A framework for adopting Twitter data in emergency response

Abstract

Timely access to up-to-date, relevant and the geographically referenced emergency information is considered as the essential requirement of the emergency response. The considerable role of social media in providing a source of timely emergency information and supporting the emergency response activities has been demonstrated during a series of real-world events, such as earthquakes, floods, and terrorist attacks. Among the available social media platforms, Twitter has been an effective way of sharing information and communicating warnings between disaster relief organisations and citizens on the ground during many real-world incidents. However, despite the inherent capability of Twitter in meeting the timeliness requirement, there are major challenges in the identification of event-related tweets as well as the estimation of the actual location of Twitter data.

First of all, finding tweets related to a specific event from tens of thousands of tweets posted every minute is a non-trivial task. The second issue lies in the spatial aspect of Twitter data. Tweets sent from location-enabled devices can be geotagged containing the precise location coordinates. However, only about 1% to 3% of all tweets are geotagged. Systems or tools that rely on geotagging can only benefit from a small fraction of Twitter data, even though there is valuable information in the remaining non-geotagged portion. Accordingly, assessment of Twitter messages to identify the event-related tweets, along with inferring their location with the highest possible accuracy are considered as priority research areas regarding the adoption of Twitter data in emergency response.

To address the above-mentioned challenges, this research was conducted by using Design Science Research Methodology (DSRM) as the overarching research method to develop a framework for adopting Twitter in emergency response. The developed framework aims at the identification of the tweets related to an event of interest within the study area along with inferring their approximate location in a timely manner. The proposed framework for event-relatedness assessment and location inference of Twitter data was successfully implemented into a prototype system and its feasibility was verified through a case study demonstration using a sample dataset of tweets collected during a real-world emergency. The results of the prototype demonstration were used to empirically evaluate the performance of the prototype through the calculation of several performance metrics. The prototype achieved an overall success rate of about 75% in performing both the event-relatedness assessment and location inference of the sample tweets. In the end, the conclusions drawn from the proposed framework and its implementation and evaluation together with the implications for practice and future research were presented.