Adaptive topic tracking based on Dirichlet process mixture model

Abstract

This paper proposes a Dirichlet Process Mixture Model (DPMM) considering relevant topical information for adaptive topic tracking. The method has two characters: 1) It uses DPMM to implement topic tracking. Prior knowledge of known topics is combined in Gibbs sampling for model inference, and correlation between a story and each known topics can be estimated. 2) To alleviate topic excursion problem and topic deviation problem brought by existing adaptive tracking methods, the paper presents a new adaptive learning mechanism, the basic idea of which is to introduce tracking feedback with a reliability metric into the topic tracking procedure and make tracking feedback influence tracing computation under the condition of the reliability metric. The empirical results on TDT3 evaluation data show that the model, without a large scale of in-domain data, can solve topic excursion problem of topic tracking task and topic deviation problem brought by existing adaptive learning mechanisms significantly even with a few on-topic stories. © 2012 Springer-Verlag.