Automated Risk Analysis for Construction Contracts Using Neural Networks

Artificial intelligence (AI) application has been recently utilized in various commercial trades. In the past 20 years, researchers made various attempts in applying AI as a supporting tool in construction management, unfortunately, most of these attempts are neither mature enough nor well developed, due to the sophisticated nature of the construction industry considering its diversified fields. Overviewing construction projects and shedding light on the construction management stages, it can be clear, identification, categorization, and impact assessment of contractual risks consumes extensive amounts of effort and time during the estimation process in the tendering stage. This process has proven to be very challenging and risky due to the tight duration usually allocated for such crucial activities. Consequently, hurrying the aforesaid bidders most likely leads to inaccurate pricing leading to unavoidable legal disputes threatening construction projects' success. Therefore, developing an AI model for bidders to support them in proper and accurate pricing, and the early stage enhancement of the risk management, in addition to a potential reduction in the possible disputes that might arise between contracting parties at a later stage. This research presents a supervised machine learning model programmed using Python language, adopting artificial neural networks (ANN), established, and trained to identify the risky clauses, their level of severity, and their expected impact (time, cost, or both). The collected data set extracted from real five construction contracts generating 486 clauses; these clauses were analyzed by eight domain experts (through two-stage interviews) to provide risk ranking and its probable impact through a predetermined question. A preprocessing stage is conducted for utilizing the collected interview replies in a suitable format for the ANN model. The python-based model uses transformers to predict clauses' risk rank and their probable impact. The data set is split into 80% training and 20% validation, results show high validation percentages for risk impact and risk ranking.