README.md

metadata

language:
  - en
task_categories:
  - question-answering
tags:
  - safety
  - workplace-safety
  - occupational-health
  - BERT
  - industrial-safety
pretty_name: Safety QA Dataset for BERT Fine-tuning

Safety QA Dataset

Dataset Description

There are two dataset that is publicaly available dataset from Mine Safety and Health Administration (MSHA). The 'seed_annotated_data.csv' dataset contains seed annotated data where the answer to the safety related questions are annotated in the accident narratives for initial training. The main 'training data.csv' data is used during the active learning (AL) process for question answering tasks in occupational safety and health domains. The dataset is specifically curated to fine-tune SafetyBERT model for safety-specific question answering applications.

Dataset Summary

This dataset supports the development of AI systems for understanding and analyzing workplace safety incidents. It includes detailed incident narratives with extracted information about body parts affected, work activities, and accident causes.

Supported Tasks

• Question Answering: The dataset is designed for extractive and abstractive question answering about workplace safety incidents using active learning (AL) approach.
• Text Classification: Can be used for categorizing safety incidents by type, cause, or affected body part
• Information Extraction: Extracting key safety-related entities from incident narratives

Languages

• English (en)

Dataset Structure

Data Files

The dataset consists of two main files:

1. seed_annotated_data.csv: Contains 149 manually annotated safety incident records.
2. training_data.csv: Extended training dataset for model fine-tuning

Data Fields

seed_annotated_data.csv:

• narrative (string): Full text description of the safety incident
• body_part (string): Body part(s) affected by the incident
• work_activity (string): Work activity being performed when incident occurred
• accident_cause (string): Primary cause or mechanism of the accident
• cluster (int): Cluster assignment based on similarity analysis
• distance_to_centroid (float): Distance metric from cluster centroid
• rank_in_cluster (int): Ranking within the assigned cluster

training_data.csv:

• [Describe the structure of your training data file]

Data Instances

Example from seed_annotated_data.csv:

{ "narrative": "Employee was carrying 5-gallon fuel cans up top the warming heaters and strained shoulder.", "body_part": "strained shoulder", "work_activity": "carrying 5-gallon fuel cans", "accident_cause": "up top the warming heaters", "cluster": 0, "distance_to_centroid": 0.09138608, "rank_in_cluster": 1 }

Annotations

Annotation Process

• Body Part: Manually extracted from incident narratives, identifying the specific body parts affected by the domain expert.
• Work Activity: Extracted descriptions of the task or activity being performed by the domain expert.
• Accident Cause: Identified root causes or contributing factors by the domain expert.
• Clustering: Applied unsupervised learning techniques to group similar incidents using the SafetyBERT model. More information is available in the relevant published paper.

Who are the annotators?

Annotations were performed by safety domain experts with experience in occupational health and safety analysis.

Dataset Curators

[Abid Ali Khan Danish (Michigan Technological University)]