Knowledge-enhanced ontology-to-vector for automated ontology concept enrichment in BIM

Abstract

Building Information Modeling (BIM) relies on standardized ontologies like IfcOWL to address interoperability. However, the increasing complexity and diversity of construction information requirements demand automated enrichment of BIM ontologies, which is hindered by several factors, including complexity in ontology structure, scalability limitations, and domain-specific issues. Manual curation and maintenance of ontologies are labor-intensive and time-consuming, particularly as the scope of BIM projects expands. Despite these challenges, the construction industry lacks an effective automated approach for ontology concept enrichment. Thus, this study proposes a knowledge-enhanced ontology-to-vector (Keno2Vec) approach for automated BIM ontology concept enrichment, which can (1) encode ontology elements into meaningful and semantically rich embeddings by employing the BERT model to integrate both ontological information (names and labels) and external knowledge (definitions from authoritative knowledge bases), effectively addressing the domain expression specificity and complexity of BIM ontologies; and (2) provide a flexible framework that supports various downstream tasks of ontology concept enrichment by utilizing the resulting embeddings, thereby improving the task-specific adaptability and variability. Experimental results on datasets derived from the large-scale ifcOWL and two smaller BIM ontologies demonstrate that Keno2Vec significantly outperforms existing ontology embedding approaches in terms of accuracy and adaptability. For example, Keno2Vec achieves F1 scores on ifcOWL of nearly 87 % for subsumption prediction, 60 % for property identification, 95 % for membership recognition, and 100 % and 90 % for category-based and schema-based concept classification, respectively. Additional analysis highlights the potential of Keno2Vec for improving BIM ontology encoding and benefiting downstream applications.