Event relatedness assessment of Twitter messages for emergency response

Abstract The ubiquity of smartphones and social media such as Twitter is clearly blurring traditional boundaries between producers and consumers of information. This is especially the case in emergency situations where people in the scene create and share on-the-spot information about the incident in real time. However, despite the proven importance of such platforms, finding event-related information from the real-time feeds of thousands of tweets is a significant challenge. This paper introduces a novel method for detecting event-specific and informative tweets that are likely to be beneficial for emergency response. The method investigates a sample dataset of tweets which was collected during a storm event passing over a specific area. The sample is manually labelled by three emergency management experts who annotated the sample dataset to obtain the ground truth through identification of the event-related tweets. A selected number of representative event-related tweets are used to extract the common patterns and to define event related term-classes based on term frequency analysis. The term-classes are used to evaluate the event relatedness of a sample dataset through a relationship scoring process. Consequently, each sample tweet is given an event-relatedness score which indicates how related a tweet is to the storm event. The results are compared with the ground truth to determine the cut-off relatedness score and to evaluate the performance of the method. The results of the evaluation indicate that the proposed method is able to detect event-related tweets with about 87% accuracy in a timely manner.