concerns about the reliability of generated answers. This study investigates how different adaptation strategies shape the behavior of small language models in educational Question Answering (QA) tasks in Portuguese. To support this analysis, we built a question-answer dataset derived from an NLP textbook and compared base models and the Retrieval-Augmented Generation (RAG) pipeline with models adapted through supervised fine-tuning and Continued Pretraining. The evaluation relies on questions from the LARI dataset, which has been validated by human specialists, and combines automatic and qualitative assessment procedures. The findings indicate that small models tuned with structured instructional knowledge achieve stronger semantic alignment and produce more pertinent answers in Portuguese educational QA scenarios.

Este trabalho propõe a identificação de vieses sociais em português nos modelos GPT-4o, GPT-4o-mini, Sabiá-3 e Sabiázinho-3, utilizando a métrica de estima a fim de avaliar o nível de respeito e deferência dos modelos sobre diferentes grupos demográficos. A avaliação abrange sujeitos com marcadores sociais explícitos de género, raça e região brasileira, em condições com e sem o uso de uma técnica de contorno das restrições de moderação (jailbreaking). Os achados mostram que os modelos de linguagem avaliados reproduzem padrões sistemáticos de valoração diferenciada entre grupos sociais, revelando vieses de estima associados a marcadores de gênero, raça e região no português brasileiro. Sujeitos com marcadores sociais enfatizados, especialmente os de raça, tendem a receber estimas mais baixas. A utilização da técnica de jailbreaking não apresentou um impacto uniforme, podendo tanto ampliar quanto reduzir as diferenças de estima.