Incident Linking: Assigning Tweets to Entries in a Disaster Database

Abstract

Twitter offers outstanding potential for collecting information about disaster events. However, these disaster events usually are recorded in incident databases. Consequently, we want to extract the structured data from tweets to supplement the entries in the traditionally acquired incident databases if some of the metadata (impact, location, description, time) is missing and create new entries, if necessary. To accomplish this, there is a need for a linking method that allows analyzing the relation between tweet texts and incident metadata. Unfortunately, established algorithms for event detection and entity linking cannot solve this problem. Event detection algorithms require a large volume of tweets for each event, which is often unavailable, and they do not attempt to connect tweets and incident databases. Entity linking algorithms require named entities in the databases, which are often missing in event databases. A new Incident Linking Framework (ILF) is proposed and evaluated in this thesis to accomplish this task automatically. This method is a two-step approach that includes individual components like candidate generation and candidate ranking. The candidate generation step returns the possible incidents that match a tweet. The candidate ranking ranks these candidates individually using scoring metrics and produces the best-fit pair of tweets and incidents. There are two different experiments conducted to evaluate this approach. Results for ILF shows that tweets can automatically get assigned to the incident database only if it matches the metadata. In addition, the candidate generation shows promising results compared to the candidate ranking.