An Automated Relationship Classification to Support Semi-Automated IFC Extension
Jiansong Zhang, Nora El-Gohary · Construction Research Congress 2016 · 2016
The most popular building information modeling (BIM) data schema, the industry foundation classes (IFC), does not cover all necessary concepts and relations for representing information for supporting automated compliance checking (ACC). The IFC schema needs to be extended to fully support ACC. Previous IFC extension efforts typically extended the IFC schema in a manual (which entailed an effort-intensive and largely subjective) manner. A more automated (thus, a less effort-intensive and more objective) method for extending the IFC is needed. To address this gap, a new method for finding concepts from domain documents to incorporate into the current IFC class hierarchy is proposed. In this new IFC extension method, machine learning (ML)-based relationship classification is used to predict the relationship between each pair of matched concepts. This paper focuses on presenting the ML-based relationship classification method and its testing results. In the proposed method, a supervised ML-based approach is taken to train a classifier to automatically identify the relationship (equivalent, super, sub, or associated concept) between a pair of concepts. Four main ML algorithms were tested, including Naïve Bayes, decision tree, k-nearest neighbors, and support vector machines. Eight syntactic and semantic features of concept pairs were used in training the classifiers. The performances of the different ML algorithms were compared based on their classification precision. The best performing algorithm achieved 88.2% precision.