Topic Detection through Statistical Methods

A system is developed to group news stories together according to topic. Several clustering algorithms can be used to group related stories into clusters. The clustering algorithms used require two types of metrics: metrics that, given a story and a set of clusters, can find the most topical cluster for that story; or metrics that can help decide whether or not a given story is on the same topic as a cluster. These metrics are derived by combining simple similarity metrics that compare stories and groups of stories. Finally, methods are proposed for evaluating the story groupings, and experimental results are reported based on these methods. Thesis Supervisor: Victor Zue Title: Associate Director of LCS and Senior Research Scientist