Abstract:

A Dirichlet process mixture model which can make use of information of known topics efficiently is proposed for topic tracking. Prior knowledge of known topics is combined in Gibbs sampling for model inference, and similarities between new story and known topics can be gained. Experiments show that the model, without a large scale of indomain data, can improve the performance of topic tracking significantly even with a few ontopic stories.

Key words: topic tracking, Dirichlet process mixture model, Gibbs sampling, known topics