Development of knowledge graphs for emergency events utilizing pre-trained language models
- 1Northwest Institute of Eco-Environment and Resources, Chinese Academy of Sciences
- 2National Science Library, Chinese Academy of Sciences
Abstract
Emergency management requires structured representations that can integrate heterogeneous event reports and support cross-event analysis. This study develops an emergency event knowledge graph (EE-KG) by combining fine-grained event classification, lifecycle-oriented ontology modeling, and ontology-guided knowledge extraction. We collected 6,418 Chinese emergency-related reports from public government portals, news websites, and safety information platforms published between 2000 and 2024, and aggregated them into 1,971 event records. A manually annotated multi-label dataset was constructed according to four primary categories and 36 subcategories. Four classification strategies were evaluated: fine-tuned ERNIE 3.0, prompt-based GLM-4-9B-Chat, agent-based GLM-4-9B-Chat with reflective reasoning, and fine-tuned GLM-4-9B-Chat. The fine-tuned GLM-4-9B-Chat achieved the best performance, with an accuracy of 0.80 and a micro-F1 score of 0.85. Classification outputs were used to assign standardized event types to event nodes, while a five-dimensional ontology guided the extraction of event description, impact assessment, emergency response, recovery and reconstruction, and evaluation information. The resulting EE-KG contains 16,057 nodes and 18,428 relationships. Expert review indicates that the proposed framework provides a reusable structure for emergency event knowledge organization and analysis.
Related
- A Reproducible Decomposed Large Language Model Pipeline for Structured Abstraction of Multimodal Neuroimaging Reports: Development and Independent Validation in 1,669 Emergency Department Encounters for Dizziness and Vertigo
- Development and validation of a nomogram-based predictive model for infectious diseases of the spine utilizing peripheral blood laboratory parameters
- Utilizing the Contrastive Language-Image Pre-Training Model in Chest Radiography: Current Capabilities and Practical Guidance.
- A systematic workflow for utilizing and evaluating large language models in public health policy framing analysis.
- A comparative study of knowledge, attitude, and practice of rapid trauma survey using traditional and online methods among interns being trained at a rural tertiary health care center
BibTeX
@article{yang2026development,
title = {{Development of knowledge graphs for emergency events utilizing pre-trained language models}},
author = {Heng Yang and Guangzu Bai and Nan Wang and Sili Wang},
journal = {Scientific Reports},
year = {2026},
month = jul,
doi = {10.1038/s41598-026-57570-7},
url = {https://doi.org/10.1038/s41598-026-57570-7},
note = {CC BY},
}
RIS
TY - JOUR
TI - Development of knowledge graphs for emergency events utilizing pre-trained language models
AU - Heng Yang
AU - Guangzu Bai
AU - Nan Wang
AU - Sili Wang
PY - 2026
DA - 2026/07/01
JO - Scientific Reports
DO - 10.1038/s41598-026-57570-7
UR - https://doi.org/10.1038/s41598-026-57570-7
AB - Emergency management requires structured representations that can integrate heterogeneous event reports and support cross-event analysis. This study develops an emergency event knowledge graph (EE-KG) by combining fine-grained event classification, lifecycle-oriented ontology modeling, and ontology-guided knowledge extraction. We collected 6,418 Chinese emergency-related reports from public government portals, news websites, and safety information platforms published between 2000 and 2024, and aggregated them into 1,971 event records. A manually annotated multi-label dataset was constructed according to four primary categories and 36 subcategories. Four classification strategies were evaluated: fine-tuned ERNIE 3.0, prompt-based GLM-4-9B-Chat, agent-based GLM-4-9B-Chat with reflective reasoning, and fine-tuned GLM-4-9B-Chat. The fine-tuned GLM-4-9B-Chat achieved the best performance, with an accuracy of 0.80 and a micro-F1 score of 0.85. Classification outputs were used to assign standardized event types to event nodes, while a five-dimensional ontology guided the extraction of event description, impact assessment, emergency response, recovery and reconstruction, and evaluation information. The resulting EE-KG contains 16,057 nodes and 18,428 relationships. Expert review indicates that the proposed framework provides a reusable structure for emergency event knowledge organization and analysis.
SN - 2045-2322
ER -