Bridge Damage Recognition from Inspection Reports Using NER Based on Recurrent Neural Network with Active Learning

Abstract

A deep understanding of the cause-effect relationship of bridge damages provides an opportunity to design, construct, and maintain bridge structures more effectively. The damage factors (i.e., bridge element, damage, and cause) and their complex relationships can be extracted from bridge inspection reports; however, it is not practical to manually read a considerable number of inspection documents and extract such valuable information. Although existing studies attempted to automatically analyze inspection reports, they require a large amount of human effort for data labeling and model development. To overcome the limitations, the authors propose an efficient information acquisition approach that extracts damage factors and causal relationships from bridge inspection reports. The named entity recognition (NER) model was developed based on a recurrent neural network (RNN) and was trained with the active learning method. In the experiments performed with 1,650 sentences (i.e., 1,300 for training and 350 for testing), the developed model successfully classified categories of text words (i.e., damage factors) and captured their causal relationship with 0.927 accuracy and 0.860 F1 score. Besides, the active learning method could significantly reduce the human effort required for data labeling and model development. The developed model achieved 0.778 F1 score only using 140 sentences, requiring less than an hour for manual labeling. These results meant that the model was able to successfully extract major damage factors and their cause-effect relationships from a set of text sentences with little effort. Consequently, the findings of this study can help field engineers to design, construct, and maintain bridge structures.