Using CNNs to Detect Graphical Representations of Structural Equation Models in IS Papers

Literature reviews are an essential but time-consuming part of every research endeavor and play an important role in the quality of the research findings. Traditional tools and literature databases only make use of the textual information and do not consider graphical representations like figures of structural equation models (SEMs). These models are often used in empirical studies to visualize theoretical models and key results. We design and implement an application for image recognition to simplify the search for relevant papers, by automatically recognizing SEM figures in scientific papers stored as PDF files. To classify whether a page in a paper contains an SEM figure we make use of convolutional neural networks and achieve an F1 score of 98,7% together with a recall of 100% for the SEM class. We further describe how we intend to automatically extract information from these SEM figures.