Abstract:
This study investigates the application of Large Language Models (LLMs) to Unsupervised Aspect-Based Sentiment Analysis (ABSA) in Portuguese, focusing on consumer reviews of Brazilian beers. We construct a novel domain-specific dataset comprising nearly 60,000 filtered reviews collected from a major Brazilian beer forum, along with a manually annotated gold-standard subset containing 1,712 labeled Beer Characteristics (BC), categories, and sentiment polarities. Two LLM families—one monolingual (Sabiá-3) and one multilingual (GPT-4o mini)—are systematically evaluated under zero-shot, one-shot, and few-shot prompting strategies. Results show that the best configuration achieves F1-scores of 0.857 for aspect extraction, 0.826 for category identification, and 0.552 for sentiment classification, with sentiment detection proving the most sensitive to prompt design. Using the optimal model and prompt configuration, a largescale annotated dataset of over 880,000 extracted BC instances is generated, enabling longitudinal analysis of consumer perceptions from 2008 to 2025. Findings indicate that positive attributes such as refreshing mouthfeel and persistent foam dominate consumer praise. The proposed pipeline demonstrates that LLMbased ABSA can effectively uncover fine-grained consumer preferences in low-resource languages without extensive labeled data, offering a scalable and cost-effective tool for market intelligence and product development in the brewing industry.
Referência:
EIRAS, Denis M.A.; BRITO, Adriana Camargo de. Unsupervised aspect-based sentimento analysisthrough LLM: a case study of na unlabeled portuguese beer database. IEEE Access, 2026.
Acesso ao artigo no site do Periódico:
https://ieeexplore.ieee.org/abstract/document/11505998/authors