Deep Learning-Based Coreference Resolution for Bridge Report Analytics

Bridge textual reports include massive amounts of data/information about bridge conditions. Extracting and analyzing this data offers opportunities to enhance bridge maintenance decision making. However, utilizing the data/information embedded in textual bridge reports for bridge data analytics remains to be a challenge due to the inherent coreference ambiguities in the text. There is a need for semantic analysis methods to disambiguate coreferences inside sentences to enhance the performance of data/information extraction from multiple types of reports. There is a lack of efforts that addressed coreference resolution in this domain. To address this gap, this paper proposes a deep learning-based coreference resolution model that automatically disambiguates the coreference ambiguities from multiple types of unstructured textual bridge reports and represents the extracted data/information in a structured format to support further data analytics. The proposed model utilizes a deep learning-based bidirectional long short-term memory (BiLSTM); a begin, inside, and outside (BIO) encoding schema; convolutional neural network (CNN) features; and span representations for coreference resolution. The model was tested on multiple types of textual bridge reports, including bridge repair manuals, bridge assessment reports, and bridge rehabilitation reports. The proposed model showed promising coreference resolution performance and generalizability across different types of reports.