Scholarlink

<p>We present a dataset of over 3,000 global disaster events from 2014 to 2024, derived from the Emergency Events Database (EM-DAT). Events are extracted from news using a pipeline combining Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) for semantic extraction. The corpus is the Europe Media Monitor (EMM), aggregating content from millions of news outlets. For each event, structured storylines are automatically generated, summarizing hazard characteristics, drivers, impacts, and responses, and transformed into knowledge graphs. This enables analysis of relationships, inter-hazard dynamics, and human-environment interactions often missed in traditional records. A small subset of knowledge graphs was evaluated by domain experts in a workshop, while a larger sample of extracted triplets was independently assessed to quantify precision and inter-annotator agreement. The dataset supports retrospective analysis and multi-hazard risk assessment, complementing resources like the Hazard Information Profiles (HIPs). All data, code, and workflows are openly available, with an interactive dashboard for exploration. This resource advances data-driven approaches to disaster scenario modeling, impact analysis, and decision support in disaster risk management.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Disaster Storylines and Knowledge Graphs from Global News with Large Language Models and Retrieval-Augmented Generation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月17日

查看全文
  • Michele Ronco,
  • Luca Bandelli,
  • Lorenzo Bertolini,
  • Sergio Consoli,
  • Damien Delforge,
  • Alessio Spadaro,
  • Marco Verile,
  • Christina Corbane

摘要

We present a dataset of over 3,000 global disaster events from 2014 to 2024, derived from the Emergency Events Database (EM-DAT). Events are extracted from news using a pipeline combining Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) for semantic extraction. The corpus is the Europe Media Monitor (EMM), aggregating content from millions of news outlets. For each event, structured storylines are automatically generated, summarizing hazard characteristics, drivers, impacts, and responses, and transformed into knowledge graphs. This enables analysis of relationships, inter-hazard dynamics, and human-environment interactions often missed in traditional records. A small subset of knowledge graphs was evaluated by domain experts in a workshop, while a larger sample of extracted triplets was independently assessed to quantify precision and inter-annotator agreement. The dataset supports retrospective analysis and multi-hazard risk assessment, complementing resources like the Hazard Information Profiles (HIPs). All data, code, and workflows are openly available, with an interactive dashboard for exploration. This resource advances data-driven approaches to disaster scenario modeling, impact analysis, and decision support in disaster risk management.