TwitterSports: Real Time Detection of Key Events from Sports Tweets

Abstract

Twitter users play a role of human sensors and update information about real-life events by posting their tweets about them. Event detection in Twitter is the process of detecting an event which is an occurrence causing change in the volume of tweets that discuss the associated topic at a specific time and a location by Twitter users. Twitter has been extensively used to detect major social and physical events such as earthquakes, celebrity deaths, presidential elections, traffic jam and others. Real time event detection in Twitter is detecting real-life events from live tweets instantly as soon as the event has occurred. Real time event detection from Cricket sports using Twitter media is an interesting, yet a complex problem. Because, event detection algorithm needs live tweets streamed at real-time about the game and should detect events such as boundary and sixer, at near real-time within few seconds from their occurrences. In this paper, a novel real-time event detection approach is proposed for the Cricket sports domain. The proposed approach first computes the post rate of an adaptive window, which is the ratio between the volumes of tweets in the second half window and the volume of tweets in the first half. An event has occurred if the post rate is above the pre-defined threshold, otherwise the algorithm selects the next big window in an adaptive manner. The predefined threshold helps to filter out the small spikes in the streaming tweets volume. Once an event is detected in a time window along the tweet stream, the event represented inside the window is recognized using the event lexicon representing different events of a cricket game. The proposed real-time event detection algorithm is extensively evaluated on 2017 IPL T20 Cricket sports dataset using ROC and AUC evaluation measures. The experimental results on the performance of the proposed approach show that the adaptive sliding window detects sports events with over 80% true positives and around 15% false positive rates.