Dragon’s Tracking and Detection Systems for the TDT2000 Evaluation

We describe the tracking and detection systems submitted by Dragon for the TDT2000 evaluation. Our research focus is on improving the distance measure between story and story collection, a computation which is central to many of the TDT tasks. In our tracking engine, we improved the measure by strengthening our targeting procedure, and introduced unsupervised adaptation on high-scoring test stories. Our detection engine uses a new measure, developed under tracking experiments, that performs 15–20% better than the previous measure, and makes use of targeting in a computationally tractable way. In both tasks, we explore the effect of automatically segmenting broadcast news stories and of introducing word stemming.