Classificação da utilidade de respostas em saúde com base em votos comunitários: uma abordagem automatizada para sistemas de qa

Abstract

This work investigates the classification of answer usefulness in health-related Question and Answer services. We propose a supervised method that combines Natural Language Processing techniques with community interaction signals to prioritize useful answers. We constructed the SaudeBR-QA corpus from CatálogoMed, from which 3,010 question–answer pairs were selected and labeled based on votes/likes. Seven classifiers were optimized via stratified cross-validation. Random Forest achieved the best balance between accuracy and F1 on the test set (79.07%) accuracy and 79% F1macro). A user study, conducted through an online form, collected judgments from both the general public and health professionals, revealing alignment between the model and human consensus as well as occasional divergences. The results demonstrate the feasibility of lightweight models to support answer ranking and improve access to medical information, and they also provide a public corpus and a reproducible protocol for future research.