Mapping textual descriptions to condition ratings to assist bridge inspection and condition assessment using hierarchical attention

Current bridge management strategies rely on experience-driven manually assigned condition ratings that are vulnerable to human subjectivity and experience variance. To improve the consistency of the condition rating practices, this study identifies narrative descriptions from bridge inspection reports as an untapped data source and proposes a data-driven framework as a supportive tool for two applications: automated condition recommendation and real-time quality control. A hierarchical architecture employing recurrent neural network encoders with an attention mechanism was developed using a collection of reports from the Virginia Department of Transportation. The condition recommendation application performed a classification task and demonstrated improved performance over a variety of baseline systems. The quality control application learns a data-driven decision threshold to decide whether to accept or reject an inspector-provided rating, which provides a cyberhuman collaboration route for condition assessment. Visualization of the resulting attention patterns was shown to provide interpretable insights which highlight potentially-overlooked indicators.

Automated summary

This study improves the consistency of bridge condition rating by using narrative descriptions from inspection reports, proposing a data-driven framework for two applications. By employing recurrent neural networks and an attention mechanism, a hierarchical architecture was developed to enhance the performance of condition recommendation.