Modeling Anticipatory Event Transitions

Abstract

Major world events such as terrorist attacks, natural disasters, wars, etc. typically progress through various representative stages/states in time. For example, a volcano eruption could lead to earthquakes, tsunamis, aftershocks, evacuation, rescue efforts, international relief support, rebuilding, and resettlement, etc. By analyzing various types of catastrophical and historical events, we can derive corresponding event transition models to embed useful information at each state. The knowledge embedded in these models can be extremely valuable. For instance, a transition model of the 1918-1920 flu pandemic could be used for the planning and allocation of resources to decisively respond to future occurrences of similar outbreaks such as the SARS (severe acute respiratory syndrome) incident in 2003, and a future H5N1 bird-flue pandemic. In this chapter, we study the Anticipatory Event Detection (AED) framework for modeling a general event from online news articles. We analyze each news document using a combination of features including text content, term burstiness, and date/time stamp. Machine learning techniques such as classification, clustering, and natural language understanding are applied to extract the semantics embedded in each news article. Real world events are used to illustrate the effectiveness and practicality of our approach.KeywordsNews ArticleClass EntropyTerm Weighting SchemeLexical ChainTopic TrackThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.