Crisis Events Template Generation and Information Extraction using LLM

Abstract

In an age where real-time access to reliable information is critical, crisis events such as hurricanes, earthquakes, and public safety incidents often generate a deluge of fragmented, unstructured data spread across numerous online sources. This makes it difficult for emergency responders, researchers, and decision-makers to quickly extract accurate, actionable insights. Our capstone project, Crisis Events Template Generation and Information Extraction using LLM, addresses this challenge by developing a web application that automates the process of summarizing crisis events and extracting key information from a set of user-provided webpages. The core functionality of the application is powered by a Large Language Model (LLM), which enables both the generation of structured templates for different types of crises (e.g., natural disasters, mass casualty events) and the automatic filling of those templates with relevant facts derived from online content. Users select a crisis category and upload URLs related to a specific event; the system then scrapes and cleans the text using BeautifulSoup [8], passes the content to the LLM for inference, and returns a populated template highlighting key data points such as dates, locations, intensities, and impacts. The project combines a modern web technology stack: a responsive frontend built with Next.js [1] and Tailwind CSS [2], secure authentication via GitHub OAuth [10] using NextAuth.js [3], and a backend implemented with Flask and MongoDB. The backend handles LLM-based processing and web scraping, exposing endpoints that communicate with the frontend via JSON. Docker is used to containerize and streamline deployment [6]. To date, our team has successfully built a functioning prototype with support for initial template generation, web scraping from multiple URLs, and text-to-slot mapping using LLMs. We’ve implemented role-based access control, enabling admins to manage templates while allowing end-users to generate and store summaries. We have also addressed key technical hurdles such as asynchronous scraping performance. Looking ahead, we plan to enhance the LLM prompting strategy, increase template flexibility, and broaden the range of supported crisis event types. Additionally, we will polish the user interface, improve error handling, and complete user and developer documentation. By project’s end, our goal is to deliver a scalable, user-friendly solution that enables quick and structured extraction of critical crisis event information paving the way for faster, data-informed responses to emergencies.