Semantic Neural Network Ensemble for Automated Dependency Relation Extraction from Bridge Inspection Reports

Abstract

Bridge inspection reports are important sources of technically detailed data/information about bridge conditions and maintenance history, yet remain untapped for bridge deterioration prediction. To capitalize on these reports for improved bridge deterioration prediction, there is a need for dependency parsing methods, in order to extract dependency relations from the reports for linking the isolated words into concepts and representing the semantically low concepts in a semantically rich structured way. To address this need, this paper proposes a novel semantic neural network ensemble (NNE)–based dependency parsing methodology. It uses a similarity-based method to sample similarly distributed configurations into the same clusters, a set of constituent neural network (NN) classifiers to learn from both the syntactic and semantic text features of the similarly distributed and therefore more easily separable configurations, and a combiner classifier to capture the classification patterns of the NN classifiers to make final predictions on the transition types. The proposed dependency parsing methodology was evaluated in extracting dependency relations from bridge inspection reports for representing information—about bridge conditions and maintenance actions—in a semantically rich structured way. It achieved a precision, recall, and F-1 measure of 96.6%, 90.4%, and 93.3% with a margin of error of 3.8%, 4.4%, and 3.8% at the semantic information element level, and 88.2%, 81.5%, and 84.7% with a margin of error of 5.4%, 5.8%, and 5.4% at the semantic information set level, respectively.