Ont4RE, a computation framework for distant supervision and entity-property relation extraction in construction documents

Engineering documents in construction management contain rich information, including various project entities and their corresponding properties, forming entity-relation-entity triples. This information is valuable for many management activities, such as quality inspection and compliance Checking. However, it is challenging to process due to the unstructured form of the original texts. Thus, triple extraction enabled by Deep Learning (DL) has become an urgent need in the sector. However, fully supervised methods are impractical as they require enormous amounts of labelled data. Distant supervision alleviates the burden of manual labelling but introduces much noise and requires establishing priori large-scale knowledge bases. Thus, the study proposes a novel framework for entity-property triple extraction, i.e., Ontology for Relation Extraction (Ont4RE). The Ont4RE leverages an ontology to integrate domain knowledge and automatically annotate a construction corpus without manual intervention, before a DL model is employed to extract triples. In experiments, the study verified that Ont4RE can achieve precise auto-labelling while being adaptable to different extractors with varying model structures. Specifically, the Area Under Precision-Recall Curve (AUPRC), mean P@N, and F1-score among tested extractors improved by 6.1%, 3.3%, and 11.7% on average, respectively. The best performance was achieved with BERT as the relation extractor (92.3% AUPRC and 93.5% F1-score). The results demonstrate that when searching for project information, the Ont4RE was superior to manual searching and artificial intelligence tools (ChatGPT-4) in terms of precision, recall, and robustness. Furthermore, it reduces information extraction time from 98 s to 2.5 s. Thus, Ont4RE can substantially improve information extraction efficiency in construction management, thereby facilitating management tasks and decision-making.