Entity Linking for real-time geolocation of natural disasters from social network posts

Area of Science:

• Natural Language Processing
• Crisis Informatics
• Geospatial Analysis

Background:

• Effective crisis management during natural disasters requires rapid situational awareness, often aided by social media data.
• Existing information extraction models struggle with real-time detection and geolocalization of disaster-related events.
• Most research focuses on English-language social media, leaving a gap for other languages.

Purpose of the Study:

• To develop and evaluate a system for detecting and geolocating natural disasters from French social media (Twitter/X) posts.
• To address the challenge of on-the-fly information extraction and mapping for crisis managers.
• To adapt existing natural language processing techniques for French disaster-related content.

Main Methods:

• Creation of a diverse French dataset including Wikipedia, CAp 2017 challenge data, and custom-annotated Twitter data from French natural disasters.
• Development of an Entity-Linking pipeline specifically designed for real-time prediction and peak resiliency applications.
• Evaluation of the system's performance against state-of-the-art methods.

Main Results:

• The developed Entity-Linking pipeline demonstrates performance comparable to existing state-of-the-art systems.
• Geolocated entities show strong coherence with the spatiotemporal patterns of the natural disasters analyzed.
• The system successfully detects and geolocates disaster information from French Twitter/X messages.

Conclusions:

• The proposed system offers a viable solution for automatic social network analysis in crisis management.
• It can significantly contribute to improving situational awareness for crisis managers dealing with French natural disasters.
• The approach validates the potential of geolocating social media data for real-time disaster response.