Abstract

Structured Sentiment Analysis (SSA) aims to extract fine-grained opinion structures as tuples (holder, target, expression, polarity). While recent advances have improved SSA for English, Brazilian Portuguese lacks dedicated resources. This paper presents an exploratory study introducing a manually annotated dataset of hotel reviews for SSA in Brazilian Portuguese. We propose a baseline approach fine-tuning the BERTimbau model under a BIO tagging scheme to extract sentiment spans. Unlike traditional approaches that model relations explicitly, we assess the viability of span-level extraction as a first step for SSA in this language. Experimental results using a strict train/validation/test split show that our approach achieves a span-level F1-score of 48.41 for holder extraction and a macro F1-score of 61.52. We also discuss the linguistic challenges of holder extraction in Portuguese, specifically regarding implicit subjects (pro-drop), and provide a detailed error analysis. These results establish a preliminary baseline for future relation-aware models in Portuguese.