Location Extraction in Disaster Tweets with a Model Trained on Past Data: Diverse Analysis

Rapid and accurate collection and dissemination of information is essential to minimize damage in a large-scale disaster. Governments and local authorities responsible for these tasks actively use immediate platforms like Twitter (now X) to gather and share information. However, the volume of information on social media increases rapidly during a large-scale disaster. It becomes necessary to swiftly select crucial, urgent information from many tweets. Identifying the locations mentioned in these tweets is also essential to facilitate decision-making by disaster responders. Considering these perspectives, quickly and manually sorting through the massive volume of posts is not easy, and attempts are being made to employ machine learning models for the sorting process. However, disaster response requires a rapid reaction, while machine learning models need high-quality training data to perform effectively. This study considers using posts circulated during past disasters to resolve these conflicting issues. A research question addressed is whether the type of disaster affects the accuracy of extracting location mentions using data from past disasters. This paper reports the verification results for the same and different disaster types.