From literature to research-based learning: An AI-powered information extraction system to enhance undergraduate thesis completion

Texto completo

Undergraduate thesis completion in STEM education is often constrained by limited timeframes, insufficient training, and the difficulty of navigating unstructured scientific literature. This study describes the design and evaluation of an AI-powered information extraction system that converts research publications into structured, traceable, and comparable datasets to support literature-driven comparison. The system is positioned as an epistemic scaffold that enables the inspection of evidence claim relationships through source traceability and structured comparisons while reducing low-level data-handling demands. A mixed-methods pilot study was conducted with 20 undergraduates from four schools and four academic supervisors within the same university, combining system log analysis, post-task surveys, and semi-structured interviews. Across 80 uploaded documents, students successfully extracted and standardized over 90% of the targeted experimental parameters and performance entries. Students' self-reported literature review time decreased by approximately 65%, and their ability to identify at least three influential variables in experimental designs increased by 50% in a brief pre/post variable-identification prompt. Survey and interview findings indicated that the structured outputs enabled students to transition from isolated reading toward cross-study comparison and evidence-based justification. Meanwhile, supervisors reported a reduced need for repeated instruction and improved students’ reasoning capacity. The pedagogical implications for data-informed supervision are discussed, along with limitations related to sample size, domain specificity, and the need to mitigate overreliance on automation through explicit verification routines. The modular architecture of the system suggests its transfer potential to other STEM domains in which parameter-based comparison is central. Given the exploratory, single-institution pilot design, quantitative results are interpreted as descriptive and mechanism-oriented rather than statistically generalizable.