Leveraging Transformers for Improved Decision-Making in Road Maintenance

Abstract

Transportation performance is heavily influenced by the overall quality and the effectiveness of road maintenance. However, this remains an expert-dependent activity, despite recent efforts to digitalize road geometries and management processes. Road maintenance knowledge accumulated through addressing relevant enquiries is inexplicitly learned by experts and transferred into experience, which contributes very little to developing maintenance digitalization techniques and automated decision-making processes. In this case, fully utilizing historical maintenance records and turning them into computer-readable knowledge is a crucial task to be solved. This paper aims to extract key information from road maintenance request texts and then implement step-by-step thinking to make road maintenance decisions. This chain of thought is first proposed by reviewing the key elements and logical flow of road maintenance decision-making. Then, a cross-attention mechanism based on a transformer architecture is implemented on maintenance record texts and target knowledge element sequences. The result of this experiment overperforms on a pre-trained BERT model and demonstrates a valid performance on the text-knowledge alignment in road maintenance domain. The method proposed in this paper provides a solution for reliable and traceable decision-making and shows a promising application in domain-specific knowledge management.