Innovation tournaments require expert evaluators to assess large volumes of lengthy, unstructured proposals under significant time pressure and cognitive constraints. This study investigates how transformer-based natural language processing (NLP) models can be integrated into this context as decision-support tools within the evaluation process. Using real-world data from a national Brazilian innovation tournament comprising 1151 startup proposals, we examine whether NLP models can learn patterns associated with expert-assigned scores and support early-stage screening. We fine-tune Portuguese-language transformer models (BERT) for classification tasks and refine long-text processing through chunking and document-level aggregation. To contextualize performance, we benchmark the neural models against lighter architectures and classical machine-learning baselines. Although the models exhibit persistent ambiguity in distinguishing intermediate-quality proposals, the best-performing configuration consistently avoids extreme misclassifications and adopts a conservative screening profile that prioritizes the retention of high-potential submissions.
Santiago et al. (Fri,) studied this question.