The Effect of Knowledge Graph Schema on Classifying Future Research Suggestions

Dimitrios Alivanistos, Seth van der Bijl, Michael Cochez, Frank van Harmelen

Abstract: The output of research doubles at least every 20 years and in most research fields the number of research papers has become overwhelming. A critical task for researchers is to find promising future directions and interesting scientific challenges in the literature. To tackle this problem, we hypothesize that structured representations of information in the literature can be used to identify these elements. Specifically, we look at structured representations in the form of Knowledge Graphs (KGs) and we investigate how using different input schemas for extraction impacts the performance on the tasks of classifying sentences as future directions. Our results show that the MECHANIC-Granular schema yields the best performance across different settings and achieves state of the art performance when combined with pretrained embeddings. Overall, we observe that schemas with limited variation in the resulting node degrees and significant interconnectedness lead to the best downstream classification performance.