Machine-Reading Comprehension for Bridge Inspection Domain by Fusing Graph Embedding

Abstract

Bridge inspection text records are very significant for the maintenance and upkeep of bridges, which can help engineers and maintenance personnel to understand the actual condition of bridges, detect and repair problems in time, and ensure the safe operation of bridges. Currently, more and more research focuses on how to extract potentially valuable bridge-related information from bridge inspection texts. In this study, we take the bridge inspection domain machine-reading comprehension corpus as the data support for model training and performance evaluation; oriented to the bridge inspection domain data text extraction machine-reading comprehension task, on the basis of word-granularity text input, we further explore two schemes of co-occurring linkage of cross-sentence entities in the context and co-occurring linkage of entities within the sentence through graph structure, and we learn and extract naming through graph-attentive neural networks-structured semantic information between entities and fused the obtained named entity embeddings with a hidden representation of the pretrained context. Tested on the bridge inspection domain dataset, the integrated model proposed in this research improves the EM optimum by 1.4% and the mean by 2.2% and the F1 optimum by 2.2% and the mean by 1.6% on the BIQA test, compared with the better-performing baseline model RoBERTa_wwm_ext.