Deep Neural Networks for Automated Metadata Extraction

Metadata extraction is essential for the discovery of scientific information. However, it can be tedious and requires a lot of time and effort, given that the number of scientific publications is increasing rapidly. In this work, we developed a solution based on deep learning to automatically extract information relating to metadata from scientific articles. The architecture based on this solution permits segmenting and labeling the main areas of scientific documents, such as title, abstract, sections, image, etc., while simultaneously recognizing the text of the detected areas. The proposed solution is based on object detection models, specifically Faster Region-based Convolutional Neural Networks (Faster R-CNN), Mask Region-based Convolutional Neural Networks (Mask R-CNN), and Cascade Mask Region-based Convolutional Neural Networks (Cascade Mask R-CNN), in order to detect the layout. The areas of interest found are then passed through an OCR (Optical Character Recognition) for text recognition. The model was trained using annotated data from the SciBank dataset. During this study, we had the opportunity to test three different models and compare the results obtained. The results from the proposed solution show promise.