Knowledge-informed semantic alignment and rule interpretation for automated compliance checking

As an essential prodecure to improve design quality in the construction industry, automated rule checking (ARC) requires intelligent rule interpretation from regulatory texts and precise alignment of concepts from different sources. However, there still exists semantic gaps between design models and regulatory texts, hindering the exploitation of ARC. Thus, a knowledge-informed framework for improved ARC is proposed based on natural language processing. Within the framework, an ontology is first established to represent domain knowledge, including concepts, synonyms, relationships, constraints, etc. Then, semantic alignment and conflict resolution are introduced to enhance the rule interpretation process based on predefined domain knowledge and unsu-pervised learning techniques. Finally, an algorithm is developed to identify the proper SPARQL function for each rule, and then to generate SPARQL-based queries for model checking purposes, thereby making it possible to interpret complex rules where extra implicit data needs to be inferred. Experiments show that the proposed framework and methods successfully filled the semantic gaps between design models and regulatory texts with domain knowledge, which achieves a 90.1% accuracy and substantially outperforms the commonly used keyword matching method. In addition, the proposed rule interpretation method proves to be 5 times faster than the manual interpretation by domain experts. This research contributes to the body of knowledge of a novel framework and the corresponding methods to enhance automated rule checking with domain knowledge.

Automated summary

This study proposes a knowledge-informed framework based on natural language processing to enhance automated rule checking. By establishing an ontology to represent domain knowledge and introducing semantic alignment and conflict resolution, the semantic gaps between design models and regulatory texts are filled. Experimental results show that the framework and methods achieve an accuracy of 90.1% and are 5 times faster than manual interpretation by domain experts.