Processing Research-related Information with Machine learning and Knowledge Graphs

This dissertation is devoted to machine learning in scientific knowledge graphs, a field concerned with constructing and studying methods that can learn from scientific data (especially scholarly documents) and the resulting graphs. The contributions of this dissertation are divided into four parts. An essential input for the automated processing of scholarly documents is their thematic classification, which needs to be more accurate and complete in the metadata. In the first part, we classify professional documents into thematic categories related to COVID-19. In the thesis, we evaluate several machine learning methods, while in addition to the correctness of the classification, we also focus on the interpretation of the created machine learning models. We also analyzed typical classification errors, which may contribute to further developing the respective methods. A large number of scholarly documents are currently available. As part of the dissertation, machine learning methods were investigated to determine the meaning or the importance of individual documents derived from their citation rate. The dissertation examined several factors that could improve the quality of the prediction of the importance of scientific documents, especially the use of thematic categorization of documents. As part of the experiments, a meaningful classifier was created separately for each thematic group, which made it possible to reflect domain specificities. The results were compared with a classifier built for all documents without categorizing. Relevant information for the automated processing of scholarly documents can also be obtained from external knowledge graphs, which can be used to enrich the content of scholarly documents. Therefore, a study was conducted regarding the impact of the selected knowledge graph (DBpedia) on classification accuracy. This study only analyzes domain-independent knowledge base effects, but in the future, we plan to investigate the impact of domain-specific knowledge bases containing domain-specific information. An essential aspect of the scholarly knowledge graph is the practical use of scientific information. The last part of the dissertation deals with the possibilities of incorporating the results of machine learning tasks into the Open Research Knowledge Graph, a knowledge graph describing scientific documents developed at the Leibniz Information Center for Science and Technology.