Crisis analysis based on tweets

Information from the public during a crisis is often limited to people calling emergency services. Social media provides new opportunities to get input from the public during times of crisis. To avoid reading through massive amounts of social media data a system automatically detecting events is advantageous. Twitter’s simple format makes it easy to create tweets, although analyzing them is more of a challenge. The tweet stream’s high noise ratio combined with the shear amount of tweets makes detecting events a formidable challenge. To be able to detect any crisis events, the solution of this thesis is constructed as a general event detector, but with emphasis on spatial detection. This makes it possible not only to detect events, but in many cases also to estimate the location of these events. This approach combines features from crisis centric event detectors with general event detectors. The solution is constructed as a three part pipeline. The first part retrieves tweets. The second part detects events and is called the detection pipeline. The last part is a website called Grapher. It visualizes the detected events. The detection pipeline is the core of the solution. It consists of a temporal, word density and two spatial detection methods. In addition the detection pipeline clusters the suggested words from the methods. The detection methods are based on comparing two statistical models based on historic data and new data. The two spatial methods and the temporal method detects words and locations by comparing kernel density estimates with a state-of-the-art method. The solution pipeline has been extensively tested on real data. It is able to detect both crisis events and events of a more general character. For general events it has an event noise ratio of 65%. For crisis events it has an event noise ratio of 94%. The results show the proposed detection methods are viable and thus impacting the field of social media event detection. The solution could be applied by crisis handling teams and organizations monitoring social media in a specific area