Automatic bridge inspection database construction through hybrid information extraction and large language models

Regular bridge inspections generate extensive reports that, while critical for maintenance, often remain underutilized due to their unstructured format. Traditional information extraction methods depend on intricate labeling systems that commonly require time-consuming and labor-intensive labeling. This paper presents a novel bridge inspection database construction method leveraging LLM-assisted information extraction. First, we introduce the pseudo-labelling method using a closed-source LLM to generate high-quality data. Then we propose the hybrid extraction pipeline to extract relevant information segments and process them by a generation-based IE model, fine-tuned on pseudo-labeled data. Finally, the extracted data is used to construct the bridge inspection database. The proposed method, validated with real-world data, not only demonstrates higher extraction precision than the closed-source LLM used for pseudo-labeling but also outperforms traditional methods in both data preparation time and extraction accuracy. This approach provides a scalable solution for more proactive and data-driven bridge maintenance strategies. • High-quality pseudo-labeled are generated with closed-source large language models to facilitate downstream model training. • A hybrid information extraction pipeline, consisting of both rule-based and intelligent generation-based information extraction model, is proposed to effectively extract data from bridge inspection report documents. • The generation-based information extraction model, finetuned with pseudo-labeled data, outperforms the closed-source large language models, as well as traditional classification-based information extraction model.