Social event detection on Twitter

Abstract

Various applications are developed today on top of microblogging services like Twitter. In order to engineer Web applications which operate on microblogging data, there is a need for appropriate filtering techniques to identify messages. In this paper, we focus on detecting Twitter messages (tweets) that report on social events. We introduce a filtering pipeline that exploits textual features and n-grams to classify messages into event related and non-event related tweets. We analyze the impact of preprocessing techniques, achieving accuracies higher than 80%. Further, we present a strategy to automate labeling of training data, since our proposed filtering pipeline requires training data. When testing on our dataset, this semi-automated method achieves an accuracy of 79% and results comparable to the manual labeling approach. © 2012 Springer-Verlag.