Generation for Configuration: A Conceptual Paradigm of a Natural Language-Based Configurator for Modular Buildings with ChatGPT

Modular construction methods have demonstrated notable improvements in productivity and quality control compared with conventional methods. To further increase efficiency and flexibility in the construction industry, the concept of mass customization through a configurator has been adopted from the manufacturing industry. Previous efforts in configurators rely heavily on direct client involvement for configuration. However, clients’ inherent semantic gap and knowledge lacuna form a natural barrier to promoting configurator efficiency. Additionally, data deficiency and system maintenance hardship hinder the creation of a robust configurator. To ameliorate these gaps, this work proposes a conceptual paradigm of a natural language-based configurator with the help of ChatGPT, a state-of-the-art generative model. The configurator's primary strengths lie in its simplicity and generalizability, as it makes decisions based solely on natural language expressions provided by clients rather than on pre-defined options and components. To obtain an adequate amount of data for supervised learning, ChatGPT is utilized to generate vivid user requirements. Deep learning methods are then applied to characterize the relationship between user requirements and existing variants. As a practical implementation, a configurator for selecting suitable modular houses is developed. This research contributes to the field by offering a novel conceptual model design, realistic data collection, and model construction. The proposed paradigm illustrates its superiority and potential to facilitate decision-making while effectively fulfilling client needs, as demonstrated through a concrete experiment and visualization.