A dataset from ten disasters for studying location descriptions and training AI models

Area of Science:

• Natural Language Processing
• Disaster Informatics
• Social Media Analysis

Background:

• Social media is crucial for disaster communication, detailing locations of victims, damage, and resources.
• Existing datasets lack detailed, multi-entity location descriptions from disaster messages.

Purpose of the Study:

• To address the scarcity of labeled data for disaster-related location descriptions.
• To facilitate research on how locations are described during natural disasters.
• To enable the development of AI models for automatic location extraction.

Main Methods:

• Collected Twitter/X messages related to ten U.S. disasters (hurricanes, floods, wildfires, tornados, winter storms).
• Developed a dataset with labeled detailed location descriptions (addresses, intersections, exits).
• Described the data collection, annotation, and validation processes.

Main Results:

• A novel dataset of disaster-related social media messages with labeled detailed locations is now available.
• The dataset covers diverse disaster types and geographic regions within the United States.
• The methodology for creating the dataset is detailed, ensuring transparency and reproducibility.

Conclusions:

• The created dataset fills a critical gap for studying disaster communication and training AI.
• This resource will advance the automatic extraction of vital location information from social media during emergencies.
• Future research can leverage this dataset to improve disaster response and situational awareness.