This study evaluates the ability of large language models (LLMs) to detect incoherence between the text of product reviews and their assigned rating (1 or 5 stars). Using popular LLMs such as GPT-5, Llama-4 and DeepSeek-3.2, and models optimized for Brazilian Portuguese, Sabiá-3.1 and Bode-3.1, we show that some are capable of detecting incoherence among texts and ratings (F1 > 90%) in a zero-shot protocol. Models also present a high agreement in the predictions, where several prediction rounds led to low variability (Fleiss’ κ> 0.95). With the demonstrated incoherence present in all product categories (aprox. 10% of comments), the results suggest that LLMs are very promising to perform this high semantic interpretation task, and they can be used as valuable tools for online monitoring and recommendation systems.

The analysis of unstructured civil petitions is often hindered by procedural noise and verbose argumentation. To address this, we propose a pipeline composed of LLM-based fact extraction followed by legal-domain embeddings of texts for unsupervised density clustering. We employ Large Language Models to isolate factual narratives from raw texts, which are then encoded using domain-specific representations (Legal-BERT) and grouped via UMAP dimensionality reduction and the HDBSCAN algorithm. Comparative experiments on a Brazilian judicial corpus reveal that clustering based solely on extracted yields significantly more cohesive and semantically well-defined groups than, which suffer from fragmentation due to content variability. Results indicate that the proposed method is a promising approach for thematic organization, procedural triage support, and large-scale discovery of legal patterns.

This study analyzes gender identification in Brazilian Portuguese using Amazon reviews drawn from ten product categories. Nine models were evaluated: three classical classifiers (Logistic Regression, Random Forest, and SVM), a multilingual BERT, and five LLMs (ChatGPT 4o, ChatGPT 3.5, DeepSeek, Sabia3, and Sabiazinho). Experiments show that BERT achieved the best performance (macro-F1 = 0.634), outperforming ChatGPT 4o and Logistic Regression by less than one percentage point. Reviews authored by women reach an average F1 of 0.654—four points higher than those by men. Performance also varies by domain: books and automotive are easier, whereas baby and pets are more challenging.