Comparação de estratégias de prompting em LLMS para classificação de texto em português: um estudo empírico entre zero-shot, few-shot e chain-of-thought

Abstract

The field of Natural Language Processing (NLP) has been profoundly impacted by the emergence of Large Language Models (LLMs), which introduced the in-context learning paradigm. This approach allows for performing complex tasks via prompts, reducing the need for traditional supervised training. However, for specific tasks in Portuguese, questions remain regarding when these new strategies justify their computational cost compared to established classical methods. This work investigates this scenario by empirically comparing different prompting strategies — including zero-shot (via textual inference and embeddings), few-shot, and Chain-of-Thought (CoT) — in the sentiment classification task. Using a balanced synthetic dataset, the study evaluated not only efficacy (Macro-F1) but also CPU latency, benchmarking LLMs against a supervised baseline (TF-IDF + SVM). Results indicate that, for data with clear lexical cues, classical techniques still offer superior performance and maximum efficiency. Among unsupervised approaches, the use of Natural Language Inference (NLI) proved to be the most robust alternative in terms of quality, while the embeddings-based strategy stood out for its speed. Generative few-shot approaches, in turn, revealed stability challenges in adhering to rigid labels. It is concluded that the adoption of LLMs must be judicious, weighing the availability of labeled data against infrastructure constraints.