Enhancing Bridge Inspection Analysis through NLP and LLM: A Case Study using Ohio's NBI Reports

Abstract: Bridges are critical components of transportation infrastructure, so they need robust management systems to ensure their safety and longevity. The National Bridge Inventory (NBI) conducts comprehensive biennial inspections, creating extensive datasets with over two hundred features per bridge. These datasets are essential for assessing bridge conditions and monitoring their performance over time. However, a significant challenge in bridge assessment is the effective integration and analysis of subjective inspector notes and reports. These notes are rich with insights but are often underutilized due to their unstructured nature. Previous research has predominantly focused on evaluating bridge conditions using quantitative data from these datasets, overlooking the textual data that contain valuable qualitative information. This study harnesses Natural Language Processing (NLP) and advanced Large Language Models (LLM) to effectively mine textual descriptions. Employing unsupervised learning techniques such as Latent Semantic Analysis (LSA), Latent Dirichlet Allocation (LDA), and the BERTopic model, we have extracted meaningful topics from inspector notes within the Ohio Department of Transportation (ODOT) dataset. Our framework demonstrates how these models can enhance the precision of bridge evaluations by incorporating qualitative insights, offering a scalable and replicable approach for scholars and transportation agencies to gain a deeper and more efficient understanding of bridge conditions.