In this paper we introduce a novel approach for incrementally building aspect models, and use it to dynamically discover underlying themes from document streams. Using the new approach we present an application which we call \query-line tracking" i.e., we automatically discover and summarize different themes or stories that appear over time, and that relate to a particular query. We present evaluation on news corpora to demonstrate the strength of our method for both query-line tracking, online indexing and clustering.

PDF file

Publisher: Springer-Verlag
All copyrights reserved by Springer 2004.

Details