BEST-KAG Enhances QA for Building Standards with Multimodal Knowledge Graphs

Construction standards are vital for ensuring building safety and sustainability, yet current workflows for applying these standards are often inefficient. They typically rely on keyword-based document retrieval and manual interpretation, which struggle with multi-clause reasoning, multimodal information, and traceable evidence linking. To overcome these limitations, a new framework called BEST-KAG (Knowledge-Augmented Generation for Building Engineering STandards) has been developed. BEST-KAG introduces three key components. First, a multimodal knowledge graph (MKG) provides a unified representation of document hierarchy and diverse standard knowledge, including various connections between elements. Second, a rule-LLM hybrid knowledge construction pipeline enables scalable extraction of multimodal knowledge, resulting in a large MKG built from 251 building engineering standards, comprising over 171,000 nodes and 310,000 edges. Finally, the framework incorporates a graph-retrieval-based knowledge-augmented generation architecture, which facilitates clause-grounded and traceable question answering. Experimental evaluations demonstrate that BEST-KAG consistently outperforms multiple mainstream large language models (LLMs) in expert assessments and metrics like BLEU and ROUGE, showing improvements of up to 74.01% over baselines. This advancement offers a more reliable and efficient way to access and apply complex building engineering knowledge.