Desenvolvimento de IA para correção de avaliações de questões subjetivas

Abstract

This study develops and evaluates an Artificial Intelligence system for the automated grading of short-answer conceptual questions, utilizing Natural Language Processing and Deep Learning techniques. The system employs the BERTimbau model—a BERT variant pre-trained on Brazilian Portuguese—combined with a Multi-Layer Perceptron classifier layer, fine-tuned to categorize answers into three classes: correct, partially correct, and incorrect. To overcome the scarcity of labeled data, a local Retrieval-Augmented Generation (RAG) architecture is adopted, based on TF-IDF vectorization and cosine similarity, which selects relevant excerpts from the textbook to guide the generation of synthetic data via the Gemini model API. The system's interface was developed using the Streamlit library, allowing instructors to configure the question, the answer key, and the reference material, initiate on-demand training, and obtain the class grade report within a single session. An Active Learning cycle enables instructor-corrected discrepancies to be reinjected into the training set, promoting the continuous evolution of the model. Experiments demonstrated that the pipeline is technically feasible, achieving an accuracy close to 100% on synthetic data, a grade distribution semantically coherent with the answer key, and a total execution time between two and five minutes on consumer hardware. It is concluded that the solution represents a relevant proof of concept for the automation of the assessment process, with the potential to reduce the time spent by teachers on routine grading and to be adapted to different disciplines without requiring a prior database.