Transfer Learning-Based Question Generation for Building a Construction Safety Chatbot

Fall hazard is one of the leading causes of fatalities in the construction industry. To alleviate fall risks, the US Occupational Safety and Health Administration (OSHA) agency developed safety regulations for construction project team members. However, retrieving relevant information on fall protection from OSHA’s text-heavy standards is a challenge for personnel with limited safety knowledge and experience. With the rise of Industry 4.0, chatbots have become increasingly popular, offering a solution to the information overload in the architecture, engineering, construction, and operations (AECO) industry. This research aims to develop a dataset for building and validating intelligent construction safety chatbots (CSCs) to improve the efficiency of information retrieval from safety regulations (i.e., OSHA) using fall protection as an example. This research uses two state-of-the-art language models, Text-To-Text Transfer Transformer (T5) and Generative Pre-trained Transformer 3 (GPT-3), to generate natural language questions from US OSHA textual regulations about fall protection. The performance of the two models in question generation is compared and evaluated. The evaluation shows that T5 generates questions that are more relevant and diverse in structure and question type in comparison to GPT-3. Based on the evaluation results, the T5-generated questions are used to build a CSC dataset. The developed dataset can be used to train and evaluate a CSC, and the developed method can be extended to other fields, such as building codes and contracts. The contribution of this research is that it facilitates the adoption of transfer learning and chatbot technologies in the AECO industry.