Application of natural language processing in residential building defects analysis: Australian stakeholders' perceptions, causes and types

Abstract

The issue of defects in residential buildings has been reported to impact the performance of the architecture, engineering, and construction (AEC) industry. Defects can increase the construction cost, significantly contribute to the increment of construction waste and cause stress to home occupants. To minimise these effects the first step is to understand the main causes of building defects (Building defects). To date, the research specific to building defects in the Australian context is scant. Limited research considered the perceptions of all the possible stakeholders that are responsible for the generation and management of defects. Hence, this study aims to explore the causes of building defects by considering the perceptions of various stakeholders using the machine learning method. The research employed a mixed approach that involve qualitative content analysis and natural language processing (NLP) of court cases obtained from the Victorian Civil and Administrative Tribunal, the legal entity that deals with building-related disputes. NLP resulted in extracting defect sentences based on defect keywords using the KeyBERT algorithm and pre-trained deep learning embedding models including BERT-Base, RoBERTa-Base, and fastText. The content analysis showed that the top three reasons for building defects are related to workmanship, design, and materials. The three main stakeholder groups involved in building defect management were builders, owners and sub-contractors. Drawing on these findings, a proactive defect prevention framework was developed to guide building defect risk management.