Knowledge Graph Enrichment of a Semantic Search System for Construction Safety

Minimising accident risk for new construction projects requires a thorough analysis of previous accidents, including an examination of circumstances and reasons for their occurrence, their consequences, and measures used for future mitigation. Such information is often recorded only within the huge amounts of unstructured textual documentation that are typically generated for large-scale construction projects. Accordingly, the process of locating, understanding, analysing and combining sufficient intelligence from documentation about previous projects can be overwhelming. Previously, text mining (TM) technology has been used to develop a semantically-enhanced search system over a repository of workplace accident reports, to increase the efficiency and effectiveness of locating important information relating to construction safety. In this article, we describe our enhancement of this system, through the generation and integration of a knowledge graph (KG). We extract triples consisting of subject, predicate and object (SPO) from unstructured text and map them to concepts in a knowledge base (KB). The mapping is carried out by comparing the contextualised representations of the text surrounding the SPO elements with concept descriptions in the KB. Subsequently, a Coreference Resolution method is used to detect mentions in text that refer to the same pronominal references that occur in SPO triples, to ensure that relevant knowledge is not overlooked. Finally, SPO triples are filtered to retain only those containing knowledge that is relevant to construction safety. We show that our approach can significantly outperform related methods, both in terms of detecting the elements of triples and linking them to entries in a KB.