In the Information Systems (IS) discipline, central contributions of research projects are often represented in graphical research models, clearly illustrating constructs and their relationships. Although thousands of such representations exist, methods for extracting this source of knowledge are still in an early stage. We present a method for (1) extracting graphical research models from articles, (2) generating synthetic training data for (3) performing object detection with a neural network, and (4) a graph reconstruction algorithm to (5) storing results into a designated research model format. We trained YOLOv7 on 20,000 generated diagrams and evaluated its performance on 100 manually reconstructed diagrams from the Senior Scholars' Basket. The results for extracting graphical research models show a F1-score of 0.82 for nodes, 0.72 for links, and an accuracy of 0.72 for labels, indicating the applicability for supporting the population of knowledge repositories contributing to knowledge synthesis
Huettemann, Sebastian; Mueller, Roland M.; Larsen, Kai R.; Dinter, Barbara; and Campos Chiny, Joshua, "How Best to Hunt a Mammoth - Toward Automated Knowledge Extraction From Graphical Research Models" (2023). Wirtschaftsinformatik 2023 Proceedings. 87.