Language-Agnostic Event Detection Across Sports from Twitter Using Temporal Features

A significant body of work focuses on detecting events in social media, much of which relies on restrictive assumptions like prior event-centric filtering or known language models. Rather than build these expensive pre-processing and filtering pipelines, we explore language-agnostic techniques using only temporal features like token frequency. We construct temporal features for tokens in an unfiltered Twitter stream, build a classifier to recognize tokens experiencing bursts in usage, and relate those bursts back to known occurrences within sports. Our results demonstrate preliminary feasibility in rapidly detecting tokens that correspond to highimpact occurrences within sporting events. We also successfully transfer a model trained on American football and hockey data to previously unseen types of sports like the 2014 World Cup and the 2014 Kentucky Derby/Belmont Stakes horse races. This model has less success at identifying events in regular-season Major League Baseball, so more research is required to enhance these capabilities.