Unsupervised Discovery of Event Scenarios from Texts
Cosmin Adrian Bejan · 2008
We propose a new unsupervised learning approach for discovering event scenarios from texts. We interpret an event scenario as a collection of related events that characterize a specific situation. The approach uses the Latent Dirichlet Allocation (LDA) probabilistic model described in (Blei, Ng, & Jordan 2003) to automatically learn the probability distribution of events correspond-ing to event scenarios. We performed experiments on an event annotated cor-pus and compared the automatically extracted event scenarios with frame scenarios defined in FrameNet (Baker, Fillmore, & Lowe 1998). The results show a better coverage for those event scenarios that are de-scribed in more detail in the event annotated corpus. When compared with a smoothed unigram model, the event scenario model achieves a perplexity reduction of 93.46%.