- Keywords: Ontology, Education, Lanugage Models, BERT
- TL;DR: Baseline performance of BERT on entity and extraction from textbooks, and a novel labeling task for improving its performance
- Abstract: Ontology graphs are graphs in which the nodes are generic classes and edges have labels that specify the relationships between the classes. In this paper, we address the question:to what extent can automated extraction and crowdsourcing techniques be combined to boostrap the creation of comprehensive and accurate ontology knowledge graphs? By adapting the state-of-the-art language model BERT to the task, and leveraging a novel relationship selection task, we show that even though it is difficult to achieve a high precision and recall, automated term extraction and crowd sourcing provide a way to bootstrap the ontology graph creation for further refinement and improvement through human effort.
- Archival Status: Archival