Data and Code for: From Unstructured Injury Narratives to Structured Risk Pathways
Description
This dataset provides supporting data and code for the manuscript “From Unstructured Injury Narratives to Structured Risk Pathways: Revealing Recurrent Work–Hazard–Accident Mechanisms.” The study uses publicly available OSHA Severe Injury Reports from the U.S. Department of Labor as the original data source. The original OSHA data are not redistributed in full in this repository. Instead, this dataset provides materials that support the understanding and partial reproduction of the data-processing, annotation, structuring, and evaluation workflow reported in the manuscript. The repository includes Jupyter notebooks for converting Doccano annotations into NER training data, training the NER model, profiling the graph-based accident representation, comparing text-based and graph-based scenario similarity, and generating representative subgraph visualizations. It also includes the label schema, example normalization rules, graph schema description, sample processed data, and summary evaluation results. The full derived dataset generated in this study, including complete NER outputs, normalized graph-ready tables, and scenario-similarity results, is not publicly deposited because it contains author-generated annotations, intermediate modeling outputs, and research-specific normalization rules that are part of an ongoing research program. The full processed dataset may be made available from the corresponding author upon reasonable request, subject to institutional approval and research-use conditions. Original data source: OSHA Severe Injury Reports, U.S. Department of Labor.
Files
Steps to reproduce
1. Download the original OSHA Severe Injury Reports from the U.S. Department of Labor. The manuscript used the January 2015 to December 2024 release as the source population. 2. Review the README file to understand the repository structure, public source data, shared sample files, and limitations of the minimal reproducibility package. 3. Install the required Python packages using requirements.txt. 4. Run code/01_doccano_to_ner_dataset_v1.ipynb to convert the sample Doccano annotation file into the NER training format. 5. Run code/02_train_ner_model_v1.ipynb to reproduce the NER training and evaluation workflow on the provided sample and schema files. 6. Run code/03_graph_profile_OSHA.ipynb to inspect the graph schema, node/edge summaries, and representative Work–Hazard–Accident patterns using the provided sample processed data. 7. Run code/04_scenario_similarity_OSHA.ipynb to reproduce the text-based and graph-based scenario similarity comparison on the sample data. 8. Run code/05_visual_subgraph_OSHA.ipynb, if needed, to generate representative subgraph visualizations. The repository provides code, schema files, sample processed data, and summary evaluation results. The full derived dataset is not publicly deposited and may be made available from the corresponding author upon reasonable request, subject to institutional approval and research-use conditions.
Institutions
- Ewha Womans UniversitySeoul, Seoul