Finding Fires with Twitter

Abstract

This paper presents a notification system to identify in near-real-time Tweets de-scribing fire events in Australia. The sys-tem identifies fire related ‘alert words’ published on Twitter which are further processed by a classifier to determine if they correspond to an actual fire event. We describe how the classifier has been estab-lished and report preliminary results. The original notification system did not in-clude a classifier and could not discrimi-nate between messages unrelated to ‘real’ fire events. In the first three months of operation, the system generated 42 ‘fire’ email notifications of which 20 related to actual fires and 12 of those contained Tweets that may have been of interest to fire fighting agencies. If the classifier had been used, 21 emails would have been is-sued: an improvement in accuracy from 48 % to 78%. However, the recall score reduced from 1 to 0.8 which is not desir-able for this particular task. We propose extensions to address this short coming. 1