Anticipatory Event Detection for Bursty Events

摘要

Anticipatory Event Detection (AED) is a framework for monitoring and tracking important and relevant news events at a fine grain resolution. AED has been previously tested successfully on news topics like NBA basketball match scores and mergers and acquisitions, but were limited to a static event representation model. In this paper, we discuss two recent attempts of adding content burstiness to AED. A burst is intuitively a sudden surge in frequency of some quantifiable measure, in our case, the document frequency. We examine two schemes for utilizing the burstiness of individual words, one for revamping the static document representation, and the other for extracting bursty and discriminatory words from the two states of the AED Event Transition Graph.