The CMU-developed system performed the best in TDT2002. The guidelines for detecting a new event were the similarity of the story with past history, and similarity with others in a deferral window. The latter however didn’t carry much influence. The novelty of a story was characterized by a score calculated using logistic regression. The features used were (a) best similarity between the current cluster and history clusters (b) cluster size of current and history clusters (c) time distance between the two clusters and (d) best similarity score between the current story and future story in the forward window. A novel story will be dissimilar to the stories in the past history, and will be similar to stories in the deferral window. A score for the former is calculated by determining the cosine similarity of the current story with the centroids produced by incremental clustering of the historical story sequence. The latter’s score was calculated by agglomerative clustering of the look-ahead window stories and similarity comparison of those more limited centroids with the new story.

-- GiridharKumaran - 26 Mar 2003
to top