A Multi-Task Information Extraction Framework for Bridge Inspection Based on Joint Neural Networks

Abstract

Focused on the issue that insufficient information extraction and knowledge services in the bridge management and maintenance domain, a multi-task information extraction framework for bridge inspection based on joint neural networks is proposed. Firstly, a multi-task information extraction training dataset for bridge inspection is constructed and a distributed representation of the text is obtained using BERT as the embedding layer. Secondly, the subtasks of topic word detection and other bridge inspection information extraction are jointly learned by sharing BERT weights and fine-tuning, and the context features are further extracted in depth. Finally, the bridge inspection knowledge service is used as application examples to verify the effectiveness of the bridge inspection information extraction model in actual application scenarios such as bridge domain question answering. In the comparison experiments with mainstream models, the proposed method outperforms the mainstream models with F1-score of 85.27%, 72.73%, and 90.76% for the NER, RE, and topic word detection respectively. The experimental results show that the model can meet the requirements of a variety of practical tasks for information extraction of bridge inspection.